Paper deep dive

Weave Forcing: Compositional Memory Routing for Interactive Long Video Generation

Ziyi Wang, Junchi Yao, Heqian Qiu, Wenbo Shi, Chengjiu Wang, Jinyang He, Binkai Hong, Hongliang Li

Year: 2026Venue: arXiv preprintArea: cs.CVType: PreprintEmbeddings: 83

Intelligence

Status: not_run | Model: - | Prompt: - | Confidence: 0%

Entities (0)

No extracted entities yet.

Relation Signals (0)

No relation signals yet.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Recent advances in autoregressive video generation have improved temporal consistency over extended durations, yet interactive storytelling requires more than continuous scene extension: a new shot may combine characters and backgrounds from different historical shots. Whole prompt retrieval can overlook the distinct reference needs of individual components, while directly combining all historical memories may introduce unrelated visual content. To address these problems, we present Weave Forcing, a training-free framework for compositional memory reuse in interactive long video generation. First, we use an LLM for semantic slot routing to decompose user prompts into character and background descriptions and explicitly select suitable historical references for each component. To isolate the required content, masked memory weaving uses contrasting attention maps conditioned on semantic slots to construct refined semantic masks, selectively exposing relevant tokens from compressed historical KV memories to guide the generation of the current shot. We further introduce coverage adaptive RoPE to adjust temporal offsets and memory retention according to no, partial, or full reference coverage, addressing visual artifacts observed when incomplete historical references are positioned close to the current generation. Extensive experiments demonstrate that Weave Forcing improves cross-shot subject and background consistency while maintaining competitive visual quality and text alignment.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

79,322 characters extracted from source content.

Expand or collapse full text

Weave Forcing: Compositional Memory Routing for Interactive Long Video Generation Ziyi Wang Affiliation: University of Electronic Science and Technology of China (UESTC) Junchi Yao Affiliation: Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) Heqian Qiu Affiliation: University of Electronic Science and Technology of China (UESTC) Wenbo Shi Affiliation: University of Electronic Science and Technology of China (UESTC) Chengjiu Wang Affiliation: University of Electronic Science and Technology of China (UESTC) Jinyang He Affiliation: University of Electronic Science and Technology of China (UESTC) Binkai Hong, Hongliang Li †thanks: Corresponding author. Affiliation: University of Electronic Science and Technology of China (UESTC) Abstract Recent advances in autoregressive video generation have improved temporal consistency over extended durations, yet interactive storytelling requires more than continuous scene extension: a new shot may combine characters and backgrounds from different historical shots. Whole prompt retrieval can overlook the distinct reference needs of individual components, while directly combining all historical memories may introduce unrelated visual content. To address these problems, we present Weave Forcing, a training-free framework for compositional memory reuse in interactive long video generation. First, we use an LLM for semantic slot routing to decompose user prompts into character and background descriptions and explicitly select suitable historical references for each component. To isolate the required content, masked memory weaving uses contrasting attention maps conditioned on semantic slots to construct refined semantic masks, selectively exposing relevant tokens from compressed historical KV memories to guide the generation of the current shot. We further introduce coverage adaptive RoPE to adjust temporal offsets and memory retention according to no, partial, or full reference coverage, addressing visual artifacts observed when incomplete historical references are positioned close to the current generation. Extensive experiments demonstrate that Weave Forcing improves cross-shot subject and background consistency while maintaining competitive visual quality and text alignment. 1 Introduction Video diffusion models have rapidly advanced visual realism, motion complexity, and text alignment (Kong et al., 2024; Wan et al., 2025). Autoregressive models further enable efficient, temporally consistent long video generation through sequential prediction, historical context reuse, and improved training and denoising (Deng et al., 2025; Huang et al., 2026). However, filmmaking and interactive storytelling require users to revise instructions, switch shots, and recall earlier characters or backgrounds. Interactive multi-shot generation must therefore respond to new instructions while preserving continuity with relevant historical content (Lin et al., 2023). Existing approaches can be distinguished by the temporal attention structure of their video generators. Methods built on bidirectional diffusion transformers (DiTs) produce high-quality clips and support composition across shots through joint generation or conditioning on historical frames. However, repeatedly denoising each target clip with bidirectional attention incurs substantial computational cost and high generation latency, limiting rapid user feedback (Guo et al., 2025; Luo et al., 2026a). Causal autoregressive approaches instead generate sequentially and reuse historical KV states; combined with few-step sampling, they reduce latency while preserving long-range context and motion dynamics (Yin et al., 2025; Meng et al., 2026). Within this causal paradigm, global historical context and local within-shot context, together with shot-transition control, further support cross-shot consistency and interactive video generation (Lu et al., 2026; Luo et al., 2026b). These advances improve interactive generation, but precise reuse of historical content for each new shot remains unresolved. Two limitations arise when instructions combine content from different narrative moments, as shown in Figure 1. First, historical references can provide too much unrelated information or too little relevant evidence. Reusing compressed or sampled representations of preceding frames preserves whole-scene content, including unrequested contents (Zhang et al., 2025). Conversely, selecting a single source by global prompt similarity, as in the optional automatic router of Echo Forcing, can miss required components across shots (Wu et al., 2026a). For example, an instruction may bring a previously seen girl and boy into a library, with each character and the background originating from a different shot. Reusing all three source memories also imports unrelated content, whereas the most similar single shot may omit a required reference. Additional sources alone cannot resolve within-shot selection. Second, a uniform reference use policy struggles to reconcile two competing demands: preserving content supported by historical references and generating new content without such support. Strong historical conditioning can constrain unreferenced components, whereas indiscriminately weakening it can compromise the consistency of referenced characters and backgrounds. Infinity-RoPE demonstrates that manipulating temporal rotary position embedding (RoPE) coordinates can regulate temporal continuity and enable scene transitions (Yesiltepe et al., 2026). Inspired by this positional control, we adapt temporal RoPE settings to the availability of character and background references, balancing reference preservation with flexible generation of new content. Figure 1: Motivation for compositional memory reuse. Broad history reuse introduces irrelevant content (too much), while global prompt similarity retrieval can miss required references (too little). Weave Forcing selectively reuses relevant historical content (just right). To address these challenges, we propose Weave Forcing, a training-free framework for compositional memory reuse in interactive long video generation. Semantic slot routing uses a large language model (LLM) to first perform slot prompt splitting to identify the requested characters and background, and then apply cross-shot routing to associate each component with a historical source and its corresponding target slot prompt. Masked memory weaving uses contrastive slot attention (CSA) to distinguish target content from other content within the same historical memory, producing a semantic KV mask that selectively exposes relevant tokens. The references are combined in the cross-shot cache to guide the new shot. Coverage adaptive RoPE then adapts temporal positioning and reference retention to the available character and background support, while progressive memory decay smooths the transition toward the current shot’s own context. Together, these components enable selective recombination of historical content as user instructions evolve. Experiments on 100 six-shot prompts demonstrate the effectiveness of Weave Forcing, showing improved cross-shot subject and background consistency over the Self Forcing backbone while achieving competitive visual quality and text alignment. Our main contributions are summarized as follows: • We introduce Semantic Slot Routing, which decomposes each shot prompt into character and background slots and independently matches them to available historical source slots, enabling composition of references from different shots. • We propose Masked Memory Weaving, which constructs semantic KV masks through CSA, then combines the independently masked historical KV memories to retain requested content while suppressing unrelated characters and backgrounds. • We develop Coverage Adaptive RoPE, which adapts temporal offsets and reference retention to no, partial, or full character and background reference coverage, balancing historical guidance with the generation of newly requested content. 