Paper deep dive
HGA-Quant: Heterogeneity-Aware Quantization for AutoRegressive Visual Generation Models
Xinmeng Liu, Xianglong Yan, Kaicheng Yang, Weisheng Dong, Yulun Zhang
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Recently, AutoRegressive Visual Generation (ARVG) models have emerged as a competitive paradigm for high-quality image synthesis. However, their substantial computational and memory demands make them difficult to deploy on resource-constrained edge devices. Post-Training Quantization (PTQ) provides a promising solution to reduce the computational and memory costs of ARVG models, yet existing methods often suffer from substantial image quality degradation under low-bit quantization. In this paper, we propose a training-free PTQ framework for ARVG, consisting of Tail-Aware Bounded-init Grid Refinement (T-BGR) and Coarse-to-Fine Activation Precision Scheduling (CAPS). T-BGR introduces a tail-aware reconstruction objective for weight quantizer calibration, assigning greater importance to distribution-tail coefficients and preserving fine-grained information without introducing additional parameters or high-precision branches. CAPS performs coarse-to-fine activation precision scheduling by jointly considering the importance and computational cost of different generation stages, enabling efficient mixed-precision allocation under a given bit budget. Extensive experiments on multiple ARVG backbones demonstrate that our framework consistently improves robustness under aggressive low-bit quantization. Notably, on Infinity, our framework achieves near full-precision image generation quality at an effective W4A5.6 activation precision, while existing methods at W4A6 still exhibit noticeable quality degradation or inconsistent performance. Code is available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2610.15297v1
- Canonical: https://arxiv.org/abs/2610.15297v1
Trouble viewing inline? Open PDF directly →
Full Text
50,895 characters extracted from source content.
Expand or collapse full text
HGA-Quant: Heterogeneity-Aware Quantization for AutoRegressive Visual Generation Models Xinmeng Liu †thanks: Equal contribution Xianglong Yan11footnotemark: 1 Kaicheng Yang Weisheng Dong Affiliation: Shanghai Jiao Tong University, Xi’an University of Electronic Science and Technology Yulun Zhang †thanks: Corresponding author: Yulun Zhang, yulun100@gmail.com Abstract Recently, AutoRegressive Visual Generation (ARVG) models have emerged as a competitive paradigm for high-quality image synthesis. However, their substantial computational and memory demands make them difficult to deploy on resource-constrained edge devices. Post-Training Quantization (PTQ) provides a promising solution to reduce the computational and memory costs of ARVG models, yet existing methods often suffer from substantial image quality degradation under low-bit quantization. In this paper, we propose a training-free PTQ framework for ARVG, consisting of Tail-Aware Bounded-init Grid Refinement (T-BGR) and Coarse-to-Fine Activation Precision Scheduling (CAPS). T-BGR introduces a tail-aware reconstruction objective for weight quantizer calibration, assigning greater importance to distribution-tail coefficients and preserving fine-grained information without introducing additional parameters or high-precision branches. CAPS performs coarse-to-fine activation precision scheduling by jointly considering the importance and computational cost of different generation stages, enabling efficient mixed-precision allocation under a given bit budget. Extensive experiments on multiple ARVG backbones demonstrate that our framework consistently improves robustness under aggressive low-bit quantization. Notably, on Infinity, our framework achieves near full-precision image generation quality at an effective W4A5.6 activation precision, while existing methods at W4A6 still exhibit noticeable quality degradation or inconsistent performance. Code is available at here. 1 Introduction Autoregressive visual generation (ARVG) models have recently emerged as a competitive paradigm for high-quality image synthesis. A representative line of work, Visual Autoregressive Modeling (VAR), reformulates image generation as next-scale prediction and constructs images from coarse to fine (Tian et al., 2024). Subsequent models have explored residual refinement, flexible scale schedules, and bitwise token prediction, as exemplified by HART (Tang et al., 2024), FlexVAR (Jiao et al., 2025), and Infinity (Han et al., 2025). These models provide a compelling alternative to diffusion-based generation, but their large Transformer backbones and repeated token prediction still impose substantial computational and memory demands. Such costs become particularly restrictive when ARVG models are deployed outside large-scale data-center settings. Post-Training Quantization (PTQ) is an attractive way to reduce the cost of pretrained generative models without retraining them. By representing weights and activations with low-bit integers, PTQ can reduce the numerical precision required by the dominant linear operations while retaining the original model structure. However, directly applying generic PTQ methods to ARVG models often causes severe image-quality degradation at aggressive bit-widths. Recent studies have identified channel-wise outliers, token-wise activation dynamics, and calibration-distribution mismatch as important sources of this difficulty (Liu et al., 2026). Other methods mitigate quantization errors through equivalent transformations (Xiao et al., 2023), rotations (Ashkboos et al., 2024), or auxiliary low-rank compensation (Li et al., 2025a). Despite these advances, the quality of visual autoregressive generation can still collapse when both weights and activations are quantized to low precision, indicating that the non-uniform structure of ARVG quantization errors requires more specialized treatment (Moon and Ham, 2026). We identify two complementary forms of heterogeneity that are especially relevant to low-bit ARVG quantization. First, weight coefficients are distributed unevenly across a large central mass and a much smaller tail region. Under