Paper deep dive
Form and Void: Entangled Composition through an Autonomous AI Agent
Shiwen Wang, Jian Yang, Xu Wang, Xincan Wang, Weiming Dong
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 10/2/2026, 8:23:49 AM
Summary
The paper introduces FaV-A (Form and Void Agent), a multimodal agent designed to generate positive-negative space compositions. Unlike direct zero-shot prompting of MLLMs, FaV-A uses a staged workflow: semantic planning to create a base object, spatial parsing to analyze shape and structure, and guided synthesis to entangle the positive and negative spaces along a shared boundary. Experimental results show FaV-A outperforms baselines like Gemini, GPT, and Claude in producing visually coherent and semantically aligned compositions.
Entities (14)
Relation Signals (10)
Weiming Dong → affiliatedwith → Institute of Automation, Chinese Academy of Sciences
confidence 95% · Weiming Dong 1,4... 4 MAIS, Institute of Automation, Chinese Academy of Sciences
Jian Yang → affiliatedwith → Renmin University of China
confidence 95% · Jian Yang 2... 2 Renmin University of China
Xu Wang → affiliatedwith → University of Chinese Academy of Sciences
confidence 95% · Xu Wang 1... 1 School of AI, University of Chinese Academy of Sciences
Xincan Wang → affiliatedwith → Shanghai Theatre Academy
confidence 95% · Xincan Wang 3... 3 Shanghai Theatre Academy
Weiming Dong → affiliatedwith → University of Chinese Academy of Sciences
confidence 95% · Weiming Dong 1... 1 School of AI, University of Chinese Academy of Sciences
Shiwen Wang → affiliatedwith → University of Chinese Academy of Sciences
confidence 95% · Shiwen Wang 1... 1 School of AI, University of Chinese Academy of Sciences
FaV-A → usesworkflow → Staged Generation
confidence 95% · FaV-A follows a progressive workflow: it first generates a base object, then analyzes its shape and spatial structure... and finally produces compositional instructions for the final image generation stage.
FaV-A → outperforms →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Positive and negative space is a fundamental principle in visual composition, supporting visually coherent forms and layered semantic relationships. Generating such compositions is challenging because it requires coordinated control over two semantic concepts that share a common boundary. Although recent text-to-image models and multimodal large language models (MLLMs) have achieved strong performance in image generation and visual understanding, positive-negative space generation remains difficult, particularly under direct single-pass prompting. In this work, we present the \textbf{F}orm \textbf{a}nd \textbf{V}oid \textbf{A}gent (\textbf{FaV-A}), a multimodal agent designed for staged positive-negative space generation. FaV-A follows a progressive workflow: it first generates a base object, then analyzes its shape and spatial structure to identify candidate negative-space semantics, and finally produces compositional instructions for the final image generation stage. Experimental results and ablation analyses suggest that FaV-A provides a more effective framework than direct zero-shot MLLM baselines for producing visually coherent and semantically aligned positive-negative space compositions.
Tags
Links
- Source: https://arxiv.org/abs/2610.02045v1
- Canonical: https://arxiv.org/abs/2610.02045v1
Trouble viewing inline? Open PDF directly →
Full Text
40,102 characters extracted from source content.
Expand or collapse full text
Form and Void: Entangled Composition through an Autonomous AI Agent Shiwen Wang 1,∗ Jian Yang 2,∗ Xu Wang 1 Xincan Wang 3,† Weiming Dong 1,4,† 1 School of AI, University of Chinese Academy of Sciences 2 Renmin University of China 3 Shanghai Theatre Academy 4 MAIS, Institute of Automation, Chinese Academy of Sciences wangxincan@sta.edu.cn weiming.dong@ia.ac.cn Topic: Music and Nature “Shattered Grief – The Whisper of the Guardian of the Withered Tree” Design: A broken plate serves as the visual focal point, with the shape of a tree emerging from its cracks. The fractured design symbolizes concern about human destruction, raising awareness for environmental protection. “Hug of Cats – The Gaze of Harmony and Coexistence” Design: The first sight is the white outline of a cat. As the gaze deepens, the outline of another cat gradually emerges in the negative space, with both animals gazing forward together. This intertwined design embodies the friendship and harmony between the animals. “Bird Sing in the Notes – The Echos of Music and Nature” Design: A note serves as the silhouette, while a bird, whose wings, beak are delicately embodied through and space of the note. This symbolizes the melodic beauty of birdsong, and conveying the harmonious unity of nature and art. Topic: Cats and Friendship Topic: Destruction and Protection Figure 1. Given a topic, our FaV-A autonomously generates an artwork accompanied by a caption and a detailed description, both grounded in the entangled duality “Form and Void”. The first two lines within the shadow box present the artwork’s caption. The “Design” section offers a description of the composition and the layered meanings embedded within the piece. Abstract Positive and negative space is a fundamental principle in visual composition, supporting visually coherent forms and layered semantic relationships. Generating such