2 Related Work Early work on autoregressive video generation connected causal sequence modeling with diffusion through independent per-token noise levels, enabling generation beyond the training horizon (Chen et al., 2024). Subsequent approaches reduced sampling cost through distillation or adversarial post-training (Yin et al., 2025; Lin et al., 2026), while training on self-generated histories and rolling-window denoising helped reduce error accumulation (Huang et al., 2026; Liu et al., 2026). Complementary system-level optimizations improve the efficiency of streaming deployment and long video training (Feng et al., 2026; Chen et al., 2026). Interactive generation additionally requires responsiveness to changing instructions while maintaining temporal coherence. Chunk-wise prompting controls successive segments (Teng et al., 2025), while LongLive (Yang et al., 2025) refreshes cached states under new prompts to improve transitions. Other work adapts positional encoding and cache management for extended rollouts and scene transitions (Yesiltepe et al., 2026), or constrains teacher supervision to balance responsiveness with stability (Wu et al., 2026c). Extending interaction to multi-shot narratives further requires memory across shot boundaries: global and local contexts support streaming generation across shots (Luo et al., 2026b), and Echo Forcing (Wu et al., 2026a) uses hierarchical memory for scene recall. In contrast, our work addresses compositional recall by routing character and background slots to historical sources, masking unrelated KV content, and adapting memory positioning. 3 Preliminaries 3.1 Autoregressive Long Video Generation Autoregressive video diffusion generates a video progressively as latent blocks, conditioning each block on the text instruction and previously generated content (Yin et al., 2025; Huang et al., 2026). Let =bb=1Bx=\x_b\_b=1^B denote a sequence of B latent blocks and P its text prompt. The generation distribution factorizes as pθ(∣P)=∏b=1Bpθ(b∣P,<b).p_θ(x P)= _b=1^Bp_θ(x_b P,x_<b). (1) Each conditional distribution is realized by iteratively denoising (Ho et al., 2020) the current block with a causal DiT (Peebles & Xie, 2023). Historical key–value (KV) states are cached and reused during attention, avoiding repeated computation over completed blocks. Once a block is generated, its clean KV states are added to the cache to support subsequent generation. This incremental process enables streaming output and extension beyond a fixed clip duration. 3.2 3D RoPE Video DiTs use 3D RoPE to encode temporal and spatial token coordinates (Wan et al., 2025; Su et al., 2024). For a token at =(pt,ph,pw)p=(p^t,p^h,p^w), the coordinates specify its temporal, height, and width positions. Query and key channels are partitioned into three subspaces, each transformed by axis-specific two-dimensional rotations whose angles depend on the coordinate and channel frequency. The combined rotation, transformed queries and keys, and resulting relative position relation are expressed together as ⁡()=diag⁡(t(pt),h(ph),w(pw)),~=(),~′=(′)′,~⊤~′=⊤(′−)′. gatheredR(p)=diag\! (R_t(p^t),R_h(p^h),R_w(p^w) ),\\ q_p=R(p)q_p, k_p =R(p )k_p ,\\ q_p k_p =q_p R(p -p)k_p . gathered (2) Thus, the rotary contribution to the attention score depends on the relative spatiotemporal displacement. Temporal repositioning changes this relation without modifying spatial coordinates: applying an offset J to a historical token gives ^′=(pt′−J,ph′,pw′) p =(p^t -J,p^h ,p^w ). A uniform offset preserves the memory’s internal relative geometry while changing its temporal phase relative to the current queries. Such coordinate adjustments provide an inference-time means of controlling temporal relationships, as illustrated by Infinity-RoPE’s use of temporal repositioning and discontinuities for long video generation and scene transitions (Yesiltepe et al., 2026). 4 Method 4.1 Semantic Slot Routing Cross-shot reference reuse requires identifying the historical content needed by the current shot. Retrieval based on whole prompt similarity can miss characters and backgrounds originating from different shots. Selecting the top-k shots alone does not resolve this issue, since global similarity neither assigns sources to individual components nor specifies which content to retain. We therefore use an LLM to establish these correspondences: slot prompt splitting extracts character and background descriptions, and cross-shot routing matches each component to a historical source and its target slot prompt as shown in Figure 2 (left). This assignment supports composition across shots and identifies the relevant content within each source. Figure 2: Overview of Weave Forcing. Left: Semantic Slot Routing uses slot prompt splitting and cross-shot routing to get historical sources. Right: Masked Memory Weaving employs contrastive slot attention (CSA) to selectively retain relevant historical tokens and integrates the masked memoriesal tokens and integrates the masked memories. Slot Prompt Splitting. Let PtP_t denote the instruction for shot t. The LLM processes PtP_t independently and decomposes it into character and background prompts: (pt,cii=1Nt,pt,b)=SplitLLM⁡(Pt).(\p_t,c_i\_i=1^N_t,p_t,b)=Split_LLM(P_t). (3) Here, NtN_t is the number of visible characters, while cic_i and bb denote the i-th character and background, respectively. Each pt,cip_t,c_i retains the identity and appearance attributes of one character, while pt,bp_t,b describes the environment. This step yields separate character and background descriptions that serve as the semantic queries for subsequent source routing. Cross-Shot Routing. Let ℋt−1H_t-1 denote the descriptions and indices of preceding shots. The LLM independently matches each character or background slot to an eligible historical source: jt,s=RouteLLM⁡(pt,s,ℋt−1),s∈c1,…,cNt,b.j_t,s=Route_LLM(p_t,s,H_t-1), s∈\c_1,…,c_N_t,b\. (4) Here, jt,sj_t,s identifies the source shot for slot s. Only shots preceding t are eligible as sources. The LLM identifies the matching slot within that source, whose extracted description serves as the target slot prompt. The source index locates the compressed KV memory, while the target slot prompt specifies the character or background content to retain. If no suitable source exists, jt,s=∅j_t,s= and no reference is assigned. All detailed prompt instructions are provided in Appendix A. 4.2 Masked Memory Weaving Historical shots selected by the router can contain relevant characters or backgrounds alongside unrelated content. Directly reusing their complete memories may therefore introduce visual information that conflicts with the current instruction. We address this interference through slot-specific semantic KV masks that isolate the requested content within each retrieved memory. At the beginning of a new shot, the masked historical tokens form KV references that guide the first generated blocks, as illustrated in Figure 2 (right). Memory compression with query weighting. To make references available at shot boundaries, we compress the KV representations of each completed shot j into a compact memory. Let Q¯j Q_j be the mean query over the shot. For each spatial token u, we normalize its query–key affinities across candidate frames f and aggregate the K/V pairs using the weights: αj,f,u=softmaxf⁡(⟨Q¯j,Kj,f,u⟩),(K¯j,u,V¯j,u)=∑fαj,f,u(Kj,f,u,Vj,f,u). _j,f,u=softmax_f\! ( Q_j,K_j,f,u ), ( K_j,u, V_j,u)= _f _j,f,u(K_j,f,u,V_j,f,u). (5) The memory is indexed by shot for retrieval via jt,sj_t,s. Contrastive slot attention (CSA). A target prompt alone can activate both the requested component and its correlated surroundings. We use CSA to discriminate the target from competing content through attention map differencing. For each character or background slot s, let jt,sj_t,s denote its selected source shot. For slots with a valid source, we use two complementary textual conditions: the target prompt Pt,starP_t,s^tar specifies the content to retain, while the other prompt Pt,sothP_t,s^oth describes non-target content to suppress within the selected source (Appendix B.2). Both prompts are evaluated on the same historical memory, yielding aligned token-level attention maps: At,star A_t,s^tar =cross(Pt,star,K¯jt,s), =A_cross (P_t,s^tar, K_j_t,s ), (6) At,soth A_t,s^oth =cross(Pt,soth,K¯jt,s). =A_cross (P_t,s^oth, K_j_t,s ). (7) We subtract the other content response from the target response and threshold the resulting contrastive attention score to obtain an initial