an unweighted reconstruction loss, the central mass of a channel usually dominates the calibration objective, while tail coefficients contribute relatively little despite their larger magnitudes. Consequently, a calibration objective that minimizes average coefficient error may favor central-mass fidelity while leaving the tail reconstruction trade-off insufficiently controlled. Existing outlier-handling approaches often redistribute the quantization difficulty or preserve outliers through additional transformations or compensation paths. While these designs can improve numerical reconstruction, they do not directly expose how the tail-versus-central reconstruction trade-off should affect the calibration of the scale and zero-point. Second, the effect of activation quantization can vary along the ARVG generation trajectory. In the coarse-to-fine process of VAR-style models, early stages establish the global structural context on which later stages condition, while later stages operate on denser token grids and mainly refine local content (Tian et al., 2024). Uniform activation precision therefore ignores the position-dependent sensitivity of the generation trajectory as well as its computational cost. Protecting all stages equally can allocate precision to late refinements, whereas assigning precision only according to token count may overlook the different consequences of perturbing early and late stages. These observations motivate a quantization framework that answers two different questions: which values in a weight distribution deserve greater reconstruction priority, and when along the generation trajectory should higher activation precision be used? To address these challenges, we propose Heterogeneity-Aware Quantization for AutoRegressive Visual Generation models (HGA-Quant), a training-free PTQ framework composed of two complementary components. Tail-Aware Bounded-init Grid Refinement (T-BGR) extends bounded grid refinement with a tail-aware reconstruction objective. It identifies distribution-tail coefficients using channel statistics and assigns them larger weights during the search for the quantization interval, scale, and zero point. T-BGR keeps the standard quantized-weight representation while directing the calibration objective toward tail reconstruction. Coarse-to-Fine Activation Precision Scheduling (CAPS) exploits the generation-order prior of ARVG models. It estimates a stage-wise floating-point operations (FLOPs) proxy for the targeted linear projections and allocates the available high-precision budget to an early prefix of coarse stages, while using lower precision for the remaining fine stages. Thus, T-BGR targets non-uniform reconstruction priorities within the weight distribution, whereas CAPS targets non-uniform sensitivity along the generation trajectory. We evaluate HGA-Quant on four ARVG backbones, including Infinity-2B (Han et al., 2025), HART (Tang et al., 2024), VAR-d30 (Tian et al., 2024), and FlexVAR-d24 (Jiao et al., 2025), using WbAaW_bA_a to denote b-bit weights and a-bit activations, under W8A8, W4A8, W4A6, and W4A4 configurations. Across models and evaluation metrics, T-BGR remains competitive at higher precision and becomes increasingly robust relative to existing baselines as the bit-width decreases. On Infinity-2B, the stage-wise analysis in Figure 4 shows that the same A4 perturbation is substantially more damaging at early stages than at late stages. Accordingly, assigning W4A8 to the first 11 stages and W4A4 to the final two stages yields a FLOPs-weighted effective activation precision of 5.5711 bits, denoted W4A5.6 below, while maintaining performance comparable to uniform W4A6. The main tables report both uniform T-BGR (Ours) and the CAPS variant (Ours (CAPS)) where evaluated; the stage-wise and budgeted allocation studies further isolate the role of precision placement. Figure 1: Motivation for HGA-Quant. Left: most weights concentrate near the central mass, but a non-negligible fraction lies in the tails, motivating tail-aware calibration. Right: token counts and linear-layer FLOPs grow sharply in later stages, motivating coarse-to-fine precision scheduling. 2 Related Works 2.1 AutoRegressive Visual Generation models Recent ARVG models have emerged as a competitive paradigm for image synthesis. Representative approaches explore diverse autoregressive formulations, including scale-wise prediction (Tian et al., 2024), continuous-token modeling (Li et al., 2024b), randomized ordering (Yu et al., 2025), parallelized generation (Wang et al., 2024), and hybrid tokenization (Tang et al., 2024). VAR (Tian et al., 2024) is a foundational work that reformulates image generation as coarse-to-fine next-scale prediction, achieving strong image quality, inference efficiency, data efficiency, and scalability. HART (Tang et al., 2024) combines discrete autoregression with lightweight residual diffusion to improve reconstruction quality and enable efficient high-resolution generation. FlexVAR (Jiao et al., 2025) replaces residual prediction with ground-truth prediction at each scale, improving flexibility in resolution, aspect ratio, and inference steps. Infinity (Han et al., 2025) further scales ARVG through bitwise token prediction and an effectively enlarged token vocabulary. Its extensions have also explored unified image-video generation and image editing (Liu et al., 2025; Wang et al., 2025). Alongside architectural advances, recent studies have investigated efficient decoding (Chen et al., 2025; He et al., 2024), attention compression (Xie et al., 2024), and low-bit quantization (Xie et al., 2024; Wei et al., 2025) to reduce the computational and memory costs of ARVG models. 