acceptodds
Under review as a conference paper at ICLR 2027

VISTA: Test-Time Compositional Alignment for Visual Autoregressive Generation

Abstract

Visual autoregressive (VAR) models offer fast, high-quality text-to-image generation but still fail at attribute binding and spatial relations. Existing VAR test-time methods search sampled trajectories or apply generic logit guidance but do not directly optimize compositional constraints; gradient-based diffusion methods do not transfer to stateful, discrete, multi-resolution VAR sampling. We introduce (sual Autoregressive emantic est-time lignment), the first gradient-based test-time framework for compositional alignment in next-scale autoregressive image generation. Built on Infinity, VISTA optimizes intermediate representations through the frozen transformer without parameter updates or training. It stabilizes optimization across scales and exposes a common interface into which any differentiable constraint on cross-attention can be plugged. Across two benchmarks and model scales, VISTA improves every targeted category, raising the mean by nearly 20% on a 2B backbone and almost 6% on an 8B backbone, especially on spatial relations. In human evaluation, VISTA wins 83% of decisive compositionality comparisons by plurality vote, with no significant preference in perceived quality; independently, ImageReward scores its outputs 20% higher. VISTA enables the 2B model to surpass a backbone four times its size, showing much of the compositional gap is recoverable at test time.

Primary area:
generative models
Keywords:
visual autoregressive models, text-to-image generation, compositional generation, test-time alignment, inference-time optimization, spatial reasoning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_fGzhOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces VISTA, a training-free, gradient-based test-time alignment framework for next-scale visual autoregressive (VAR) text-to-image models, aiming to improve attribute binding and spatial relationships. At selected early, coarse scales of Infinity, VISTA optimizes the continuous scale input through backpropagation. It uses compositional losses computed from differentiable cross-attention maps, combining CONFORM-style binding and PSG-style planar objectives with a newly proposed boundary-containment depth loss. To support optimization during stateful multi-scale sampling, the method incorporates KV-cache isolation, per-token gradient normalization, scale-aware step sizes, and an adaptive step budget. Experiments on T2I-CompBench and GenEval show substantial improvements over Infinity-2B/8B: the targeted average increases from 0.480 to 0.575 on 2B, while GenEval Position improves from 0.25 to 0.745. ImageReward scores and human preferences for compositionality also show improvements.

Critical strengths
  1. The paper identifies an important gap: while test-time attention steering has been widely explored in diffusion models, it remains relatively underdeveloped for next-scale VAR models. The paper also clearly explains why existing diffusion-based approaches cannot be directly applied.

  2. The proposed stabilization techniques, including cache isolation, per-token gradient normalization, and scale-aware step sizes, are each designed to address specific challenges arising from stateful multi-scale sampling.

  3. The framework supports modular objectives, allowing existing attention-based losses from diffusion models to be incorporated easily. Among 17 tested configurations, 15 outperform the original backbone.

  4. The improvements are substantial and consistent across random seeds, including a +20% gain in the targeted average on 2B. Notably, Infinity-2B with VISTA even outperforms the original 8B model.

  5. The experimental evaluation is relatively comprehensive, covering human studies with statistical significance tests, ablation experiments, and runtime and memory analysis. The paper also acknowledges limitations such as increased latency and reduced aesthetic quality.

Critical weaknesses
  1. The methodological novelty is relatively limited. The binding and planar losses are directly adopted from CONFORM and PSG, respectively. The primary contribution lies in adapting existing objectives to VAR models through several engineering techniques, along with a heuristic depth objective that provides relatively small improvements (+0.047 on 2B and +0.010 on 8B for 3D). The claim of being the “first gradient-based” approach also needs more careful discussion in relation to SSG and VPG, as their comparisons rely only on results reported in the original papers.

  2. The baseline comparisons are not matched in computational cost or experimental protocol. VISTA is evaluated over 4 seeds, whereas BoN and TTS-VAR use only one, and TTS-VAR is tested only with N = 1 and 2. Interestingly, BoN (N = 8) performs worse than the original backbone, without further explanation. The paper also lacks comparisons under matched wall-clock time and does not investigate whether search-based methods could be combined with gradient-based steering.