binary mask Mt,sM_t,s: Mt,s,u=[(At,s,utar−At,s,uoth)>τs].M_t,s,u=I\! [ (A_t,s,u^tar-A_t,s,u^oth )> _s ]. (8) Here, u indexes a token in the shared memory layout and τs _s is the threshold for slot s. The contrast is computed at inference time and introduces no contrastive training objective. For example, differencing the heatmaps for the girl and the other content highlights the girl while suppressing the library surroundings. The initial binary mask is reshaped to the spatial grid associated with the compressed memory and refined using morphological opening and closing (Soille et al., 1999) to obtain the final semantic KV mask M~t,s M_t,s, as follows: M~t,s=ℱmorph(Mt,s)=Close⁡(Open⁡(Mt,s)). M_t,s=F_morph(M_t,s)=Close\! (Open(M_t,s) ). (9) Opening removes isolated high-response regions, whereas closing fills small holes and connects nearby target regions. In combination, these operations reduce spurious activations and improve the spatial coherence of the mask, supporting more complete retention of the target content. The semantic KV mask M~t,s M_t,s is then mapped back to the memory token layout to control which historical KV tokens are accessible to the current block. Cache initialization with semantic masks. At shot t, we reset the context and load compressed memories from valid source indices, retaining independently masked views for slots sharing a source. The bounded context has four windows: sink initially stores historical references, then rolling anchors; compressed retains compressed within-shot history; recent holds the last completed block; and current holds the block being denoised. This layout treats cross-shot references as temporary contents, not a separate window. The current block supplies queries, while occupied window entries supply keys and values: Kall K_all =[Ksink;Kcomp;Krecent;Kcur], =[K_sink;K_comp;K_recent;K_cur], (10) Vall V_all =[Vsink;Vcomp;Vrecent;Vcur]. =[V_sink;V_comp;V_recent;V_cur]. In the Causal DiT, self-attention (Vaswani et al., 2017) integrates these states and gates each loaded cross-shot view with its slot-specific mask M~t,s M_t,s. The semantic bias is zero for all current-shot tokens, including subsequent sink entries, rather than permanently masking the sink window. For loaded reference tokens: Ot=Softmax⁡(QtKall⊤d+Bsem)Vall,Bsem(q,s,u)=0,M~t,s,u=1,−∞,M~t,s,u=0. gatheredO_t=Softmax\! ( Q_tK_all d+B_sem )V_all,\\ B_sem(q,s,u)= cases0,& M_t,s,u=1,\\ -∞,& M_t,s,u=0. cases gathered (11) This masked attention supplies relevant cross-shot references to guide generation as the current shot develops its own context. Figure 3: Coverage adaptive RoPE and cache evolution. Top: no coverage. Middle: partial coverage. Bottom: full coverage. 4.3 Coverage Adaptive RoPE As the current shot accumulates context, historical guidance must be regulated. Following the transient reference reuse of Echo Forcing (Wu et al., 2026a), we remove loaded cross-shot references after a prescribed number of completed blocks, rather than upon reaching the window capacity: one block for no coverage and five for partial or full coverage in our configuration. This eviction frees space for within-shot history without removing the sink window itself. The sink, compressed, and recent windows maintain rolling anchors, compressed history, and recent context, respectively, while the current window holds the block being denoised. Their updates support the transition from historical guidance to the current shot’s own context. The similarity-adaptive decay rule is detailed in Appendix B.5. However, a uniform retention policy is insufficient because referenced and unreferenced content require different positional treatment. We therefore distinguish no coverage, partial coverage, and full coverage according to character and background reference availability, with corresponding temporal RoPE configurations and retention policies (Figure 4.2). No coverage. When both character and background reference sets are empty, we use the preceding shot’s complete compressed memory as a transient style reference without semantic masking. We assign a temporal RoPE offset of 4545 to separate it from the current queries and remove it after the first block. Subsequent generation relies on the Intra-Shot Cache, preserving stylistic continuity while limiting unrelated historical influence. Partial coverage. When only one semantic category has references, we insert the corresponding masked memories into the Cross-Shot Cache. Since uneven reference support across semantic components may cause drift, we use a temporal RoPE offset of 4545 to increase its positional separation from the current queries. The reference then decays progressively until eviction, preserving guidance for the supported component while allowing missing content to develop from the current prompt. Full coverage. When both semantic categories have references, we use their masked memories with similarity-dependent temporal RoPE offsets. We set the offset within [1,6][1,6] based on the similarity between the reference and current-shot prompt embeddings. This close positional coupling facilitates the joint reuse of character and background references, while progressive decay before eviction supports the transition to the current shot’s own context. 5 Experiments Dataset. Using Gemini 2.5 Pro, we construct 100 test narratives by generating concise prompts and expanding each into six detailed shot prompts. Each shot lasts 10 seconds, yielding 60-second videos. Appendix A.1 provides the expansion template. Implementation Details. We build Weave Forcing upon Self Forcing (Huang et al., 2026) with the Wan2.1-T2V-1.3B backbone (Wan et al., 2025) and use Gemini 2.5 Pro (Google, 2025) for slot prompt decomposition and causal cross-shot reference routing. We generate videos at a resolution of 832×480832× 480 pixels with random seed 00 and configured denoising timesteps [1000,750,500,250][1000,750,500,250]. The model generates three latent frames per autoregressive block, with a total KV cache capacity of 21 latent frames, allocated to the sink, compressed, recent, and current windows with capacities of 12, 3, 3, and 3 latent frames, respectively. All experiments are conducted on an NVIDIA RTX PRO 6000 GPU with 96 GB of memory. Evaluation Metrics. We evaluate subject and background consistency at both intra-shot and inter-shot levels. We sample frames separately for intra-shot and inter-shot evaluation. We then use YOLO (Redmon et al., 2016) for subject localization and SAM (Kirillov et al., 2023) for segmentation, separating subject regions. DINOv2 (Oquab et al., 2023) features are then extracted from the corresponding regions and compared using cosine similarity to measure visual consistency. We use VBench-Long (Huang et al., 2024; Huang et al., 2025) to assess aesthetic and imaging quality on each complete video. For text alignment, we apply the VBench ViCLIP metric (Wang et al., 2024) to each shot and its corresponding short raw prompt (Appendix C). These metrics jointly characterize visual quality, prompt alignment, and consistency across temporal scales. Table 1: Quantitative comparison with existing methods. ↑ indicates that higher is better. Sub., Bg., and Cons. denote subject, background, and consistency, respectively. Best results are in bold, and second-best results are underlined. Method Intra-shot Cons. Inter-shot Cons. Aesthetic Quality↑ Imaging Quality↑ Text Alignment↑ Sub.↑ Bg.↑ Sub.↑ Bg.↑ Bidirectional models EchoShot (Wang et al., 2026) 0.776 0.950 0.338 0.652 0.594 0.687 0.231 CineTrans (Wu et al., 2026b) 0.845 0.983 0.354 0.652 0.609 0.712 0.227 Autoregressive models Self Forcing (Huang et al., 2026) 0.702 0.878 0.405 0.648 0.590 0.682 0.160 LongLive (Yang et al., 2025) 0.758 0.894 0.489 0.769 0.596 0.734 0.182 Rolling Forcing (Liu et al., 2026) 0.814 0.941 0.435 0.671 0.598 0.748 0.239 Infinity-RoPE (Yesiltepe et al., 2026) 0.805 0.903 0.495 0.735 0.589 0.732 0.209 Reward Forcing (Lu et al., 2026) 0.719 0.845 0.472 0.749 0.590 0.750 0.176 ShotStream (Luo et al., 2026b) 0.802 0.885 0.478 0.662 0.595 0.705 0.218 Echo Forcing (Wu et al., 2026a) 0.799 0.935 0.431 0.708 0.605 0.731 0.227 Ours 0.771 0.923 0.513 0.811 0.611 0.750 0.237 5.1 Quantitative Results Table 1 compares Weave Forcing with representative bidirectional and autoregressive baselines. Bidirectional models exhibit stronger intra-shot consistency, likely benefiting from access to the full shot context when modeling dependencies among frames during generation. Although Weave Forcing does not lead in intra-shot consistency, it achieves the highest inter-shot subject and background consistency among the compared methods, demonstrating its strength in preserving both characters and backgrounds across shot boundaries. This supports the effectiveness of selectively reusing cache components relevant to the current instruction, rather than relying on a single historical source or uniformly sampled historical frames. Alongside these cross-shot consistency gains, it maintains competitive video quality, including imaging quality and text alignment. We attribute this in part to our coverage adaptive RoPE strategy, which adjusts the use and progressive decay of historical references according to the available character and background support, helping mitigate visual drift and unstable transitions. Overall, the quantitative results demonstrate the effectiveness of selective cross-shot memory reuse in preserving historical content while maintaining competitive visual quality and alignment with the current prompt. Figure 4: Qualitative comparison across six sequential shots. Red circles (○ ) mark prompt mismatches, and yellow boxes (□ ) mark inconsistencies with earlier shots. 