2.2 Post-Training Quantization Quantization is a standard compression technique for large generative models. It reduces model size, memory traffic, and inference latency. PTQ enables efficient deployment without retraining. GPTQ (Frantar et al., 2022) uses Hessian-guided error compensation to minimize quantization error. AWQ (Lin et al., 2024) protects salient weights based on activation statistics. Methods based on scale redistribution and equivalent transformations (Shao et al., 2023; Xiao et al., 2023; Wei et al., 2023) mitigate activation outliers by shifting or redistributing quantization difficulty across weights and activations. Rotation-based methods (Ashkboos et al., 2024; Li et al., 2025b) further transform the representation space to disperse outliers across dimensions, facilitating low-bit quantization. SVDQuant (Li et al., 2025a) and Q-DiT4SR (Zhang et al., 2026) use a high-precision low-rank branch to compensate for quantization errors after outlier transformation. PTQ4ARVG (Liu et al., 2026) specifically studies PTQ for autoregressive visual generation, addressing channel-wise outliers, token-wise activation dynamics, and sample-wise distribution mismatch. Some recent works (Moon and Ham, 2026; Vezzali et al., 2026) have begun to explore PTQ and quantization specifically for visual autoregressive models, targeting different aspects of efficient deployment. Despite these advances, significant image quality degradation remains under aggressive quantization configurations such as 4-bit weights and 6-bit activations in visual autoregressive models. 3 Method Figure 2: Overview of HGA-Quant. T-BGR performs tail-aware calibration for robust low-bit quantization, while CAPS assigns stage-wise precision in a coarse-to-fine manner for high-quality and efficient ARVG generation. 3.1 Preliminaries Autoregressive image generators such as Infinity synthesize an image through a sequence of coarse-to-fine stages. Let =ii=0S−1S=\p_i\_i=0^S-1 denote the scale schedule, where i=(pi,1,…,pi,D)p_i=(p_i,1,…,p_i,D) specifies the token grid at stage i. The number of image tokens processed at this stage is Ti=∏d=1Dpi,dT_i= _d=1^Dp_i,d. This generation process exposes two forms of non-uniformity that are not captured by conventional uniform post-training quantization (PTQ): different regions of a weight distribution can have different effects on the reconstructed output, and errors introduced at different generation stages need not propagate equally to the final image. We address these two sources of heterogeneity with T-BGR for tail-aware weight-quantizer calibration and CAPS for activation precision scheduling. Post-training quantization converts a pretrained full-precision model to a low-bit model by determining quantization parameters from the weights and, when applicable, a small calibration set, without updating model parameters. For a scalar x, an unsigned b-bit asymmetric affine quantizer is written as b(x,s,z)=s(clip(⌊xs⌉+z,0,2b−1)−z),Q_b(x;s,z)=s (clip ( xs +z,0,2^b-1 )-z ), (1) Here, s is the scale and z is the zero point. The corresponding integer code is q=clip(⌊x/s⌉+z,0,2b−1)q=clip( x/s +z,0,2^b-1), and the expression in Eq. equation 1 is its dequantized value. 3.2 Tail-Aware Bounded-init Grid Refinement 3.2.1 Motivation Bounded-init Grid Refinement (BGR) searches candidate quantization intervals and refines the corresponding affine scale and zero point to reduce weight reconstruction error (Li et al., 2025b). Its conventional reconstruction objective weights all coefficients equally. Under this objective, the large central mass dominates the total error, while a small number of tail coefficients contributes relatively little to the loss despite their larger magnitudes. Therefore, preserving these coefficients might be important for retaining fine-grained responses under aggressive quantization. We therefore propose Tail-Aware Bounded-init Grid Refinement (T-BGR), a tail-sensitive BGR extension. It modifies the calibration objective so that distribution-tail errors carry greater weight when selecting the quantization interval, scale, and zero point. This preserves the BGR deployment form while biasing its search toward a controlled tail-versus-central reconstruction trade-off. 3.2.2 Tail-Aware Reconstruction Objective Consider one output channel c∈ℝCinw_c ^C_in of a weight matrix, where wc,jw_c,j denotes its j-th weight coefficient. We compute its channel-wise mean and standard deviation as μc=1Cin∑jwc,j,σc=1Cin∑j(wc,j−μc)2. _c= 1C_in _jw_c,j, _c= 1C_in _j (w_c,j- _c )^2. (2) We identify a tail coefficient using a κσκσ criterion, mc,j=[|wc,j−μc|>κσc],m_c,j=1\! [ |w_c,j- _c |>κ _c ], (3) and assign it the reconstruction weight ρc,j=1+(γ−1)mc,j. _c,j=1+(γ-1)m_c,j. (4) Here, κ controls the statistical tail boundary and γ controls the relative priority assigned to tail coefficients. A larger γ increases the influence of tail reconstruction on the selected quantization parameters, while a smaller κ identifies a larger portion of the channel as tail. Let w^c,j=bw(wc,j,sc,zc) w_c,j=Q_b_w(w_c,j;s_c,z_c) be the dequantized value of wc,jw_c,j, where bwb_w denotes the weight bit-width. T-BGR minimizes the normalized tail-aware reconstruction error ℒT-BGR(c)(sc,zc)=∑jρc,j(w^c,j−wc,j)2∑jρc,j.L_T -BGR^(c)(s_c,z_c)= _j _c,j ( w_c,j-w_c,j )^2 _j _c,j. (5) When γ=1γ=1, Eq. equation 5 reduces to the uniformly weighted reconstruction criterion used by the original BGR calibration. For γ>1γ>1, the selected quantization parameters place greater emphasis on the reconstruction of tail coefficients, while the normalization keeps the loss scale comparable across channels with different tail ratios. 3.2.3 Weighted Grid Refinement T-BGR retains the bounded grid-search structure of BGR. For each channel, we generate a set of candidate clipping intervals [lc(r),rc(r)][l_c^(r),r_c^(r)] by progressively shrinking the original range [minjwc,j,maxjwc,j][ _jw_c,j, _jw_c,j]. For each candidate interval, we initialize an asymmetric affine quantizer and alternately update the scale and zero point. Given the current integer code qc,jq_c,j, the weighted scale update is sc←∑jρc,j(qc,j−zc)wc,j∑jρc,j(qc,j−zc)2,s_c← _j _c,j(q_c,j-z_c)w_c,j _j _c,j(q_c,j-z_c)^2, (6) followed by the weighted zero-point update zc←clip(⌊∑jρc,j(qc,j−wc,j/sc)∑jρc,j⌉,0,2bw−1).z_c ( _j _c,j (q_c,j-w_c,j/s_c ) _j _c,j ,0,2^b_w-1 ). (7) The codes are then requantized under the updated (sc,zc)(s_c,z_c), and the refined parameters with the smallest value of Eq. equation 5 are selected. 