composi- tions is challenging because it requires coordinated control over two semantic concepts that share a common bound- ary. Although recent text-to-image models and multimodal large language models (MLLMs) have achieved strong per- formance in image generation and visual understanding, positive-negative space generation remains difficult, partic- ularly under direct single-pass prompting. In this work, we present the Form and Void Agent (FaV-A), a multimodal agent designed for staged positive-negative space genera- tion. FaV-A follows a progressive workflow: it first gen- erates a base object, then analyzes its shape and spatial structure to identify candidate negative-space semantics, ∗ Equal contribution. † Corresponding authors. and finally produces compositional instructions for the fi- nal image generation stage. Experimental results and abla- tion analyses suggest that FaV-A provides a more effective framework than direct zero-shot MLLM baselines for pro- ducing visually coherent and semantically aligned positive- negative space compositions. 1. Introduction The interplay between form and void lies at the core of vi- sual composition and is classically expressed through pos- itive and negative space. In art and design, negative space refers to the empty regions around and between the sub- jects of an image [4, 16, 17]. This principle plays an im- portant role across visual art, photography [2], graphic de- sign [14], and sculpture [25]. Designing with negative space requires careful coordination of visual elements, as figure- ground relationships affect not only compositional structure 1 arXiv:2610.02045v1 [cs.CV] 1 Oct 2026 but also semantic interpretation. From a computational per- spective, generating such compositions is challenging be- cause the model must simultaneously represent an explicit foreground form and an implicit secondary shape induced by the surrounding space. Recent text-to-image models have shown strong capa- bilities in visual generation. However, existing generation pipelines often rely on explicit spatial controls, such as masks, layouts, or other manually specified conditions, to manage figure-ground relationships. With the emergence of multimodal large language models (MLLMs) [5, 19, 22], visual understanding has improved substantially. Never- theless, a gap remains between understanding and gener- ation: although MLLMs can often recognize and describe positive-negative space compositions, generating such ef- fects in a zero-shot, single-pass setting remains difficult. In our preliminary observations, directly prompting an end- to-end model to synthesize two overlapping semantic con- cepts with a shared boundary often results in structurally unstable compositions or semantically ambiguous outputs. These observations motivate a staged generation framework in which structure discovery and semantic refinement are separated into successive steps. To address this gap, we adopt a staged generation per- spective motivated by common design workflows.In positive-negative space composition, a plausible workflow is to first establish a primary subject, then inspect its con- tour and spatial structure, and finally refine the image by introducing secondary semantics in the negative space. Fol- lowing this intuition, we model the task as a sequence of in- termediate steps in which earlier outputs provide structural guidance for later stages. This decomposition reduces the burden of requiring a model to resolve shared-boundary ge- ometry and dual semantic alignment in a single generation step. In addition, recent work on MLLM-based agents sug- gests that explicit intermediate reasoning can be beneficial for complex multimodal generation tasks [6–8]. Motivated by this perspective, we present the Form and Void Agent (FaV-A), a multimodal agent for positive- negative space generation from abstract topics. Rather than relying on one-shot synthesis, FaV-A adopts a staged work- flow. Given a topic, it first generates a base object informed by design-oriented priors. It then analyzes the base image to identify shape and spatial cues that could support a sec- ondary interpretation in the negative space. Finally, these observations are translated into compositional instructions for the final image generation step. As shown in Figure 1, this staged process yields more coherent positive-negative space compositions in our examples. These results suggest that multimodal agents are a promising direction for design- oriented generation tasks that require coordinated spatial and semantic control. To summarize, this work makes three contributions. First, we formulate positive-negative space generation as a challenging multimodal design problem that requires co- ordinated spatial and semantic reasoning under shared- boundary constraints. Second, we propose the Form and Void Agent (FaV-A), a staged multimodal agent that de- composes this task into three successive phases: base- object generation, image-aware structural analysis, and instruction-guided final synthesis. Third, through qualita- tive comparisons, user evaluation, and ablation analysis, we show that this staged framework produces more structurally coherent and semantically aligned positive-negative space compositions than direct zero-shot MLLM prompting. 