5.2 Qualitative Results Figure 4 shows a representative interactive generation example. The first three shots establish individual characters, while later prompts require switching scenes and combining different characters. Self Forcing does not explicitly support free shot-level switching; in this example, its generated frames are dark, making their visual content difficult to discern. Methods such as LongLive can generate content that clearly deviates from the prompt, as marked by the red circles. We further observe that methods such as EchoShot and Rolling Forcing may match the requested scene but still alter referenced details such as clothing materials or trouser colors, as highlighted by the yellow boxes. This suggests that they may rely more on prompt-level matching than on faithful reuse of earlier visual content. In contrast, Weave Forcing better preserves these appearance details when composing characters drawn from different historical shots. This behavior is consistent with semantic slot routing and masked memory reuse, which provide relevant visual references while limiting interference from unrelated content. Additional qualitative results and analyses are provided in Appendix D and Appendix E. 5.3 Ablation Study To verify the contribution of each component, we conduct ablation studies over cache fusion, reference routing, semantic isolation, cross-shot guidance duration, and intra-shot history retention. Table 2: Cross-shot cache fusion strategy. Align. denotes alignment. Best results are in bold. Variant Intra-shot↑ Inter-shot↑ Text Align.↑ Sub. Bg. Sub. Bg. Concat 0.778 0.929 0.493 0.809 0.224 Weighted 0.793 0.933 0.414 0.687 0.228 Ours 0.771 0.923 0.513 0.811 0.237 Figure 5: Qualitative comparison of weighted fusion, concatenation, and Ours. Cross-shot cache fusion strategy. Table 5 and Figure 5 compare cache fusion strategies quantitatively and qualitatively. Ours (semantic KV mask) outperforms direct concatenation and weighted fusion in cross-shot consistency and text alignment. The qualitative example shows that semantic masks reduce background mixing and better preserve reference details across shots when combining memories from different historical sources, supporting the effectiveness of our semantic masks. Individual components. Table 3: Component ablations. Align. denotes alignment. Best results are in bold. Variant Intra-shot↑ Inter-shot↑ Text Align.↑ Sub. Bg. Sub. Bg. Reference routing w/o LLM 0.790 0.938 0.442 0.734 0.230 Semantic separation w/o separation 0.777 0.927 0.501 0.778 0.220 Memory decay w/o decay 0.765 0.919 0.471 0.780 0.219 Intra-shot compression w/o compression 0.765 0.916 0.475 0.766 0.224 Ours 0.771 0.923 0.513 0.811 0.237 Table 3 ablates reference routing, semantic isolation, cross-shot guidance duration, and intra-shot history retention (implementation details in Appendix B). Each component supports stable long-form generation. Without LLM-assisted routing, cross-shot preservation declines because global prompt similarity can miss required characters or backgrounds. Removing semantic masking reduces inter-shot consistency and text alignment, supporting separate isolation even for components from the same source frame. Without progressive decay, historical references persist excessively, hindering the transition to current-shot context. Removing intra-shot compression discards compact history, weakening both local consistency and cross-shot preservation. The full method achieves the best inter-shot consistency and text alignment, highlighting the need to jointly assess intra-shot and cross-shot criteria. 6 Conclusion We presented Weave Forcing, a training-free framework for compositional memory reuse in interactive long video generation. Semantic slot routing uses an LLM to select historical sources for individual characters and backgrounds, while CSA constructs semantic KV masks to combine relevant memory content with reduced interference from unrelated components. Coverage adaptive RoPE adjusts temporal offsets and reference retention to the available support, balancing historical continuity with new content generation as memory decay shifts guidance toward the current shot’s own context. Experiments demonstrate improved inter-shot subject consistency while maintaining competitive visual quality and text alignment. These findings suggest that effective memory reuse for interactive video generation should prioritize the relevance of individual semantic components to the current instruction, rather than treating historical scenes as indivisible visual references. AI use statement In this work, we used generative AI tools for writing assistance and research execution. Specifically, ChatGPT was used to improve the clarity, grammar, and overall presentation of the manuscript. Gemini 2.5 Pro was used to assist the development and implementation of the semantic slot routing component, as well as to generate prompts for constructing the synthetic test datasets used in our experiments. All AI-assisted outputs were carefully reviewed, verified, and revised by the authors. We take responsibility for the final content of this work. Reproducibility statement To facilitate reproducibility, Section 4 describes the proposed method, while Appendices A, B, and C provide the prompt templates, implementation details, and evaluation procedures, respectively. These details provide the information needed to reproduce our work. References Chefer et al. (2023) Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM transactions on Graphics (TOG), 42(4):1–10, 2023. Chen et al. (2024) Boyuan Chen, Diego Martí Monsó, Yilun Du, Max Simchowitz, Russ Tedrake, and Vincent Sitzmann. Diffusion forcing: Next-token prediction meets full-sequence diffusion. Advances in Neural Information Processing Systems, 37:24081–24125, 2024. Chen et al. (2026) Yukang Chen, Luozhou Wang, Wei Huang, Shuai Yang, Bohan Zhang, Yicheng Xiao, Ruihang Chu, Weian Mao, Qixin Hu, Shaoteng Liu, et al. Longlive-2.0: An nvfp4 parallel infrastructure for long video generation. arXiv preprint arXiv:2605.18739, 2026. Cheng et al. (2022) Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 1280–1289. IEEE, 2022. Deng et al. (2025) Haoge Deng, Ting Pan, Haiwen Diao, Zhengxiong Luo, Yufeng Cui, Huchuan Lu, Shiguang Shan, Yonggang Qi, and Xinlong Wang. Autoregressive video generation without vector quantization. In International Conference on Learning Representations, volume 2025, p. 44730–44745, 2025. Dosovitskiy (2020) Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. Feng et al. (2026) Tianrui Feng, Zhi Li, Shuo Yang, Haocheng Xi, Muyang Li, Xiuyu Li, Lvmin Zhang, Keting Yang, Kelly Peng, Song Han, et al. Streamdiffusionv2: A streaming system for dynamic and interactive video generation. Proceedings of Machine Learning and Systems, 8:1656–1669, 2026. Google (2025) Google. Gemini 2.5 Pro model card, June 2025. URL https://modelcards.withgoogle.com/assets/documents/gemini-2.5-pro.pdf. Guo et al. (2025) Yuwei Guo, Ceyuan Yang, Ziyan Yang, Zhibei Ma, Zhijie Lin, Zhenheng Yang, Dahua Lin, and Lu Jiang. Long context tuning for video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 17281–17291, 2025. Ho et al. (2020) Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020. Huang et al. (2026) Xun Huang, Zhengqi Li, Guande He, Mingyuan Zhou, and Eli Shechtman. Self forcing: Bridging the train-test gap in autoregressive video diffusion. Advances in Neural Information Processing Systems, 38:167283–167308, 2026. Huang et al. (2024) Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 21807–21818. IEEE, 2024. Huang et al. (2025) Ziqi Huang, Fan Zhang, Xiaojie Xu, Yinan He, Jiashuo Yu, Ziyue Dong, Qianli Ma, Nattapol Chanpaisit, Chenyang Si, Yuming Jiang, et al. Vbench++: Comprehensive and versatile benchmark suite for video generative models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. Jocher & Qiu (2024) Glenn Jocher and Jing Qiu. Ultralytics YOLO11, 2024. URL https://github.com/ultralytics/ultralytics. Software, version 11.0.0. Kirillov et al. (2023) Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In 2023 IEEE/CVF international conference on computer vision (ICCV), p. 3992–4003. IEEE, 2023. Kong et al. (2024) Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. Li et al. (2024) Yuhong Li, Yingbing Huang, Bowen Yang, Bharat Venkitesh, Acyr Locatelli, Hanchen Ye, Tianle Cai, Patrick Lewis, and Deming Chen. Snapkv: Llm knows what you are looking for before generation. Advances in Neural Information Processing Systems, 37:22947–22970, 2024. Lin et al. (2023) Han Lin, Abhay Zala, Jaemin Cho, and Mohit Bansal. Videodirectorgpt: Consistent multi-scene video generation via llm-guided planning. arXiv preprint arXiv:2309.15091, 2023. Lin et al. (2026) Shanchuan Lin, Ceyuan Yang, Hao He, Jianwen Jiang, Yuxi Ren, Xin Xia, Yang Zhao, Xuefeng Xiao, and Lu Jiang. Autoregressive adversarial post-training for real-time interactive video generation. Advances in Neural Information Processing Systems, 38:41061–41086, 2026. Liu et al. (2026) Kunhao Liu, Wenbo Hu, Jiale Xu, Ying Shan, and Shijian Lu. Rolling forcing: Autoregressive long video diffusion in real time. In International Conference on Learning Representations, volume 2026, p. 91177–91196, 2026. Lu et al. (2026) Yunhong Lu, Yanhong Zeng, Haobo Li, Hao Ouyang, Qiuyu Wang, Ka Leong Cheng, Jiapeng Zhu, Hengyuan Cao, Zhipeng Zhang, Xing Zhu, et al. Reward forcing: Efficient streaming video generation with rewarded distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 34385–34397, 2026. Luo et al. (2026a) Xiangyang Luo, Qingyu Li, Xiaokun Liu, Wenyu Qin, Miao Yang, Meng Wang, Pengfei Wan, Di Zhang, Kun Gai, and Shao-Lun Huang. Filmweaver: Weaving consistent multi-shot videos with cache-guided autoregressive diffusion. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, p. 7689–7697, 2026a. Luo et al. (2026b) Yawen Luo, Xiaoyu Shi, Junhao Zhuang, Yutian Chen, Quande Liu, Xintao Wang, Pengfei Wan, and Tianfan Xue. Shotstream: Streaming multi-shot video generation for interactive storytelling. In European Conference on Computer Vision, p. 193–209. Springer, 2026b. Meng et al. (2026) Yihao Meng, Zichen Liu, Hao Ouyang, Qiuyu Wang, Ka Leong Cheng, Yue Yu, Hanlin Wang, Haobo Li, Jiapeng Zhu, Yanhong Zeng, et al. Causalcine: Real-time autoregressive generation for multi-shot video narratives. arXiv preprint arXiv:2605.12496, 2026. Oquab et al. (2023) Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. Peebles & Xie (2023) William Peebles and Saining Xie. Scalable diffusion models with transformers. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), p. 4172–4182. IEEE, 2023. Redmon et al. (2016) Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 779–788, 2016. Rombach et al. (2022) Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In 2022 IEEE/CVF conference on computer vision and pattern recognition (CVPR), p. 10674–10685. ieee, 2022. Schütze et al. (2008) Hinrich Schütze, Christopher D Manning, and Prabhakar Raghavan. Introduction to information retrieval, volume 39. Cambridge University Press Cambridge, 2008. Soille et al. (1999) Pierre Soille et al. Morphological image analysis: principles and applications, volume 2. Springer, 1999. Su et al. (2024) Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. Tang et al. (2023) Raphael Tang, Linqing Liu, Akshat Pandey, Zhiying Jiang, Gefei Yang, Karun Kumar, Pontus Stenetorp, Jimmy Lin, and Ferhan Türe. What the daam: Interpreting stable diffusion using cross attention. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 5644–5659, 2023. Teng et al. (2025) Hansi Teng, Hongyu Jia, Lei Sun, Lingzhi Li, Maolin Li, Mingqiu Tang, Shuai Han, Tianning Zhang, WQ Zhang, Weifeng Luo, et al. Magi-1: Autoregressive video generation at scale. arXiv preprint arXiv:2505.13211, 2025. Vaswani et al. (2017) Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. Wan et al. (2025) Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025. Wang et al. (2026) Jiahao Wang, Hualian Sheng, Sijia Cai, Weizhan Zhang, Caixia Yan, Yachuang Feng, Bing Deng, and Jieping Ye. Echoshot: Multi-shot portrait video generation. Advances in Neural Information Processing Systems, 38:22058–22090, 2026. Wang et al. (2024) Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, et al. Internvid: A large-scale video-text dataset for multimodal understanding and generation. In International Conference on Learning Representations, volume 2024, p. 42055–42079, 2024. Wu et al. (2026a) Mingqiang Wu, Weilun Feng, Zhefeng Zhang, Haotong Qin, Yuqi Li, Guoxin Fan, Xiaokun Liu, Zhulin An, Libo Huang, Yongjun Xu, et al. Echo-forcing: A scene memory framework for interactive long video generation. arXiv preprint arXiv:2605.16003, 2026a. Wu et al. (2026b) Xiaoxue Wu, Bingjie Gao, Yu Qiao, Yaohui Wang, and Xinyuan Chen. Cinetrans: Learning to generate videos with cinematic transitions via masked diffusion models. In International Conference on Learning Representations, volume 2026, p. 60198–60230, 2026b. Wu et al. (2026c) Yuheng Wu, Xiangbo Gao, Tianhao Chen, Xinghao Chen, Qing Yin, Zhengzhong Tu, and Dongman Lee. Delta forcing: Trust region steering for interactive autoregressive video generation. arXiv preprint arXiv:2605.14382, 2026c. Yang et al. (2025) Shuai Yang, Wei Huang, Ruihang Chu, Yicheng Xiao, Yuyang Zhao, Xianbang Wang, Muyang Li, Enze Xie, Yingcong Chen, Yao Lu, et al. Longlive: Real-time interactive long video generation. arXiv preprint arXiv:2509.22622, 2025. Yesiltepe et al. (2026) Hidir Yesiltepe, Tuna Meral, Adil Kaan Akan, Kaan Oktay, and Pinar Yanardag. Infinity-rope: Action-controllable infinite video generation emerges from autoregressive self-rollout. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 40256–40265, 2026. Yin et al. (2025) Tianwei Yin, Qiang Zhang, Richard Zhang, William T Freeman, Fredo Durand, Eli Shechtman, and Xun Huang. From slow bidirectional to fast autoregressive video diffusion models. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 22963–22974. IEEE, 2025. Zhang & Sennrich (2019) Biao Zhang and Rico Sennrich. Root mean square layer normalization. Advances in neural information processing systems, 32, 2019. Zhang et al. (2025) Kaiwen Zhang, Liming Jiang, Angtian Wang, Jacob Zhiyuan Fang, Tiancheng Zhi, Qing Yan, Hao Kang, Xin Lu, and Xingang Pan. Storymem: Multi-shot long video storytelling with memory. arXiv preprint arXiv:2512.19539, 2025. Zhang et al. (2023) Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, et al. H2o: Heavy-hitter oracle for efficient generative inference of large language models. Advances in neural information processing systems, 36:34661–34710, 2023. Appendix Contents Appendix A LLM Prompts and Computational Cost All language-model operations described in this section—multi-shot prompt expansion, semantic-slot extraction, and cross-shot source routing—are performed with Gemini 2.5 Pro (Google, 2025). The following templates specify the corresponding inputs, constraints, and structured outputs. A.1 Multi-Shot Prompt Expansion Summary. The shot planner converts one raw narrative into chronological, directly usable video-generation prompts. It computes ⌈T/10⌉ T/10 shots for a video of length T seconds (six 10-second shots in our 60-second setting), preserves the order of events, and requires complete repeated descriptions of recurring characters and locations. The implementation uses the following template with runtime fields shown in braces. _now:NeΠ _now:NeΠYou are a cinematic multi-shot video prompt planner. Start from the raw prompt _now:NeΠand expand it into a short sequence of complete, directly usable _now:NeΠvideo-generation prompts. Stay close to the raw prompt: preserve its main _now:NeΠcharacters, setting, visual style, and central action. Add only the minimal _now:NeΠdetails needed for visual continuity between shots. Keep the story simple and _now:NeΠeasy to depict. Do not introduce complicated plots, multiple sub-events, _now:NeΠunnecessary characters, dramatic twists, or long chains of actions. Do not _now:NeΠchange the identity or defining appearance of an established character. Keep _now:NeΠevery shot visually concrete, temporally consistent, and clear, natural, and _now:NeΠgrammatically correct English. Return JSON only. _now:NeΠ _now:NeΠ# TASK _now:NeΠTransform the raw prompt into a chronological sequence of complete _now:NeΠvideo-generation scene prompts. _now:NeΠ _now:NeΠ# RAW PROMPT _now:NeΠ–raw˙prompt˝ _now:NeΠ _now:NeΠ# SHOT PLAN _now:NeΠ- Split the total duration of –duration˙seconds˝ seconds into –shot˙count˝ _now:NeΠ chronological scene shots. _now:NeΠ- Use these exact shot durations: –shot˙durations˝. _now:NeΠ- Use scene1, scene2, … in chronological order. _now:NeΠ- Plan shots strictly in the order expressed in the RAW PROMPT; do not reorder, _now:NeΠ reverse, or otherwise change events, actions, or descriptions. _now:NeΠ- Each shot depicts one simple scene: first describe the background, then the _now:NeΠ visible character(s) and their one simple action or stable state. _now:NeΠ- Do not use complex action chains, rapid camera changes, or separate actions, _now:NeΠ poses, gestures, reactions, gaze directions, or motion states for different _now:NeΠ visible characters. Multiple characters share the same action or state. _now:NeΠ _now:NeΠ# FOR BACKGROUNDS _now:NeΠIn every visible setting, give a full concrete description of the location and _now:NeΠits architecture or natural environment. Repeat it independently in every _now:NeΠshot. On recurrence, begin with ’back in’ and repeat the full supported _now:NeΠdescription; never write only ’the same place’ or ’the same’. _now:NeΠ _now:NeΠ# FOR CHARACTERS _now:NeΠOn first appearance, give all supported stable details, including approximate _now:NeΠage, gender presentation, clothing colours, and visual traits. On recurrence, _now:NeΠbegin with ’the same’ and repeat the full supported appearance; never write _now:NeΠonly ’the same person’, ’the same woman’, ’the same man’, or ’the same’. Do _now:NeΠnot invent unsupported details or change clothing, accessories, hairstyle, or _now:NeΠother defining details. Do not add a second person when the raw prompt