3.3 Coarse-to-Fine Activation Precision Scheduling 3.3.1 Motivation Uniform activation precision does not account for position-dependent sensitivity in coarse-to-fine autoregressive generation. Early stages contain relatively few tokens but establish the global structural scaffold on which all subsequent stages condition. Later stages operate over denser token grids and mainly refine local content. This ordering gives a practical prior for precision allocation: early perturbations can affect the context consumed by later stages, whereas later perturbations are more localized. CAPS uses this prior by reserving the limited high-precision budget for an early prefix of the autoregressive trajectory. 3.3.2 Stage-wise Linear FLOPs Proxy For a quantized linear projection m with input and output dimensions KmK_m and NmN_m, processing n tokens requires the linear FLOPs proxy 2nKmNm2nK_mN_m. At stage i, the scale vector ip_i defines a D-dimensional token grid, and Ti=∏d=1Dpi,dT_i= _d=1^Dp_i,d is the number of image tokens in that stage. Thus, image-token projections process BTiBT_i tokens for batch size B. We use FiF_i as the linear-layer FLOPs proxy for precision allocation. We also include text-conditioning KV projections using the number of conditional tokens TcondT_cond. Let ni,mn_i,m denote the resulting token count associated with projection m at stage i; the stage cost is Fi F_i =∑m∈ℳ2ni,mKmNm, = _m 2n_i,mK_mN_m, (8) ni,m n_i,m =BTi,mis an image-token projection,Tcond,mis a text-KV projection. = casesBT_i,&m\ is an image-token projection,\\ T_cond,&m\ is a text-KV projection. cases 3.3.3 FLOPs-Budgeted Precision Allocation CAPS assigns an activation bit-width bi∈blow,bhighb_i∈\b_low,b_high\ to each autoregressive stage while enforcing a FLOPs-weighted average-bit constraint: b¯act=∑i=1SFibi∑i=1SFi≤Btarget. b_act= _i=1^SF_ib_i _i=1^SF_i≤ B_target. (9) Equivalently, the FLOPs that can be assigned high precision are bounded by Fhighbudget=⌊(Btarget−blow)∑i=1SFibhigh−blow⌋.F_high^budget= (B_target-b_low) _i=1^SF_ib_high-b_low . (10) The choice of blowb_low, bhighb_high, and BtargetB_target specifies the available precision budget. Because stages are indivisible allocation units, the realized b¯act b_act can be lower than the target. 3.3.4 Prefix-Greedy Scheduling and Runtime Switching To instantiate the coarse-to-fine prior, CAPS uses a prefix-greedy allocation. Starting from the earliest stage, it assigns bhighb_high until the next stage would exceed Eq. equation 10; all remaining stages use blowb_low. Formally, the high-precision set is a prefix ℋ=1,…,kH=\1,…,k\, where k is the largest index satisfying ∑i=1kFi≤Fhighbudget. _i=1^kF_i≤ F_high^budget. (11) At inference time, CAPS constructs the plan once from the input scale schedule and switches the activation bit-width of all targeted quantized linear projections at each stage boundary. The prefill computation is initialized at bhighb_high to protect the initial conditioning and structural context; Eq. equation 9 is defined over the scheduled autoregressive coarse-to-fine stages. This runtime stage-adaptive behavior uses the same quantized linear modules throughout generation, while their activation precision changes as the generation stage progresses. Algorithm 1 HGA-Quant pipeline for autoregressive visual generation. 1: pretrained ARVG model fθf_θ; target modules ℳM; scale schedule S; quantizer settings C. 2: quantized model fθ^f_ θ and stage-wise activation plan Π . 3: ℳq←∅M_q← 4: apply the fixed random Hadamard representation transform when used by the target implementation 5: Offline weight calibration: 6: for each target module m∈ℳm with weight WmW_m do 7: (Qm,sm,zm)←T-BGR(Wm,)(Q_m,s_m,z_m)← T-BGR(W_m,C) 8: replace m with a quantized linear module storing (Qm,sm,zm)(Q_m,s_m,z_m) 9: ℳq←ℳq∪mM_q _q∪\m\ 10: end for 11: Runtime activation scheduling: 12: Π←CAPS(,ℳq,) ← CAPS(S,M_q,C) 13: run prefill with high-bit activations 14: for stage i=1,…,||i=1,…,|S| do 15: set all modules in ℳqM_q to activation bit-width Πi _i 16: generate the tokens of stage i 17: end for 18: return fθ^,Πf_ θ, 4 Experiments 4.1 Experimental Setup Models and Evaluation Protocols. We evaluate HGA-Quant on four ARVG backbones: Infinity-2B (Han et al., 2025), HART (Tang et al., 2024), VAR-d30 (Tian et al., 2024), and FlexVAR (Jiao et al., 2025). For Infinity, our primary testbed, we evaluate text-image alignment using GenEval (Ghosh et al., 2023) and DPG-Bench (Hu et al., 2024), and assess preference-oriented visual quality using ImageReward (Xu et al., 2023). For HART, we evaluate prompt adherence using DPG-Bench and visual quality using Fréchet Inception Distance (FID) on the MJHQ-30K benchmark (Li et al., 2024a). For VAR-d30 and FlexVAR, we generate 50K images on ImageNet (Deng et al., 2009) and evaluate generation quality using FID (Heusel et al., 2017) and Inception Score (IS) (Salimans et al., 2016). All experiments run on a single NVIDIA RTX A6000 GPU for image generation and evaluation. Compared Methods. Baselines include RTN (Round-to-Nearest), SmoothQuant (Xiao et al., 2023), SVDQuant (Li et al., 2025a), QuaRot (Ashkboos et al., 2024), and PTQ4ARVG (Liu et al., 2026); PTQ4ARVG uses 128 MS COCO captions for calibration (Lin et al., 2014). Configurations. We denote b-bit weights and a-bit activations by WbAaW_bA_a. Ours is uniform-precision T-BGR with κ=3,γ=8κ=3,γ=8; Ours (CAPS) uses the same weight calibration with FLOPs-budgeted A8/A4 stage scheduling targeting W4A6, whose realized average may fall below 6 bits because stages are indivisible. Our method and QuaRot use fixed Hadamard preprocessing; other baselines use their standard settings. Red/blue mark the best/second-best quantized results. 