2. Related Work 2.1. Text to Image Generation Recent diffusion-based methods have shown that generative models can produce images with multiple valid interpreta- tions under predefined transformations or viewing setups, such as multi-view optical illusions or physically arranged illusion patterns [3, 11]. However, these methods typically target fixed transformation rules or illusion mechanisms, rather than discovering a figure-ground composition from an abstract semantic topic. Controllable image generation with explicit structural conditions is also the research focus. Methods such as Con- trolNet, T2I-Adapter, and GLIGEN enable text-to-image models to follow masks, edges, depth maps, keypoints, bounding boxes, or other grounding signals [15, 18, 27]. These approaches provide strong spatial control once the control signal is available, but they generally assume that the relevant structure has already been specified. In con- trast, our setting begins with an abstract topic and requires the system to infer both a suitable primary form and a com- patible negative-space interpretation without manually de- signed spatial templates. 2.2. MLLM Agents for Visual Generation Recent MLLM-based agents have demonstrated strong ca- pabilities in visually grounded planning and execution [13]. In creative and design-oriented settings, PosterLLaVA for- mulates poster layout generation with multimodal large lan- guage models, GenArtist uses an MLLM agent to coor- dinate image generation and editing through planning and verification, and BannerAgency applies multimodal agents to the creation of editable advertising banners [23, 24, 26]. GraphicBench further shows that graphic design remains challenging for language agents, especially when reasoning about spatial relationships and coordinating global depen- dencies across design decisions [12]. These works motivate the use of staged multimodal rea- soning for design tasks, but they primarily focus on layout planning, tool orchestration, or editable asset generation. 2 Our work instead targets positive-negative space synthesis, where the central challenge is to derive a shared-boundary figure-ground composition from an abstract topic through intermediate image analysis and compositional anchoring. 3. Method 3.1. Overall Architecture We propose the Form and Void Agent (FaV-A), a multi- modal agent for positive-negative space generation built on a large multimodal model, denoted by M. Given an ab- stract textual topic T , the goal of FaV-A is to generate a final composition I final together with a textual description D final that captures the intended figure-ground duality. As illustrated in Figure 2, unlike layout-controlled gen- eration pipelines that depend on manually specified masks, edges, or spatial layouts, FaV-A follows a staged and mask- free generation process. Instead of directly mapping T to the final image in a single step, the agent decomposes the task into three sequential stages. In Stage I, Topic- Driven Semantic Planning (M plan ), the agent converts the input topic into a base-object prompt and a high-level com- positional plan, guided by design priors P prior . In Stage I, Multi-Modal Spatial Parsing and In-Context Anchor- ing (M parse ), the agent generates a base image and ana- lyzes its shape and spatial structure to derive constraints for negative-space composition. In Stage I, Mask-Free En- tangled Synthesis (M syn ), the agent uses the base image and the derived structural cues to produce the final positive- negative space composition. This staged generation process can be written as: (I final ,D final ) =M syn I base , A ,(1) (I base , A) =M parse M plan (T ,P prior ) ,(2) where A denotes the spatial and compositional cues ex- tracted from the intermediate analysis stage. 3.2. Stage I: Semantic Planning Abstract topicsT usually lack explicit geometric and spatial specifications. The goal of Stage I is therefore to transform the input topic into structured textual guidance for later gen- eration stages. We provide the planning model with a sys- tem prompt containing design priorsP prior , which encour- age attention to figure-ground structure, negative space, and compositional balance. The prompts, design priors and de- tailed pipeline will be included in our Supplementary Ma- terial. Conditioned on T and P prior , the planning model out- puts: (t obj , c base ) =M plan (T ,P prior ),(3) where t obj denotes the Base Object Prompt for the primary subject, and c base denotes the Base Composition Blueprint describing the intended figure-ground arrangement. This stage converts an abstract topic into a more concrete se- mantic specification for subsequent image generation and analysis. 3.3. Stage I: Spatial Anchoring Given the intermediate plan from Stage I, a text-to-image module first generates a base image, I base =G T2I (t obj ).