has no _now:NeΠtwo-person shot. A character is fully present or absent: when absent, do not _now:NeΠmention its body part, clothing, shadow, reflection, or off-screen implication. _now:NeΠUse one mostly static state per shot. In a multi-character shot, describe each _now:NeΠindividual’s appearance in one uninterrupted phrase without commas and do not _now:NeΠinterleave attributes across characters. _now:NeΠ _now:NeΠ# GENERAL RULES _now:NeΠ- Do not remove supported visual details merely to shorten a shot. _now:NeΠ- Keep every shot clear, natural, and directly usable by a video generator. _now:NeΠ- Do not add explanations, commentary, or markdown fences. _now:NeΠ _now:NeΠ# OUTPUT FORMAT _now:NeΠReturn exactly this JSON shape, with no additional keys or text: _now:NeΠ–”shots”:[–”shot˙id”:”scene1”,”duration˙seconds”:10, _now:NeΠ”prompt”:”complete shot prompt”˝]˝ _now:NeΠEach object in ”shots” contains only shot˙id, duration˙seconds, and prompt. A.2 Semantic Slot Extraction Summary. For every expanded scene independently, the semantic planner separates the description into a variable number of character slots and one background slot. Distinct visible subjects are represented by character1, character2, and so forth; a scene containing only one subject does not introduce an empty second-character field. This conservative decomposition retains static identity and appearance cues while excluding actions and poses. No cross-scene coreference resolution is performed during slot extraction. _now:NeΠ _now:NeΠYou are a video prompt semantic slot extractor. For each scene prompt, produce _now:NeΠone slot for each distinct visible character, named character1, character2, _now:NeΠand so forth, together with one background slot. Return JSON only and process _now:NeΠeach scene independently. Use only information explicitly present in that _now:NeΠscene prompt and retain its original wording whenever possible. _now:NeΠ _now:NeΠ# TASK _now:NeΠSeparate each scene prompt into character1, character2, … and background. _now:NeΠ _now:NeΠ# SCENES _now:NeΠscene1: –scene1˙prompt˝ _now:NeΠscene2: –scene2˙prompt˝ _now:NeΠ… _now:NeΠ _now:NeΠ# SLOT EXTRACTION RULES _now:NeΠ- Process each scene independently; do not use other scenes. _now:NeΠ- Use only information explicit in the current prompt; do not infer, resolve _now:NeΠ references, or invent details. _now:NeΠ- Create exactly one characterN slot for each distinct visible person or _now:NeΠ creature. Number slots by order of first mention in the scene. _now:NeΠ- Do not output characterN when the corresponding character is absent. _now:NeΠ- Each characterN slot contains only that subject’s identity, appearance, _now:NeΠ clothing, accessories, and static visual traits; exclude expression, pose, _now:NeΠ movement, behavior, actions, and attributes belonging to another subject. _now:NeΠ- background: only place, environment, props, lighting, atmosphere, and camera _now:NeΠ setting as compact noun phrases; exclude people, verbs, and prepositions. _now:NeΠ- Keep original wording and visual details whenever possible. _now:NeΠ _now:NeΠ# OUTPUT FORMAT _now:NeΠReturn exactly this JSON shape and no additional text: _now:NeΠ–”scenes”:[–”scene˙id”:”scene1”, ”character1”:”first subject”, _now:NeΠ”character2”:”second subject”, ”background”:”only environment”˝]˝ _now:NeΠOmit character2 and later character fields when those subjects are absent. A.3 Cross-Shot Source Routing Summary. Given the extracted slots, a second LLM call selects earlier source slots for each recall target. Each target character is matched independently and may be routed to a differently indexed character slot in the source scene; for example, scene2.character1 may refer to scene1.character2. Empty lists explicitly represent unavailable references. _now:NeΠ _now:NeΠYou select source scenes for semantic recall in video generation. For each _now:NeΠrecall target scene, assign suitable source scenes from earlier scenes only. _now:NeΠReturn JSON only. _now:NeΠ _now:NeΠ# TASK _now:NeΠFor each recall target, select earlier source slots for every target character _now:NeΠand for the background. Use the target scene and all eligible earlier scenes _now:NeΠas the basis for the decision. _now:NeΠ _now:NeΠ# INPUTS _now:NeΠ## SCENES _now:NeΠscene1: character1=–scene1˙character1˝; character2=–scene1˙character2˝; _now:NeΠ background=–scene1˙background˝ _now:NeΠscene2: character1=–scene2˙character1˝; background=–scene2˙background˝ _now:NeΠ… _now:NeΠ _now:NeΠ## RECALL TARGETS _now:NeΠTarget scene scene–t˝: _now:NeΠ character1=–target˙character1˝ _now:NeΠ character2=–target˙character2˝ _now:NeΠ … _now:NeΠ background=–target˙background˝ _now:NeΠOnly scenes earlier than the target may be used as sources. _now:NeΠ _now:NeΠ# IMMEDIATE PREVIOUS SCENE CONTINUITY RULE _now:NeΠFirst compare each target with its immediately preceding scene. If character _now:NeΠidentity and appearance, character count, background location, and main visual _now:NeΠelements are almost identical, choose that preceding scene for both _now:NeΠthe corresponding characterN˙sources and background˙sources. Do not choose an _now:NeΠolder scene unless the preceding scene fails one of these conditions. _now:NeΠ _now:NeΠ# CHARACTER SOURCE REQUIREMENTS _now:NeΠ1. Determine the number of distinct people or creatures in the target text. _now:NeΠ2. Prefer sources containing only one person or creature. Match identity, _now:NeΠ appearance, clothing, and other static visual features. _now:NeΠ3. If several candidates match, select the earliest valid source. _now:NeΠ4. Match every target character to the corresponding characterN slot in the _now:NeΠ selected source scene. Source and target slot indices need not be equal. _now:NeΠ5. A character source may be empty; never force a non-genuine match. _now:NeΠ _now:NeΠ# BACKGROUND SOURCE REQUIREMENTS _now:NeΠ1. Select at most one background source for each target. _now:NeΠ2. Match the location and key visual elements, including important objects, _now:NeΠ architecture, environment, lighting, and atmosphere. _now:NeΠ3. background˙sources may be empty; never force a non-genuine match. _now:NeΠ _now:NeΠ# OUTPUT FORMAT _now:NeΠReturn exactly this JSON shape and no additional text: _now:NeΠ–”sink˙assignments”:–”¡target˙scene˙id¿”: _now:NeΠ–”character1˙sources”:[–”scene˙id”:”¡source˙scene˙id¿”, _now:NeΠ”source˙role”:”character2”˝], _now:NeΠ”background˙sources”:[–”scene˙id”:”¡source˙scene˙id¿”, _now:NeΠ”source˙role”:”background”˝]˝ A.4 LLM Inference Cost Each LLM request uses approximately 4,000 tokens in total, with an average reported throughput of about 130 tokens per second (TPS). Dividing this token budget by the throughput gives a coarse latency estimate of 4,000/130≈314,000/130≈ 31 seconds, or roughly 30 seconds per call. This estimate is indicative rather than a measured end-to-end latency: TPS characterizes token generation, whereas the total token budget also includes input tokens. Actual request latency therefore depends on the input–output token split and API serving overhead. Appendix B Method Details B.1 Shot Memory Storage Each completed shot is registered as one entry in a bounded scene pool. An entry retains its chronological shot index, original prompt, normalized prompt feature, and—when semantic routing is enabled—separate character and background prompt features. The pool is ordered by generation time and discards its oldest entry when its capacity is exceeded. These lightweight fields locate historical shots; the visual content is retained separately as K/V memory. The visual component is stored per transformer layer as a compact pair of key and value tensors (Vaswani et al., 2017). Let Ks,f,u(ℓ)K^( )_s,f,u and Vs,f,u(ℓ)V^( )_s,f,u denote the key and value at layer ℓ , candidate latent frame f, and spatial token u of shot s. A latent frame contains U=1560U=1560 spatial tokens in the current implementation, and each token has 12 attention heads of dimension 128. We use the score_weighted mode to aggregate the sampled candidate frames into one K/V representative per spatial position and layer. The tensors are detached from the denoising graph and moved to CPU storage; they are copied back to the active GPU cache only when that shot is selected as a reference. In LLM-semantic mode, the pipeline additionally stores cross-attention query summaries, later used to derive slot-specific masks over the saved K/V tokens. A saved shot therefore comprises retrieval metadata, compact layer-wise visual K/V memory, and optional semantic-query statistics; decoded RGB frames are not used as cross-shot memory. For every layer, the cache accumulates the mean query Q¯(ℓ) Q^( ) and a bounded candidate K/V buffer while the shot is denoised. The affinity of candidate frame f at spatial position u is averaged over attention heads, as,f,u(ℓ)=1H∑h=1H⟨Q¯h(ℓ),Ks,f,u,h(ℓ)⟩,Q¯(ℓ)=1Nq∑i=1NqQi(ℓ).a^( )_s,f,u= 1H _h=1^H Q^( )_h,K^( )_s,f,u,h , Q^( )= 1N_q _i=1^N_qQ^( )_i. (12) These scores are normalized over candidate frames, αs,f,u(ℓ)=softmaxf⁡(as,f,u(ℓ))α^( )_s,f,u=softmax_f(a^( )_s,f,u), and used to form weighted K/V summaries (Vaswani et al., 2017). To retain a stable reference, the implementation further blends the weighted summary with the first sampled candidate frame: K^s,u(ℓ) K^( )_s,u =λKs,0,u(ℓ)+(1−λ)∑f∈sαs,f,u(ℓ)Ks,f,u(ℓ), =λ K^( )_s,0,u+(1-λ) _f _sα^( )_s,f,uK^( )_s,f,u, (13) V^s,u(ℓ) V^( )_s,u =λVs,0,u(ℓ)+(1−λ)∑f∈sαs,f,u(ℓ)Vs,f,u(ℓ), =λ V^( )_s,0,u+(1-λ) _f _sα^( )_s,f,uV^( )_s,f,u, (14) where λ is the configured base-frame weight. The same weights are applied to keys and their co-indexed values, preserving their correspondence in the compact shot memory ℳs=(K^s,u(ℓ),V^s,u(ℓ))ℓ=1,u=1L,UM_s=\( K^( )_s,u, V^( )_s,u)\_ =1,u=1^L,U. B.2 Routing Shot Prompts At each shot boundary, the semantic planner uses the LLM-produced (Google, 2025) slot and routing results to determine both the historical memory source and the text used to localize that source. For a target role rir_i in shot t, the routing result specifies a source shot s<ts<t and a matched source role πt(ri) _t(r_i). Importantly, the correspondence is not constrained to preserve the numerical slot index: a subject represented by character1 in the target may correspond to character2 in the source. The target localization prompt is therefore read directly from the matched source slot, Ptar(t,i)=Ps,πt(ri),(s,πt(ri))=LLMRoute⁡(t,ri),P_tar^(t,i)=P_s, _t(r_i), (s, _t(r_i))=LLMRoute(t,r_i), (15) rather than being reconstructed from the full target-shot description. This preserves the identity and appearance wording attached to the actual historical subject whose memory is being retrieved. The complementary prompt is constructed from the semantic slots previously extracted by the LLM from the source shot’s own description (Appendix A.2). Let sS_s denote its non-empty character and background slots. Non-target slots are the remaining slots after excluding the matched source slot πt(ri) _t(r_i); their descriptions are concatenated without an additional LLM call: Poth(t,i)=Concat⁡(Ps,rr∈s∖πt(ri)).P_oth^(t,i)=Concat\! (\P_s,r\_r _s \ _t(r_i)\ ). (16) For a character target, these slots comprise the background and all other characters; for a background target, they comprise all characters. Thus, PtarP_tar describes only the routed component, whereas PothP_oth summarizes the remaining non-target content in the same historical memory. Both prompts are then encoded to extract attention responses over the same source memory, using the pretrained query and key projections and their normalization parameters from the DiT’s cross-attention layers. Specifically, stored source queries are paired with keys projected from each prompt’s text features; scaled query–key scores are averaged over text tokens, attention heads, and layers to obtain the two spatial response maps used by CSA. No additional attention parameters are trained. B.3 Compressed Window This component computes a compact set of latent-token representations from the completed within-shot history. More precisely, the implementation operates on the K/V projections of transformer latent tokens (Peebles & Xie, 2023), rather than averaging or re-encoding VAE latents (Rombach et al., 2022). Given hidden latent tokens H(ℓ)H^( ), the stored candidate representation at layer ℓ is (Zhang & Sennrich, 2019) K(ℓ)=RMSNorm⁡(H(ℓ)WK(ℓ)),V(ℓ)=H(ℓ)WV(ℓ).K^( )=RMSNorm(H^( )W_K^( )), V^( )=H^( )W_V^( ). (17) At denoising timestep zero, each newly generated latent frame contributes its K/V tokens, together with their original temporal and spatial indices, to a bounded history buffer. The module simultaneously accumulates a calibrated mean query Q¯(ℓ) Q^( ) over the tokens generated before compression. The compressed representation is obtained by selecting informative historical K/V tokens (Zhang et al., 2023; Li et al., 2024), not by synthesizing a new latent vector. Each historical token u receives the query-guided score ρu(ℓ)=Fuseh,ω⁡[Re⁡(Q¯h(ℓ)Ku,h(ℓ)¯eiωΔtu)+Amp⁡(Q¯h(ℓ),Ku,h(ℓ))],ρ^( )_u=Fuse_h,ω\! [Re\! ( Q^( )_h K^( )_u,he^iω t_u )+Amp\! ( Q^( )_h,K^( )_u,h ) ], (18) where Δtu t_u is the token’s temporal displacement and FuseFuse aggregates frequency components and attention heads. The first term measures phase-aware query–key compatibility under temporal RoPE (Su et al., 2024); the second is an amplitude-compensation term, optionally reduced when the current query distribution drifts from its calibration. The compressed representation consists of the highest-scoring eligible history tokens, ℐcomp=TopKC⁡(ρu(ℓ)u∈ℋ(ℓ)),(K,V)comp=(K,V)ℐcomp.I_comp=TopK_C (\ρ^( )_u\_u ^( ) ), (K,V)_comp=(K,V)_I_comp. (19) Here ℋ(ℓ)H^( ) is the eligible K/V history and C is the fixed compressed-token budget. The cache limits repeated selection of the same historical token. It copies the selected K/V pairs together with their original temporal and spatial indices, so compression preserves the exact values and positions of selected latent-token representations rather than forming a blurred temporal average. B.4 Semantic Separation Semantic separation determines whether character and background content drawn from the same historical frame are treated as one shared memory source or as independent semantic components. Without separation, a source frame selected for both roles is inserted once and its K/V tokens remain jointly available. Consequently, subject appearance and scene layout are recalled together, which is appropriate when the complete historical composition should be preserved but may also transfer irrelevant background or subject attributes. With separation enabled, the shared frame is represented by role-specific views of the same stored K/V tensors. Let Ms,rM_s,r be the semantic mask for role r in source shot s. The routed memory for that role is K~s,r=Ms,r⊙Ks,V~s,r=Ms,r⊙Vs, K_s,r=M_s,r K_s, V_s,r=M_s,r V_s, (20) or, equivalently in attention implementation, excluded tokens receive a strong negative attention bias (Cheng et al., 2022). Character and background may therefore share the same source frame while contributing through different spatial token subsets. The distinction is solely whether co-located semantic content is kept coupled or separated; it does not change the selected historical frame or create new visual features. B.5 Similarity-Adaptive Decay The decay mechanism gradually reduces the influence of retained historical tokens using a rate bounded by the configured minimum and maximum values. For each spatial token u, the implementation compares the key of the first frame in the current block, kucurk^cur_u, with the corresponding historical key kuhistk^hist_u. Their cosine distance is (Schütze et al., 2008) du=1−⟨kucur,kuhist⟩∥kucur∥2∥kuhist∥2,d_u=1- k^cur_u,k^hist_u k^cur_u _2 k^hist_u _2, (21) where the comparison is averaged over attention heads. The distances are min–max normalized across the spatial tokens, d^u=du−dmindmax−dmin, d_u= d_u-d_ d_ -d_ , (22) with d^u=0 d_u=0 when the distance range is numerically degenerate. Given the configured bounds γmin _ and γmax _ , the token-wise decay rate is obtained by linear interpolation, γu=γmax−(γmax−γmin)d^u,γu∈[γmin,γmax]. _u= _ -( _ - _ ) d_u, _u∈[ _ , _ ]. (23) The historical token weight is then updated multiplicatively as wu←γuwuw_u← _uw_u. Hence, tokens that remain similar to the current content receive a rate near γmax _ and decay slowly, whereas tokens with a larger feature discrepancy receive a rate near γmin _ and are suppressed more rapidly. Appendix C Evaluation Metric Implementation C.1 Shot-Aware Video Partitioning All metrics are evaluated on the native temporal extent of each generated shot. In particular, we do not resample, pad, or truncate all methods to a common number of frames per shot. Let Vn,s(m)V_n,s^(m) be shot s of video n generated by method m, and let Ln,s(m)L_n,s^(m) denote its observed frame count after extraction. Every frame-based metric operates on the available index set ℱn,s(m)(δ)=20+kδ∣20+kδ<Ln,s(m),k∈ℕ0,F_n,s^(m)(δ)=\20+kδ 20+kδ<L_n,s^(m),\ k _0\, (24) so that its samples reflect the actual duration produced by that method. The first 20 frames are skipped to avoid the shot-boundary transient. When shots are extracted from a concatenated sequence, the default extractor additionally removes the first three transition frames from shots 2–6. These operations only define shot boundaries; they do not equalize shot lengths. For cross-shot evaluation, the extractor forms separate subject and background groups from the recorded routing relations and concatenates the native shot files in temporal order. A shot may belong to multiple groups when it reuses different sources. Consequently, a grouped video represents one semantic continuity chain rather than an artificial fixed-length clip. We use a stride of 60 frames for within-shot evaluation and 160 frames for grouped cross-shot evaluation. A score is reported only when the corresponding sequence contains the samples required by that metric. C.2 Subject and Background Consistency We use a common region-based procedure for subject and background consistency, both within shots and across the grouped cross-shot videos described above. For every sampled RGB frame, YOLO11n (Jocher & Qiu, 2024) detects all COCO person instances at confidence threshold 0.250.25. Each detection box is supplied to SAM-ViT-B (Kirillov et al., 2023), which produces one highest-IoU binary mask per person. We encode each masked person crop and the complementary background image using the [CLS][CLS] (Dosovitskiy, 2020) embedding of DINOv2 ViT-B/14 (Oquab et al., 2023), followed by ℓ2 _2 