4.2 Main Results Table 1: Unified comparison on Infinity-2B and HART. Quantization Methods Infinity-2B HART GenEval ↑ ImageReward ↑ DPG-Bench ↑ DPG-Bench ↑ MJHQ-30K FID ↓ FP – 0.71896 0.95485 83.4741 74.835 6.178 W8A8 RTN 0.72706 0.90395 83.0017 75.754 12.476 SmoothQuant 0.72192 0.92828 83.0177 75.354 7.185 SVDQuant 0.71827 0.93168 82.8461 75.411 7.477 QuaRot 0.72866 0.94160 83.2944 75.019 6.380 PTQ4ARVG 0.74081 0.89858 82.971 – – Ours 0.73216 0.94418 83.1832 75.172 6.366 W4A8 RTN 0.61900 0.33108 79.1849 62.029 92.908 SmoothQuant 0.51707 -0.14869 75.0448 49.257 117.259 SVDQuant 0.61737 0.26062 78.9916 72.656 26.802 QuaRot 0.66563 0.75864 80.9206 72.173 16.669 PTQ4ARVG 0.68797 0.64962 81.6667 – – Ours 0.69825 0.83010 82.0691 71.895 9.379 W4A6 RTN 0.44685 -0.29458 71.0519 17.199 186.260 SmoothQuant 0.28231 -0.93134 66.0039 18.593 145.355 SVDQuant 0.48178 -0.30246 75.5081 64.644 82.087 QuaRot 0.65560 0.76999 80.8647 71.386 23.916 PTQ4ARVG 0.65213 0.50648 81.2235 – – Ours 0.68003 0.82882 82.0022 73.211 13.166 Ours (CAPS) 0.69043 0.83523 82.0877 71.058 21.675 W4A4 RTN 0.00312 -2.25143 8.2155 5.340 307.111 SmoothQuant 0.00141 -2.24901 6.0742 8.749 191.014 SVDQuant 0.00469 -2.25759 4.8947 10.072 195.629 QuaRot 0.47687 0.25269 75.3323 50.814 134.535 PTQ4ARVG 0.02376 -2.18336 26.2649 – – Ours 0.60207 0.44772 80.1773 61.417 88.475 Table 2: Unified comparison on VAR-d30 and FlexVAR. Quantization Methods VAR-d30 FlexVAR FID ↓ IS ↑ FID ↓ IS ↑ FP – 1.9827 323.326 2.1598 336.262 W8A8 RTN 1.9742 312.803 2.2093 329.612 SmoothQuant 2.0138 318.025 2.1797 331.019 SVDQuant 1.9799 319.311 2.1680 329.660 QuaRot 1.9708 322.788 2.1604 335.229 PTQ4ARVG 2.4351 314.487 – – Ours 1.9631 323.172 2.1854 336.288 W4A8 RTN 54.7231 45.315 10.4348 157.526 SmoothQuant 13.1672 146.123 6.3042 209.413 SVDQuant 6.6533 208.545 5.7541 229.426 QuaRot 3.9433 255.786 3.6883 264.783 PTQ4ARVG 26.9862 85.808 – – Ours 2.5532 293.163 2.7165 297.213 W4A6 RTN 92.5568 17.685 37.4952 59.187 SmoothQuant 21.2844 102.013 16.2536 130.247 SVDQuant 12.4921 150.365 14.8319 146.591 QuaRot 4.2578 249.087 4.1719 254.551 PTQ4ARVG 37.2724 63.484 – – Ours 2.6586 290.467 2.9150 290.874 Ours (CAPS) 2.9469 283.009 8.6066 201.977 W4A4 RTN 195.3049 4.439 233.0478 3.004 SmoothQuant 173.3009 4.683 221.9740 3.719 SVDQuant 153.7270 5.418 225.9401 3.537 QuaRot 16.9405 143.519 44.4773 55.159 PTQ4ARVG 109.2316 11.218 – – Ours 5.9785 232.040 17.6975 134.295 Infinity. T-BGR is competitive as activation precision decreases: at W4A4 it obtains 0.44772 ImageReward and 80.1773 DPG-Bench, compared with 0.25269 and 75.3323 for QuaRot. Most other methods collapse on at least one metric at W4A4. At the nominal W4A6 target, Ours (CAPS) reaches 0.69043, 0.83523, and 82.0877 on GenEval, ImageReward, and DPG-Bench with only 5.5711 FLOPs-weighted activation bits, improving over uniform W4A6 T-BGR (0.68003/0.82882/82.0022). HART, VAR-d30, and FlexVAR. Uniform T-BGR is strongest at W4A6 on these backbones, while CAPS realizes 5.4963, 5.5000, and 5.5000 effective activation bits, respectively, and loses some quality from the lower final budget. On HART, CAPS reaches 71.058 DPG-Bench and 21.675 FID, versus 73.211 and 13.166 for uniform W4A6; the FID is still better than QuaRot’s 23.916, although DPG-Bench is slightly lower than QuaRot’s 71.386, and both are far better than the other baselines. On VAR-d30, CAPS obtains 2.9469/283.009 (FID/IS), below uniform W4A6’s 2.6586/290.467 but better than QuaRot on both metrics. On FlexVAR, CAPS obtains 8.6066/201.977 versus 2.9150/290.874 for uniform W4A6 and 4.1719/254.551 for QuaRot, while remaining better than other baselines. The reduced activation budget can hurt absolute quality, but coarse-to-fine allocation still outperforms most alternatives at lower compute. CAPS stage sensitivity. Lowering one consecutive stage pair from A8 to A4 reduces ImageReward by 0.17448 for the first pair but only 0.00442 for stages 11–12 (Figure 4). CAPS achieves the lowest FLOPs-weighted activation precision (5.5711 bits) and highest ImageReward (0.83523), supporting a high-precision prefix and low-precision refinement. 4.3 Ablation Study Figure 3: Effect of the tail-stretch factor in T-BGR. T-BGR. We investigate the tail-stretch factor γ in Tail-Aware Bounded-init Grid Refinement (T-BGR), which controls the contribution of distribution-tail coefficients when optimizing the scale and zero point. Here, γ=1γ=1 corresponds to ordinary BGR, while γ>1γ>1 increases the reconstruction weight of coefficients satisfying |w−μc|>3σc|w- _c|>3 _c. We evaluate W4A5.6 with coarse-to-fine mixed precision, uniform W4A6, and uniform W4A4 on Infinity-2B using ImageReward. As shown in Table 3, the main setting γ=8γ=8 improves over no BGR in all three configurations and reaches 0.83523 for W4A5.6 and 0.82882 for W4A6. The optimum is not monotonic in the tail-stretch factor: W4A4 is highest at γ=0γ=0, which assigns zero weight to tail reconstruction error and emphasizes the central mass. This suggests that the most aggressive W4A4 regime favors preserving the dense central region over the rare extremes. We further isolate the weight-side effect with FP activations and fixed Hadamard preprocessing. Under W4 weight quantization with ordinary BGR (γ=1γ=1), ImageReward is 0.79494. Restoring the detected tail coefficients to FP raises it to 0.82331 when restoring a random half of the tail entries and to 0.83709 when restoring all of them, indicating that a small tail subset has a disproportionate influence on visual quality. T-BGR then tests whether this insight can be used without an FP outlier branch: assigning a larger tail-stretch factor, γ=8γ=8, reaches 0.84753, higher than both ordinary BGR and full tail restoration under γ=1γ=1. This supports treating T-BGR as a range-allocation mechanism that protects tail information while retaining the standard low-bit weight representation. Table 3: Ablation studies on Infinity-2B. Left: ablation of the T-BGR tail-stretch factor γ. Right: ablation of CAPS. W4A5.6 combines W4A4 and W4A8 at a FLOPs-weighted effective activation precision of 5.5711 bits; “no BGR” omits bounded grid refinement. no BGR γ=0γ=0 γ=1γ=1 γ=2γ=2 γ=8γ=8 γ=16γ=16 W4A5.6 0.81594 0.81001 0.81458 0.80069 0.83523 0.81491 W4A6 0.82125 0.80062 0.80668 0.79104 0.82882 0.79840 W4A4 0.40181 0.50311 0.48880 0.49401 0.44772 0.43064 GenEval↑ ImageReward↑ DPG-Bench↑ W4A8 0.69825 0.83010 82.0691 W4A6 0.68003 0.82882 82.0022 W4A4 0.60207 0.44772 80.1773 W4A5.6 0.69043 0.83523 82.0877 CAPS. Table 3 evaluates CAPS. W4A5.6 obtains 0.69043, 0.83523, and 82.0877 on GenEval, ImageReward, and DPG-Bench, respectively, compared with 0.68003, 0.82882, and 82.0022 for uniform W4A6. Uniform W4A4 instead drops to 0.60207, 0.44772, and 80.1773. Figure 4 motivates prefix protection; Appendix D compares