(4) This base image provides an initial foreground structure, but it does not yet encode the intended positive-negative space composition. To derive guidance for the final synthe- sis stage, FaV-A performs a multimodal analysis of I base . Specifically, the agent analyzes the base image along several dimensions, including shape attributes, spatial re- lationships, visual hierarchy, potential negative-space in- terpretations, symbolic associations, and implementation- related considerations. We denote the resulting intermediate analysis by S = Φ(I base ),(5) where S summarizes the structural and semantic observa- tions extracted from the base image. To further align this analysis with the target composition, FaV-A incorporates multimodal in-context learning (ICL). The agent is provided with N = 3 visual-textual exemplars, E =(I (i) exam ,D (i) exam ) N i=1 , together with the base composition blueprint c base and the parsed representation S. Based on these inputs, the parsing module generates a detailed compositional prompt: P anchor =M parse (I base , S, c base ,E ),(6) where the resulting prompt P anchor serves as a semantic and structural anchor for the final generation stage. Rather than specifying explicit spatial masks, it describes the in- tended contour relationships, negative-space interpretation, and compositional constraints in natural language. 3.4. Stage I: Guided Final Synthesis The final stage generates the target positive-negative space composition from the intermediate outputs of the previous stages. Rather than relying on manually specified bounding boxes, masks, or edge-based control signals, FaV-A condi- tions the synthesis process on both the base image and a detailed compositional prompt. Formally, the synthesis stage is defined as (I final ,D final ) =M syn (I base ,P anchor ),(7) 3 T2I Base Object Prompt Shape features Spatial relationship Visual hierarchy Design potential Symbolic significance Technical features Image Feature Analysis Base Image Design Examples Image Feature Analysis Base Composition (I&T)2 (I&T) Output Topic Stage I: Knowledge Distillation and Base Design Generation “A chess piece is used as the positive shape, and the player's face is used as the negative shape.” “One chess piece. Black and white.” “Chess and Player” Stage I: Base Image Generation and Multi-modal Understanding Design Knowledge Topic Analysis Stage I: Multi- modal Generation Base Composition Specific and Detail Composition Image Prompt Text Prompt : The Whole Node : Processed Node : Processing Node T2I : Text to Image (I&T) 2(I&T) : Image&Text to Image&Text Design Description Input Figure 2. Multistage design process of FaV-A to create positive-negative space compositions. where I base provides the initial visual structure andP anchor provides text-based guidance for the intended figure-ground composition. The output image I final is the final synthesized result, while D final is a textual description of the semantic relationship expressed in the composition. This stage integrates the intermediate structural analysis and the compositional prompt into the final generation pro- cess, enabling mask-free positive-negative space synthesis under multimodal conditioning. 4. Results and Discussions 4.1. Comparison with State-of-the-Art Baselines To evaluate the effectiveness of our multi-stage ar- chitecture, we compare FaV-A against direct prompt- ing baselines utilizing state-of-the-art closed-source multimodalmodels,specificallyGemini(Gemini 3.1 [21]),GPT (gpt-image-1.5 [20]),and Claude (Sonnet 4.6 [1]).We conducted all experiments us- ing gemini-3.1-flash [10] as the understanding (MLLM) model and gemini-3.1-flash-image [9] as the image generation model. In this setup, the baseline models are provided with the target abstract topic alongside explicit theoretical definitions and principles of positive- negative space design.Concretely, all closed-source baselines are prompted with the same instruction: Definition of Positive-Negative Space Art (Form and Void): Positive-negative space refers to a composition where the positive shape (the subject) and the neg- ative space (the area around it) are equally impor- tant and interdependent. In advanced design, they are “entangled”—meaning they share a common, single contour line. The boundary of the positive form must perfectly define the shape of another meaningful ele- ment in the negative space, creating a visual pun or dual meaning (figure-ground reversal). Based on the definition above, please autonomously design and generate a minimalist black-and-white art- work for the topic: “T opic”. In this way, the baselines are given access to the same high- level conceptual knowledge as FaV-A during its planning stage. Behavior and Limitations of Baseline Models. When provided with the augmented prompt (topic + design knowl- edge), the baseline models are generally able to capture the target concepts and produce high-contrast black-and-white images. Nevertheless, a clear gap remains between seman- tic understanding and positive-negative space execution. As shown in Figure 3, rather than forming two mutually de- fined shapes through a shared boundary, the baselines ex- hibit three representative failure patterns: • Semantic Drift and Weak Topic Grounding (Claude): Claude exhibits the weakest alignment with the intended topics. Although it occasionally produces clean black- and-white graphics, its outputs often fail to preserve the core semantic concepts specified in the prompt. This is- sue is particularly evident in Rose and Love and War and Peace, where few clearly recognizable topic-related ele- ments remain. Hence, Claude’s limitation is not only its inability to realize positive-negative space structure, but also its insufficient topic grounding. As a result, its out- puts lack both semantic fidelity and the dual-form cou- pling required for effective positive-negative space