normalization. Person crops replace non-person pixels inside their bounding rectangles with a neutral value; for background features, the union of all person masks is filled by the median colour of the non-person pixels. Thus, the same encoder and masking protocol are used for all four settings (within-shot subject/background and cross-shot subject/background); only the input sequence changes. Let t,rz_t,r be the normalized DINOv2 feature of person r in sampled frame t, and let tb_t be the corresponding background feature. For a sequence whose first sampled frame contains reference persons ℛR, subject consistency at a later frame t is csub(t)=1|ℛ|∑r∈ℛmaxj∈t⁡0,r⊤t,j,Csub=1||∑t∈csub(t),c_sub(t)= 1|R| _r _j _tz_0,r z_t,j, C_sub= 1|T| _t c_sub(t), (25) where tP_t is the set of detected and segmented persons at t and T contains all later sampled frames. A later frame without a valid person mask contributes zero. If the reference frame has no valid person, the sequence receives zero; sequences with no later sample are undefined and are excluded from averaging. For cross-shot subject groups, the evaluator selects one reference person per group and prevents duplicate protagonists across groups of the same video: if its DINOv2 similarity to a previously selected reference exceeds 0.50.5, that group receives zero. Background consistency is the mean pairwise cosine similarity across all valid sampled background features, Cbg=2K⁡(K−1)∑1≤i<j≤Ki⊤j,C_bg= 2K(K-1) _1≤ i<j≤ Kb_i b_j, (26) where K is the number of valid sampled frames. A sequence with fewer than two valid background frames is excluded. Within-shot scores are first averaged over the available shots of a video; cross-shot scores are first averaged over the available semantic groups. Dataset scores are then unweighted means over valid videos. Runtime failures and undefined sequences are excluded, whereas the missing-person cases specified above remain penalized by zero. C.3 Video Quality and Text Alignment We assess perceptual video quality with two VBench-Long dimensions (Huang et al., 2024; Huang et al., 2025): aesthetic quality and imaging quality. Each method’s complete, unpartitioned output video is passed directly to VBench-Long; no manually extracted shot is supplied to these two metrics. VBench-Long applies its standard long video processing and returns scores for its internally produced long clips. For a video with QnQ_n such clips and dimension d, we use Qn,d=1Qn∑q=1Qnqn,d(q),Qd=1N∑n=1NQn,d.Q_n,d= 1Q_n _q=1^Q_nq_n,d^(q), Q_d= 1N _n=1^NQ_n,d. (27) The reported imaging-quality value is rescaled from VBench-Long’s percentage output to [0,1][0,1] before aggregation. Both quantities are therefore whole-video measures, while retaining VBench-Long’s native temporal treatment. For text alignment, the ViCLIP text encoder (Wang et al., 2024) has a restricted effective text length, making concatenation of all six detailed generation prompts unsuitable. We instead evaluate each extracted shot Vn,s(m)V_n,s^(m) independently with the ordinary VBench ViCLIP overall_consistency implementation (Huang et al., 2024) and its corresponding short raw prompt pn,srawp^raw_n,s (stored as viclip_prompt). If an,sa_n,s denotes the resulting ViCLIP score, the video- and dataset-level scores are An=1S∑s=1San,s,A=1N∑n=1NAn,S=6.A_n= 1S _s=1^Sa_n,s, A= 1N _n=1^NA_n, S=6. (28) This shot-wise design aligns each temporal segment with the instruction that actually specifies it, while avoiding text truncation. All reported metrics are computed from the same JSON-listed videos; result files retain per-video, per-shot, and per-group scores to permit independent verification. Figure 6: Supplementary qualitative comparison in a single-subject narrative. All methods receive the same shot-level instructions. The sequence tests preservation of the established subject while the setting and event progress change across shots. Appendix D Additional Qualitative Results D.1 Single-Subject Comparison The main paper presents a multi-person composition case, where later shots must jointly recover several identities. Figure 6 provides the complementary single-subject setting. It compares all methods under the same sequence of instructions, isolating whether a method can retain one established identity while the scene and action evolve. This setting separates subject persistence from the additional difficulty of binding attributes across multiple people. As shown in Figure 6, Weave Forcing achieves favorable qualitative results in this controlled setting. Across successive shots, it preserves the subject’s defining appearance while accommodating the requested changes in scene context and event progression. Compared with the displayed alternatives, the resulting sequence exhibits fewer conspicuous identity changes and maintains a more coherent correspondence to the previously established subject. This observation indicates that the proposed memory routing is also effective when consistency depends on a single recalled identity rather than a multi-subject composition. Figure 7 reports additional results for Weave Forcing without a baseline row, allowing the individual recall patterns to be inspected at a larger scale. The examples cover (i) scene recall, in which a previously established environment is requested again; (i) subject recall, in which an earlier character must be preserved under a new context; and (i) narrative development, in which the action or event changes while the relevant visual history remains available. Together, these cases show the range of reference-availability and story-progression conditions represented by the evaluation prompts. The qualitative results show that Weave Forcing performs consistently across these heterogeneous scenarios. It can recover previously observed scene structure, retain subject-specific appearance after a contextual change, and support continued event development without simply reproducing an earlier shot. Although the visual requirements differ across the examples, the method generally preserves the requested historical content while allowing newly specified elements to evolve, suggesting that its routing mechanism is not restricted to a single type of cross-shot dependency. Figure 7: Additional Weave Forcing results across diverse interactive scenarios. The examples include scene recall, subject recall, and event development, illustrating preservation and controlled evolution of visual content across sequential shots. D.2 Intra-Shot Consistency In addition to cross-shot recall, Figure 8 shows the within-shot behavior of the method. The examples isolate short-range temporal consistency while a subject and its local environment remain visible within one shot. They complement the cross-shot cases above by demonstrating that the routed historical memory is integrated with, rather than substituted for, the cache accumulated during the current shot. For visualization, we uniformly sample three frames from each shot and compare their subject appearance, local background, and spatial organization. The sampled frames show satisfactory intra-shot stability, with the principal visual attributes remaining coherent over the duration of the shot. We attribute this behavior to two complementary components: the compressed window retains informative latent K/V tokens from the recent generation history, while the Self Forcing backbone (Huang et al., 2026) provides a comparatively stable basis for short-range autoregressive generation. Their combination helps limit local appearance drift even when routed cross-shot memory is simultaneously introduced. Figure 8: Supplementary within-shot consistency results. Three frames are uniformly sampled from each shot to show the persistence of subject appearance and local scene content over its temporal extent. Appendix E Limitations and Failure Analysis Figure 9 illustrates a failure mode in a crowded multi-subject shot. Although the generated scene retains both people, their appearance attributes are not cleanly separated: the lower part of the rear male subject is rendered with skirt-like clothing, while the female subject inherits the male subject’s hat. This is an attribute-entanglement (Tang et al., 2023) error rather than a complete failure to render either subject. The failure has two coupled causes. First, an imperfect semantic mask can retain spatial tokens from an adjacent subject, leaving insufficient separation between their respective historical appearances. Second, the new prompt asks the generator to attend to many entities and visual attributes simultaneously. The resulting competition for attention can weaken the slot-specific routing signal and cause the model to bind an attribute to the wrong person (Chefer et al., 2023). These examples suggest that more discriminative instance-level masks and explicit multi-subject binding are needed when several visually salient subjects must be recalled and modified in the same shot. Figure 9: Failure case of multi-subject attribute binding. In the generated result, the rear male subject exhibits skirt-like lower clothing, while the female subject inherits the male subject’s hat. Incomplete semantic masking and competition among the many attributes requested by the new prompt can cause this cross-subject feature mixing.

Instant research discovery