early-high, late-high, and random schedules. Figure 4: CAPS stage sensitivity on Infinity-2B. The table reports FLOPs-weighted effective activation bits and ImageReward when changing one consecutive stage pair from A8 to A4; the curve plots the same ImageReward values with the uniform W4A8 reference as a dashed horizontal line. 5 Conclusion We presented HGA-Quant, a training-free PTQ framework for low-bit ARVG that combines tail-aware weight calibration with budgeted activation allocation. Experiments across four backbones demonstrate the benefits of T-BGR under aggressive quantization. CAPS supports protecting early coarse stages, but reduced activation budgets can compromise quality on some backbones. These findings motivate backbone-specific precision schedules to balance generation quality and efficiency. AI use statement Generative AI tools were used to assist with drafting and modifying experimental code, including initial implementations of core experimental components and auxiliary scripts. They were also used for language polishing, minor revisions to the organization and content of the manuscript, visualization code, and figure layout ideas. All AI-assisted code was carefully reviewed, tested, and modified before use, and AI-assisted manuscript content was reviewed and verified before inclusion. The core research ideas, methodology, and experimental design were developed independently of generative AI. Experimental results were analyzed and interpreted by the research team, and the conclusions were independently formulated based on these analyses. The authors take full responsibility for the final content of this work, including all text, code, results, and figures. Ethics statement This work studies post-training quantization methods for improving the efficiency of autoregressive visual generation models. Quantized models may retain the biases, limitations, and potential for misuse of their pretrained counterparts, while reduced computational and memory costs may facilitate broader deployment of such models. Our experiments focus on generation quality and quantization efficiency and do not evaluate the safety, fairness, or potential misuse of the resulting models. The experiments use publicly available models and benchmark datasets and do not involve human subjects or the collection of personal or sensitive data. Reproducibility statement We will publicly release the source code and relevant configurations used to obtain the experimental results reported in this paper. The release will cover our quantization implementations for Infinity, HART, VAR, and FlexVAR, together with the corresponding experimental configurations, to facilitate reproduction of our results. References Ashkboos et al. (2024) S. Ashkboos, A. Mohtashami, M. L. Croci, B. Li, P. Cameron, M. Jaggi, D. Alistarh, T. Hoefler, and J. Hensman QuaRot: outlier-free 4-bit inference in rotated llms. arXiv preprint arXiv:2404.00456. External Links: Link Cited by: §1, §2.2, §4.1. Chen et al. (2025) Z. Chen, X. Ma, G. Fang, and X. Wang Collaborative decoding makes visual auto-regressive modeling efficient. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), External Links: Link Cited by: §2.1. Deng et al. (2009) J. Deng, W. Dong, R. Socher, L. Li, K. Li, and L. Fei-Fei ImageNet: A large-scale hierarchical image database. In 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2009), 20-25 June 2009, Miami, Florida, USA, Cited by: §4.1. Frantar et al. (2022) E. Frantar, S. Ashkboos, T. Hoefler, and D. Alistarh GPTQ: accurate post-training quantization for generative pre-trained transformers. arXiv preprint arXiv:2210.17323. Cited by: §2.2. Ghosh et al. (2023) D. Ghosh, H. Hajishirzi, and L. Schmidt GenEval: an object-focused framework for evaluating text-to-image alignment. In Neural Information Processing Systems, External Links: Link Cited by: §4.1. Han et al. (2025) J. Han, J. Liu, Y. Jiang, B. Yan, Y. Zhang, Z. Yuan, B. Peng, and X. Liu Infinity: scaling bitwise autoregressive modeling for high-resolution image synthesis. In Computer Vision and Pattern Recognition, External Links: Link Cited by: §1, §1, §2.1, §4.1. He et al. (2024) Y. He, F. Chen, Y. He, S. He, H. Zhou, K. Zhang, and B. Zhuang ZipAR: parallel auto-regressive image generation through spatial locality. arXiv preprint arXiv:2412.04062. External Links: Link Cited by: §2.1. Heusel et al. (2017) M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Neural Information Processing Systems, Cited by: §4.1. Hu et al. (2024) X. Hu, R. Wang, Y. Fang, B. Fu, P. Cheng, and G. Yu ELLA: equip diffusion models with llm for enhanced semantic alignment. External Links: 2403.05135 Cited by: §4.1. Jiao et al. (2025) S. Jiao, G. Zhang, Y. Qian, J. Huang, Y. Zhao, H. Shi, L. Ma, Y. Wei, and Z. Jie FlexVAR: flexible visual autoregressive modeling without residual prediction. In Advances in Neural Information Processing Systems, Vol. 38, Main Conference. External Links: Link Cited by: §1, §1, §2.1, §4.1. Li et al. (2024a) D. Li, A. Kamko, E. Akhgari, A. Sabet, L. Xu, and S. Doshi Playground v2.5: three insights towards enhancing aesthetic quality in text-to-image generation. External Links: 2402.17245, Link Cited by: §4.1. Li et al. (2025a) M. Li, Y. Lin, Z. Zhang, T. Cai, X. Li, J. Guo, E. Xie, C. Meng, J. Zhu, and S. Han SVDQuant: absorbing outliers by low-rank components for 4-bit diffusion models. In The Thirteenth International Conference on Learning Representations, External Links: Link Cited by: §1, §2.2, §4.1. Li et al. (2024b) T. Li, Y. Tian, H. Li, M. Deng, and K. He Autoregressive image generation without vector quantization. Advances in Neural Information Processing Systems. Cited by: §2.1. Li et al. (2025b) Z. Li, H. Li, J. Wu, K. Liu, Q. Haotong, L. Kong, G. Chen, Y. Zhang, and X. Yang DVD-quant: data-free video diffusion transformers quantization. arXiv preprint arXiv:2505.18663. Cited by: §2.2, §3.2.1. Lin et al. (2024) J. Lin, J. Tang, H. Tang, S. Yang, W. Chen, W. Wang, G. Xiao, X. Dang, C. Gan, and S. Han AWQ: activation-aware weight quantization for llm compression and acceleration. In MLSys, Cited by: §2.2. Lin et al. (2014) T. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick Microsoft coco: common objects in context. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, External Links: Link Cited by: §4.1. Liu et al. (2025) J. Liu, J. Han, B. Yan, H. Wu, F. Zhu, X. Wang, Y. Jiang, B. Peng, and Z. Yuan InfinityStar: unified spacetime autoregressive modeling for visual generation. In Neural Information Processing Systems, External Links: Link Cited by: §2.1. Liu et al. (2026) X. Liu, Z. Li, J. Zhang, M. Chen, and Q. Gu PTQ4ARVG: post-training quantization for autoregressive visual generation models. International Conference on Learning Representations. Cited by: §1, §2.2, §4.1. Moon and Ham (2026) J. Moon and B. Ham Shift-and-sum quantization for visual autoregressive models. In International Conference on Learning Representations, External Links: Link Cited by: §1, §2.2. Salimans et al. (2016) T. Salimans, I. J. Goodfellow, W. Zaremba, V. Cheung, A. Radford, and X. Chen Improved techniques for training gans. In Neural Information Processing Systems, External Links: Link Cited by: §4.1. Shao et al. (2023) W. Shao, M. Chen, Z. Zhang, P. Xu, L. Zhao, Z. Li, K. Z. Zhang, P. Gao, Y. Qiao, and P. Luo OmniQuant: omnidirectionally calibrated quantization for large language models. arXiv preprint arXiv:2308.13137. Cited by: §2.2. Tang et al. (2024) H. Tang, Y. Wu, S. Yang, E. Xie, J. Chen, J. Chen, Z. Zhang, H. Cai, Y. Lu, and S. Han HART: efficient visual generation with hybrid autoregressive transformer. International Conference on Learning Representations. External Links: Link Cited by: §1, §1, §2.1, §4.1. Tian et al. (2024) K. Tian, Y. Jiang, Z. Yuan, B. Peng, and L. Wang Visual autoregressive modeling: scalable image generation via next-scale prediction. In Neural Information Processing Systems, External Links: Link Cited by: §1, §1, §1, §2.1, §4.1. Vezzali et al. (2026) E. Vezzali, F. Bolelli, C. Grana, L. Benini, and Y. Li Enabling 8b bitwise autoregressive image generation on edge gpus. In Pattern Recognition - 28th International Conference, ICPR 2026, Lyon, France, August 17-22, 2026, Proceedings, Part XII, Cited by: §2.2. Wang et al. (2025) J. Wang, Y. Chen, J. Yu, G. Lu, and W. Pei EditInfinity: image editing with binary-quantized generative models. Advances in Neural Information Processing Systems. Cited by: §2.1. Wang et al. (2024) Y. Wang, S. Ren, Z. Lin, Y. Han, H. Guo, Z. Yang, D. Zou, J. Feng, and X. Liu Parallelized autoregressive visual generation. arXiv preprint arXiv:2412.15119. Cited by: §2.1. Wei et al. (2025) R. Wei, S. Xu, Q. Guo, and M. Li FPQVAR: floating point quantization for visual autoregressive model with fpga hardware co-design. arXiv preprint arXiv:2505.16335. Cited by: §2.1. Wei et al. (2023) X. Wei, Y. Zhang, Y. Li, X. Zhang, R. Gong, J. Guo, and X. Liu Outlier suppression+: accurate quantization of large language models by equivalent and optimal shifting and scaling. arXiv preprint arXiv:2304.09145. Cited by: §2.2. Xiao et al. (2023) G. Xiao, J. Lin, M. Seznec, H. Wu, J. Demouth, and S. Han SmoothQuant: accurate and efficient post-training quantization for large language models. In Proceedings of the 40th International Conference on Machine Learning, External Links: Link Cited by: §1, §2.2, §4.1. Xie et al. (2024) R. Xie, T. Zhao, Z. Yuan, R. Wan, W. Gao, Z. Zhu, X. Ning, and Y. Wang LiteVAR: compressing visual autoregressive modelling with efficient attention and quantization. arXiv preprint arXiv:2411.17178. Cited by: §2.1. Xu et al. (2023) J. Xu, X. Liu, Y. Wu, Y. Tong, Q. Li, M. Ding, J. Tang, and Y. Dong ImageReward: learning and evaluating human preferences for text-to-image generation. In Proceedings of the 37th International Conference on Neural Information Processing Systems, External Links: Link Cited by: §4.1. Yu et al. (2025) Q. Yu, J. He, X. Deng, X. Shen, and L. Chen Randomized autoregressive visual generation. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). Cited by: §2.1. Zhang et al. (2026) X. Zhang, K. Yang, H. Lu, H. Qin, Y. Guo, and Y. Zhang Q-dit4sr: exploration of detail-preserving diffusion transformer quantization for real-world image super-resolution. arXiv preprint arXiv:2602.01273. Cited by: §2.2. Appendix A Algorithm Algorithm 2 Tail-Aware Bounded-init Grid Refinement (T-BGR). 1: weight matrix W; bit-width bwb_w; grid size N; tail threshold κ; tail weight γ; tolerance ϵε; iteration limit ImaxI_ . 2: integer codes Q⋆Q ; real-valued scales s⋆s ; integer zero points z⋆z . 3: qmax←2bw−1q_ ← 2^b_w-1 4: compute per-channel statistics μc,σc _c, _c of Wc,:W_c,: 5: ρc,j←1+(γ−1)(|wc,j−μc|>κσc) _c,j← 1+(γ-1)1(|w_c,j- _c|>κ _c) 6: initialize channel-wise best loss ℒc⋆←+∞L _c←+∞ 7: ℓc←minjwc,j _c← _jw_c,j, uc←maxjwc,ju_c← _jw_c,j, and ηc←(uc−ℓc)/(2N) _c←(u_c- _c)/(2N) 8: for r=0,…,N−1r=0,…,N-1 do 9: Wc,:(r)←clip(Wc,:,ℓc+rηc,uc−rηc)W^(r)_c,: (W_c,:, _c+r _c,u_c-r _c) ⊳ bounded initialization 10: initialize (s,z,Q)(s,z,Q) from the asymmetric affine quantizer induced by W(r)W^(r) 11: for t=1,…,Imaxt=1,…,I_ do 12: A←Q−zA← Q-z 13: sc←∑jρc,jAc,jwc,j∑jρc,jAc,j2s_c← _j _c,jA_c,jw_c,j _j _c,jA_c,j^2 14: zc←clip(round(∑jρc,j(qc,j−wc,j/sc)∑jρc,j),0,qmax)z_c \! (round\! ( _j _c,j(q_c,j-w_c,j/s_c) _j _c,j ),0,q_ ) 15: Q←clip(round(W/s)+z,0,qmax)Q (round(W/s)+z,0,q_ ) 16: ℒc←∑jρc,j(wc,j−(qc,j−zc)sc)2∑jρc,jL_c← _j _c,j (w_c,j-(q_c,j-z_c)s_c )^2 _j _c,j 17: if maxc|ℒc−ℒcprev|<ϵ _c|L_c-L^prev_c|<ε then 18: break 19: end if 20: ℒprev←ℒL^prev 21: end for 22: update (Qc,:⋆,sc⋆,zc⋆,ℒc⋆)(Q _c,:,s _c,z _c,L _c) for channels with ℒc<ℒc⋆L_c<L _c 23: end for 24: return Q⋆,s⋆,z⋆Q ,s ,z Algorithm 3 Coarse-to-Fine Activation Precision Scheduling (CAPS). 1: scale schedule =ii=1SS=\p_i\_i=1^S; quantized modules ℳqM_q; effective batch size B; conditional token count TcondT_cond; bit-widths blow<bhighb_low<b_high; target average bit-width btarb_tar. 2: stage-wise activation plan Π=(Π1,…,ΠS) =( _1,…, _S). 