de- sign. • Topic-Consistent but Additive Composition (Gemini): Gemini generally preserves the target concepts and pro- duces recognizable topic-relevant objects. However, its dominant strategy is additive composition, typically by stacking, juxtaposing, or locally inserting one object into another. While semantically consistent, such lay- outs rarely achieve the figure-ground coupling central to positive-negative space art.Instead, the generated forms remain illustrative rather than structurally interde- pendent. Consequently, Gemini produces coherent topic- 4 Chess and PlayerRose and LoveWar and PeaceCats/Dogs and FriendshipLetters “CAT” and catLetters “DOG” and dog Gemini gemini-3.1 Claude Claude-sonnet-4.6 Ours gemini- 3.1-flash (a)(b)(c)(d)(e)(f) GPT gpt-image-1.5 Figure 3. Visual comparison between state-of-the-art baselines (Gemini, Claude, and GPT) and our FaV-A. Given the same topic and positive-negative space design knowledge, the baselines fail to achieve true topological entanglement: Claude defaults to semantic juxta- position, GPT relies on explicit spatial partitioning, and Gemini resorts to internal part-filling. Only FaV-A successfully entangles both subjects along a continuous shared boundary, achieving true bistable perception. driven graphics, but lacks the contour sharing, ambigu- ity, and formal tension characteristic of strong positive- negative space compositions. • Graphic Plausibility without True Figure-Ground Re- versal (GPT): GPT shows the strongest overall design sensibility among the direct-prompting baselines. Rela- tive to simple object juxtaposition, it more often yields compositions with clearer visual hierarchy, stronger sym- bolic organization, and better poster-like readability. However, these results are better characterized as literal graphic design than genuine positive-negative space de- sign. Multiple concepts are typically integrated through explicit depiction or layout-level composition, rather than being jointly defined by a shared contour across posi- tive and negative regions. Therefore, although GPT ap- proaches visually compelling conceptual design, it still falls short of the stricter figure-ground entanglement re- quired in positive-negative space art. Advantages of the FaV-A Architecture. In contrast, our proposed method successfully achieves structural entan- glement by reformulating the visual illusion as a progres- sive, constrained reasoning task. Rather than expecting an MLLM to perform zero-shot spatial alignment in a sin- gle generation step, FaV-A utilizes a base image to estab- lish an initial geometric foundation. By conducting multi- dimensional visual parsing on this base image, the agent synthesizes a precise semantic spatial anchor (P anchor ). This anchor explicitly maps the semantic requirements to specific topological regions (e.g., instructing the model to align a facial profile with the jagged external edge of a pri- mary object). Guided by these rigorous spatial constraints, the final synthesis stage aligns the conflicting visual pri- ors, forcing the positive and negative spaces to share a sin- gle, continuous contour. This methodology ensures the re- liable emergence of bistable perception, realizing the “one shape, dual meaning” principle that direct prompting base- lines cannot achieve. Table 1. Quantitative results of the ablation study based on human evaluation (N = 50). Metrics are evaluated on a 5-point Likert scale (higher is better). Our full FaV-A pipeline significantly out- performs all ablated variants across Topic Adherence (TA), Gestalt Quality (GQ), and overall Aesthetic Score (AS). Model VariantTA↑GQ↑AS↑ w/o Topic Analysis1.35± 0.423.12± 0.652.85± 0.70 w/o Detailed Base Prompt4.15± 0.502.90± 0.583.25± 0.62 w/o Image Feature Analysis3.80± 0.552.15± 0.602.40± 0.65 w/o Specific Composition3.92± 0.481.85± 0.522.10± 0.55 w/o Multi-modal Generation3.45± 0.601.60± 0.451.80± 0.48 FaV-A (Ours - Full)4.65± 0.354.58± 0.404.52± 0.38 5 (1) Chess and Player (2) Petting a Cat (a) (b)(c)(d)(e)(f) BaseImage BaseImage Results Results Figure 4. Ablation study on the “Chess and Player” (top) and “Petting a Cat” (bottom) topics. (a) Full FaV-A Pipeline, showcasing strong figure-ground interplay and shared boundaries; (b) w/o Topic Analysis, resulting in a shallow vertical juxtaposition; (c) w/o Detailed Base Prompt, introducing early semantic deviation (e.g., generating a dragonfly instead of a player); (d) w/o Image Feature Analysis, leading to a loss of structural fidelity and morphological drift; (e) w/o Specific Composition, causing visual confusion and semantic bleeding; (f) w/o Multi-modal Generation, leading to complete structural dissolution (e.g., the cat arbitrarily transforms into a dog). User Study To rigorously quantify these qualitative ob- servations, we conducted a user study (N = 50) to evalu- ate the ablated variants across three dimensions: Topic Ad- herence (TA), Gestalt Quality (GQ), and overall Aesthetic Score (AS) on a 5-point Likert scale. The participants were recruited from graduate students and researchers with prior experience in visual design, computer vision, or digital art, including a subset with formal design training. During eval- uation, all results were anonymized and presented in ran- domized order, such that the raters were unaware of which method produced each image. Each participant was asked to score the generated artworks independently based on the three criteria, following a unified instruction sheet describ- ing the meaning of each metric. This procedure was adopted to reduce presentation bias and improve the reliability of subjective assessment. As reported in Table 1, the full FaV-A pipeline signif- icantly outperforms all ablated variants. The quantitative results strongly align with our visual findings: notably, the removal of topic analysis causes a severe drop in TA (1.35), reflecting the model’s inability to follow the se- mantic prompt. Meanwhile, the absence of specific com- positional constraints (GQ: 1.85) and the bypass of multi- modal generation (GQ: 1.60) incur the heaviest penalties in Gestalt Quality, statistically confirming their indispensabil- ity in maintaining the structural illusion. 4.2. Ablation Study To validate the effectiveness of each functional module within the FaV-A pipeline, we conduct a set of ablation studies using the abstract topics “Chess and Player” and “Petting a Cat”. For each configuration, we remove or sim- plify a single component while keeping the remainder of the pipeline unchanged. Figure 4 visualizes the qualitative degradation in structural coherence and positive-negative space entanglement when individual stages are ablated. Topic Analysis. The primary role of this module is to de- compose an abstract concept into candidate positive and negative entities and establish their initial spatial relation- ship. For the ablation, we bypass the semantic planning and feed the raw topic directly to the subsequent stages. The absence of this preliminary logic causes the output to de- generate into a shallow juxtaposition of elements (see Fig- ure 4(b)). In the “Petting a Cat” example, the design merely stacks a hand silhouette and a cat silhouette vertically, lack- ing intentional interaction. In contrast, the full pipeline al- lows the hand to act as a negative shape that wraps around and gently strokes the cat’s head, demonstrating a much stronger figure-ground interplay (see Figure 4(a)). Base Object Prompt. This module translates the topic into 6 morphological instructions to synthesize a base image capa- ble of supporting negative space carving. For the ablation, we replace the LLM-derived instructions with a generic manual prompt: “Positive or negative shapes representing topic”. As observed in the upper example of Figure 4(c), this underspecified prompt introduces severe semantic devi- ation at the source—generating an irrelevant dragonfly-like shape due to an unstable interpretation of the word “player”. This error propagates through the pipeline, proving that pre- cise initial morphological constraints are essential to pre- vent downstream semantic drift. Image Feature Analysis. This stage grounds the spatial reasoning in the actual geometry (e.g., shape characteristics, visual hierarchy) of the generated base image. As illustrated in Figure 4(d), removing this multi-dimensional parsing dis- connects the refinement process from the original base ob- ject. Consequently, the final outputs suffer from a loss of structural fidelity: in the upper example, the player’s hair characteristics are noticeably altered; in the lower example, the cat’s pose drifts. This analysis is particularly crucial for a human-in-the-loop workflow, where the intermediate base form selected by a designer must remain morphologically stable during final generation. Specific Composition. The purpose of this component is to synthesize a localized spatial anchor (P anchor ) that dic- tates exactly how the contours should be sculpted. We ab- late this by substituting the detailed anchor with a naive instruction: “Transform this into a POSITIVE-NEGATIVE SPACE poster. CRITICAL: Use ONLY pure black and pure white...” As shown in Figure 4(e), without the detailed topo- logical constraints, the model exhibits morphological drift and weakened visual harmony. The positive and negative shapes become visually confused (e.g., overlapping messily in the chess example) or fail to preserve the intended object configuration (e.g., the number of cats inexplicably changes in the lower example). Multi-modal Generation. The final (I&T) to (I&T) stage enforces high-frequency edge preservation by inte- grating the accumulated visual and textual priors. By by- passing this dedicated reasoning pass and directly feed- ing the intermediate inputs to a standard image generator (Figure 4f), the system fails to consolidate the spatial con- straints. This results in structural dissolution, where the vi- sual illusion collapses entirely. Most notably, in the lower example, the cat is arbitrarily transformed into a dog, in- dicating that without this final reasoning mechanism, the shape priors established in earlier stages are completely ig- nored. 