3: for i=1,…,Si=1,…,S do 4: Ti←∏dpi,dT_i← _dp_i,d 5: Fi←∑m∈ℳq2ni,mdmindmoutF_i← _m _q2\,n_i,m\,d_m^ind_m^out, where ni,m=Tcondn_i,m=T_cond for text-KV projections and ni,m=BTin_i,m=BT_i otherwise 6: end for 7: Fbudget←⌊btar−blowbhigh−blow∑i=1SFi⌋F_budget← b_tar-b_lowb_high-b_low _i=1^SF_i 8: choose the longest prefix ℋ=1,…,kH=\1,…,k\ such that ∑i∈ℋFi≤Fbudget _i F_i≤ F_budget 9: for i=1,…,Si=1,…,S do 10: Πi←bhigh _i← b_high if i∈ℋi , otherwise Πi←blow _i← b_low 11: end for 12: return Π Appendix B Weight-Only Outlier Diagnostic under W3 The main ablation uses W4 to study tail-aware weight calibration. To check whether the same intuition remains relevant under a more aggressive weight-only setting, we also run W3 quantization with FP activations and fixed Hadamard preprocessing. Without BGR, W3 collapses to -2.25053 ImageReward, while ordinary BGR (γ=1γ=1) recovers the score to 0.58618. This confirms that a range-refinement baseline is important before studying finer tail behavior in very low-bit weights. Starting from W3 with ordinary BGR, restoring all detected tail coefficients to FP further improves ImageReward from 0.58618 to 0.61887. Restoring only a random half gives 0.57379, slightly below the non-restored run, suggesting that the W3 regime is sensitive to uneven mixed-precision perturbations rather than following a simple monotonic rule. The full-restoration result nevertheless supports the same qualitative conclusion as W4: tail precision remains useful when the bit-width is pushed lower. We also evaluate the stronger T-BGR setting γ=8γ=8 under W3. It obtains 0.32632, far above the no-BGR collapse but below ordinary BGR. This does not negate the tail-aware motivation; instead, it shows that the appropriate tail emphasis depends on the available bit-width. In an extremely narrow W3 codebook, aggressively stretching the tail can overly reduce the resolution of the central mass, so a bit-width-adaptive tail factor may be preferable. Appendix C Coarse-to-Fine Generation Visualization Figure 5 provides an intuitive view of the coarse-to-fine generation process of Infinity. The early stages mainly establish global structures and compositions, while later stages progressively refine object details and visual appearance. This observation motivates stage-aware precision allocation: higher precision can be preserved for early structural generation, while lower precision can be applied to later refinement stages. Figure 5: Compact illustration of coarse-to-fine generation at selected stages. The generation process gradually evolves from coarse layouts to detailed images, with early stages determining global structures and later stages refining object-level details. To further illustrate this behavior, Figure 6 compares intermediate outputs at all 13 stages under full precision (FP), uniform W4A6, CAPS with W4A5.6, and uniform W4A4. The early stages preserve the global layout, while later stages mainly contribute incremental refinements. CAPS maintains a coherent trajectory after switching to lower precision in the final stages, whereas uniform W4A4 introduces artifacts from early stages and produces a visibly degraded trajectory. Figure 6: Stage-wise visualization of coarse-to-fine generation. We compare full precision (FP), uniform W4A6, CAPS with W4A5.6, and uniform W4A4 over all 13 generation stages. CAPS preserves the coarse structure with its high-precision prefix and maintains coherent refinement after switching to lower precision in the final stages, whereas uniform W4A4 exhibits visible artifacts throughout the trajectory. Appendix D Budgeted Allocation Comparison The stage-wise results in the main text establish that early and late stages have different precision sensitivity. We therefore add a broader placement check under the same target FLOPs budget. The early-prefix schedule is the CAPS allocation used by our method, with A8 assigned to stages 1–11 and A4 assigned to stages 12–13. The late-high schedule assigns A8 only to the final refinement stage, which has the largest token grid. For random schedules, we start from all-A4 stages, sample a deterministic random stage order, and switch a stage to A8 when doing so does not exceed the target budget; if a stage would exceed the budget, it is skipped and the remaining stages are considered. All rows use the common allocation-comparison entry; consequently, its early-high score is 0.83523. Table 4: Budgeted CAPS allocation comparison on Infinity-2B. Stage labels are one-based. The reported average bit-width is the FLOPs-weighted effective activation precision realized by each allocation, and all rows use the common allocation-comparison evaluation entry. Allocation A8 stages Avg. act. bits ImageReward ↑ Δ vs. early-high early-high (CAPS) 1–11 5.5711 0.83523 0.00000 late-high 13 5.5543 0.44577 -0.38946 random_01 1,2,3,4,5,6,8,9,11,12 5.9587 0.81186 -0.02337 random_02 1,2,3,4,5,10,13 5.9930 0.77997 -0.05526 random_03 1,2,3,4,5,6,7,8,9,10,12 5.8381 0.82448 -0.01075 random_04 1,2,3,4,5,6,7,8,13 5.9095 0.80030 -0.03493 random_05 1,2,3,4,5,8,9,13 5.9755 0.78069 -0.05454 The early-high and late-high schedules are also close in compute: their high-precision linear FLOPs are 31.20T and 30.87T, differing by 1.06% of the high-precision portions and only 0.42% of the full stage FLOPs. Nevertheless, ImageReward falls from 0.83523 to 0.44577, a decrease of 0.38946 that is close to uniform W4A4 (0.44772). Among the random schedules, the best result is random_03 at 0.82448, still 0.01075 below early-high despite using a higher realized activation precision of 5.8381 bits. Thus, neither a similar compute budget nor a larger average precision is sufficient without protecting the early stages. Appendix E Qualitative Comparisons We provide qualitative comparisons across the evaluated methods under W4A6 and coarse-to-fine W4A5.6 quantization on Infinity-2B. For each example, all methods are conditioned on the same text prompt. Our method generally preserves the global composition and fine-grained details more faithfully, while several baselines exhibit visible artifacts, structural distortion, or severe quality degradation at aggressive bit-widths. Figure 7: Comparisons across different methods. In this figure, all images were obtained with the model undergoing W4A6 quantization. Figure 8: In this figure, the images labeled ”ours*” were obtained using our coarse-to-fine W4A5.6 approach, while the images from other methods were generated using W4A6 quantization.