4.3. Discussions of Results by FaV-A Here we qualitatively evaluate the autonomous composi- tions generated by our FaV-A (see Figure 5). Rather than merely viewing these outputs as static artistic artifacts, we regard them as both exquisite artistic works and as empirical evidence of the MLLM’s capacity to execute complex zero- shot topological reasoning. In this discussion, we not only present the generated results but also provide in-depth anal- yses of the AI-generated design drafts that underpin these compositions. These results validate how the abstract du- ality of “Form and Void” is successfully grounded in pre- cise visual geometries through our multi-modal reasoning pipeline, thereby imbuing the artworks with profound inter- pretive depth. Zero-Shot Semantic-to-Spatial Entanglement Tradi- tional generative models typically struggle to map seman- tic intent to negative space, often defaulting to generating isolated foreground forms without explicit mask guidance. FaV-A, however, transforms spatial absence into a highly structured semantic entity. In Silent Duel: Between the Pieces and Their Masters (see Figure 5a (1)), the negative space surrounding the chess piece is precisely sculpted into a human profile in the latent space. Similarly, in Phantom of Peace: A Doomsday Mushroom Cloud Beneath the Olive Branch (see Figure 5a (3)), the mushroom cloud is syn- thesized entirely through the absence of surrounding tex- tures rather than explicit contours. This demonstrates the agent’s robust semantic-to-spatial entanglement, effectively translating the rich cognitive constraints of the semantic an- chor (P anchor ) into complex topological boundaries without human-engineered layouts. Structural Coherence in Multi-stable Perception A fundamental technical challenge in generating visual illu- sions is maintaining structural coherence across overlap- ping visual elements without causing morphological col- lapse or semantic bleeding. FaV-A overcomes this by en- forcing high-frequency edge preservation guided by its in- ternal cross-modal attention mechanisms. In Whispered Kiss Beneath the Rose (see Figure 5a (2)), a couple’s in- teraction subtly emerges from the intrinsic curves of the rose, effectively blending surface geometry and hidden sub- text. In Harmonious Gaze: Poetry of Cat and Cat/Dog in Silent Companionship (see Figure 5b (Left)), both the pri- mary silhouette and the embedded negative-space figures maintain high visual legibility. This confirms the agent’s capacity to simultaneously optimize two distinct visual pri- ors along a shared, continuous boundary, ensuring dual- semantics within a single structural contour. Cross-Domain Alignment and Gestalt Emergence We further explore the agent’s generalization capabilities in cross-domain scenarios involving typography.In the CAT/DOG typography series (see Figure 5b (Right)), text characters and animal silhouettes are fused into bistable configurations that resist immediate deterministic classifi- cation. For a standard vision model, textual forms and bi- 7 Design: The side profile of chess king— symbolizing power, strategy, and leadership, the side profile of a player—representing wisdom, contemplation, and engagement. The design creates a symmetric composition rich in metaphor and visual impact. Silent Duel: Between the Pieces and Their Masters Topic:ChessandPlayer Design: The positive part forms side faces of a couple sharing a kiss, while the central negative space subtly evokes the shape of a wine glass. It creates a dual- layered composition in which viewers can see both the lovers and their affections. Whispered Kiss Beneath the Rose Topic: Rose and Love Phantom of Peace: A Doomsday Mushroom Cloud Beneath the Olive Branch Design: Wheat ears, olive tree, dove of peace, and mushroom cloud—with the looming threat of war embodied by a mushroom cloud. It symbolizes the fragility of peace, aiming to provoke reflection on the value of peace and the consequences of war. Topic: War and Peace (1)(2)(3) (a) Common Design: The English word “CAT/DOG” serves as the object, while the negative space is cleverly used to shape the silhouette of a cat/dog. Common Design: A cat embraces a cat/dog as they gaze into the distance together, symbolizing friendship and harmonious coexistence. The cat appears as the figure, while the cat/dog emerges from the negative space, sharing the harmonious gaze. Harmonious Gaze – Poetry of Cat and Cat/Dog in Silent Companionship “DOG” and a dog / “CAT” and a cat Topic: Cats/Dogs and Friendship Topic: Letters “CAT/DOG” and cat/dog (b) Figure 5. Positive-negative space composition results autonomously created by FaV-A. (Top) Examples demonstrating zero-shot semantic- to-spatial entanglement and structural coherence in multi-stable perception. (Bottom) Typography series showcasing cross-domain align- ment and Gestalt emergence. ological contours typically belong to disparate latent distri- butions, making seamless integration difficult. Yet, FaV- A successfully navigates this dual constraint: it interprets textual forms as the rigid “ground” while dynamically embedding animal silhouettes into the enclosed negative spaces. This algorithmic emergence of Gestalt perception highlights the MLLM’s advanced compositional reason- ing, proving its proficiency in handling conflicting multi- objective spatial tasks. 5. Conclusion In this work, we introduced Form and Void Agent (FaV- A), a novel autonomous multi-modal system designed to synthesize complex positive-negative space compositions based on abstract textual topics. By formulating visual illu- sions as a sequence of high-level cognitive reasoning tasks, we bypass the heavy reliance on explicit spatial masks and structural templates prevalent in existing controlled diffu- sion models. Our proposed three-stage pipeline success- fully integrates topic-driven semantic planning, in-context spatial anchoring, and mask-free (I&T) to (I&T) syn- thesis. Experimental results demonstrate that FaV-A not only executes cross-modal visual alignment with high struc- tural coherence but also exhibits advanced zero-shot topo- logical reasoning capabilities, such as multi-stable percep- tion and cross-domain typographic integration. Ultimately, FaV-A highlights the immense potential of Large Multi- modal Models to serve as autonomous cognitive partners in high-level aesthetic design, paving the way for more so- phisticated, semantic-driven computational artistry. Acknowledgements This work was supported in part by the National Natural Science Foundation of China under no. 62572458, and in part by the Shanghai Pujiang Talent Program under no. 25PJC081. 8 References [1] Anthropic. Claude sonnet 4.6 system card, 2026. Anthropic. 4 [2] Geoffrey Batchen. Negative/Positive: A History of Photog- raphy. Routledge, 2020. 1 [3] Ryan Burgert, Xiang Li, Abe Leite, Kanchana Ranasinghe, and Michael S. Ryoo. Diffusion illusions: Hiding images in plain sight, 2023. 2 [4] Alessandra Cave. Shooting with Soul: 44 Photography Ex- ercises Exploring Life, Beauty and Self-Expression - From Film to Smartphones, Capture Images Using Cameras from Yesterday and Today. Quarry Books, Beverly, MA, 2013. 1 [5] Chaorui Deng, Deyao Zhu, Kunchang Li, Chenhui Gou, Feng Li, Zeyu Wang, Shu Zhong, Weihao Yu, Xiaonan Nie, Ziang Song, et al. Emerging properties in unified multimodal pretraining. arXiv preprint arXiv:2505.14683, 2025. 2 [6] Rongyao Fang, Chengqi Duan, Kun Wang, Linjiang Huang, Hao Li, Hao Tian, Shilin Yan, Weihao Yu, Xingyu Zeng, Jifeng Dai, Xihui Liu, and Hongsheng Li. GoT: Unleash- ing reasoning capability of MLLM for visual generation and editing. In The Thirty-ninth Annual Conference on Neural Information Processing Systems (NeurIPS), 2025. 2 [7] Weixi Feng, Wanrong Zhu, Tsu-Jui Fu, Varun Jampani, Ar- jun Reddy Akula, Xuehai He, S Basu, Xin Eric Wang, and William Yang Wang.LayoutGPT: Compositional visual planning and generation with large language models. In Thirty-seventh Conference on Neural Information Process- ing Systems (NeurIPS), 2023. [8] Jiahui Gao, Renjie Pi, Jipeng Zhang, Jiacheng Ye, Wanjun Zhong, Yufei Wang, Lanqing HONG, Jianhua Han, Hang Xu, Zhenguo Li, and Lingpeng Kong. G-LLaVA: Solving geometric problem with multi-modal large language model. In The Thirteenth International Conference on Learning Representations (ICLR), 2025. 2 [9] Gemini Team. Gemini 3.1 flash image model card, 2026. Google DeepMind. 4 [10] Gemini Team.Gemini 3.1 flash-lite model card, 2026. Google DeepMind. 4 [11] Daniel Geng, Inbum Park, and Andrew Owens. Visual ana- grams: Generating multi-view optical illusions with diffu- sion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24154–24163, 2024. 2 [12] Dayeon Ki, Tianyi Zhou, Marine Carpuat, Gang Wu, Puneet Mathur, and Viswanathan Swaminathan.GraphicBench: A planning benchmark for graphic design with language agents, 2025. 2 [13] Jing Yu Koh, Robert Lo, Lawrence Jang, Vikram Duvvur, Ming Chong Lim, Po-Yu Huang, Graham Neubig, Shuyan Zhou, Ruslan Salakhutdinov, and Daniel Fried. VisualWe- bArena: Evaluating multimodal agents on realistic visual web tasks. In Proceedings of the Annual Meeting of the As- sociation for Computational Linguistics (ACL), 2024. 2 [14] Dong Hyun Lee. Effective use of negative space in graphic design. Master’s thesis, Rochester Institute of Technology, 2007. 1 [15] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jian- wei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. GLIGEN: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 22511–22521, 2023. 2 [16] Edward Lucie-Smith. The Thames and Hudson Dictionary of Art Terms. Thames and Hudson, 1984. Negative volume, negative space: An enclosed empty space in architecture, sculpture or painting which makes an essential contribution to the composition. 1 [17] Ellen Lupton and Jennifer Cole Phillips. Graphic Design: The New Basics. Princeton Architectural Press, New York, 2015. 1 [18] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie. T2I- Adapter: Learning adapters to dig out more controllable abil- ity for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 4296–4304, 2024. 2 [19] OpenAI.GPT-4 technical report.arXiv preprint arXiv:2303.08774, 2023. 2 [20] OpenAI. GPT image 1.5. OpenAI API Documentation, 2025. 4 [21] Gemini Team. Gemini 3.1 pro model card. 2026. Google DeepMind. 4 [22] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean- Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 2 [23] Heng Wang, Yotaro Shimose, and Shingo Takamatsu. Ban- nerAgency: Advertising banner design with multimodal LLM agents. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4304–4329, 2025. 2 [24] Zhenyu Wang, Aoxue Li, Zhenguo Li, and Xihui Liu. Genartist: Multimodal llm as an agent for unified image gen- eration and editing. In Advances in Neural Information Pro- cessing Systems (NeurIPS), 2024. 2 [25] Peter Weibel. Negative space: Trajectories of sculpture in the 20th and 21st centuries. MIT Press, 2021. 1 [26] Tao Yang, Yingmin Luo, Zhongang Qi, Yang Wu, Ying Shan, and Chang Wen Chen. PosterLLaVa: Constructing a unified multi-modal layout generator with LLM, 2024. 2 [27] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3836–3847, 2023. 2 9