  3. The adaptive step-budget scheduler appears to have no practical effect on 8B. With τ = 1e-12, the appendix indicates that neither early stopping nor budget reduction is activated, meaning every scale uses all 5 optimization steps. Moreover, there is no adaptive-vs-fixed ablation on 2B to demonstrate the scheduler's effectiveness. The values of τ also differ by nine orders of magnitude across backbones, while lr_base differs by 9×, both selected through “brief manual exploration”. This raises questions about the method's robustness and the amount of hyperparameter tuning required for other backbones.

  4. The generalizability of the method remains unclear beyond Infinity. All experiments are conducted on Infinity with ScaleKV, and the approach depends on explicit text cross-attention and a dependency parser. The framework also does not address numeracy or non-spatial relationships, while improvements on the Complex category are quite limited (0.384 → 0.398).

Questions for authors
  1. Could you compare the adaptive scheduler with a fixed N_max budget on 2B? Was the prompt subset used to manually tune τ and lr kept separate from the evaluation prompts?

  2. Why does BoN (N = 8) perform worse than the original backbone? Could you provide comparisons with BoN or TTS-VAR under matched wall-clock time and identical random seeds?

  3. Could you evaluate SSG and VPG using the same backbone and evaluation protocol to enable a more direct comparison?

  4. Have you tested VISTA on VAR architectures beyond Infinity, particularly models that do not provide explicit text cross-attention?

5 comments

Thank you for the detailed review. We added a fixed-budget ablation, corrected the BoN result, reran TTS-VAR across four seeds, and extended the evaluation to Switti. We address your four concerns below.

Novelty, the depth objective, and comparison with SSG and VPG

Our work addresses a technically challenging problem because the discrete, multi-resolution, and stateful nature of visual autoregressive generation makes directly transferring existing alignment methods particularly difficult. To our knowledge, VISTA is the first to transfer compositional cross-attention losses originally defined for diffusion models to discrete next-scale visual autoregressive generation. We demonstrate this with CONFORM for binding, PSG for planar relations, and CountCluster for numeracy. Each scale predicts a larger token grid conditioned on earlier scales. VISTA updates the continuous input through the frozen transformer before sampling that grid, so the intervention changes the tokens available to all finer scales. In Infinity, cache isolation preserves the sampling state during auxiliary forwards. The new ablation shows that the choice of gradient normalization materially affects the result: the targeted mean is 0.5733 with per-token normalization, 0.5318 with global L2, and 0.5283 with raw gradients (Sections 3.1–3.4 and Table 17).

We narrowed the priority claim to the mechanism introduced by VISTA. To our knowledge, VISTA is the first method to backpropagate explicit compositional cross-attention losses through a frozen next-scale visual autoregressive generator to update its continuous scale inputs before discrete token sampling. SSG uses a closed-form logit update to reinforce coarse-to-fine refinement, and VPG adjusts logits to improve consistency with the generated visual prefix. Both evaluate compositional performance, but their guidance rules do not explicitly encode constraints such as attribute binding or spatial relations.

We reproduced SSG on Infinity‑2B with four T2I‑CompBench seeds. Binding scores improved, both spatial scores declined, and Overall remained essentially flat (Appendix N, Table 21). We could not locate a public VPG implementation, so Table 22 reports the authors’ GenEval results under their rewritten-prompt protocol, including gains on Two Object and Position. A matched comparison with VPG remains open. In our paired evaluation against the same Infinity‑2B backbone, VISTA improves the targeted T2I‑CompBench mean by 19.8%, with an 88.9% relative gain on 2D Spatial.

Depth objective. On the full T2I-CompBench 3D Spatial category, VISTA improves Infinity‑2B by 11.6% (0.405 → 0.452), compared with 6.7% for BoN and 3.5% and 4.9% for TTS-VAR at N=1 and N=2, respectively, under the reported settings. VISTA also improves Infinity‑8B by 2.4% (0.414 → 0.424, Table 1).

To test whether the choice of depth loss matters, we compared five depth objectives on the same 3D Spatial subset with the rest of the steering pipeline held fixed. Boundary containment achieved the highest score (0.495, +12.8% over the unsteered baseline of 0.439), followed by peak dominance (0.487, +11.0%, Appendix E, Table 8). This supports our choice of depth objective on the tested subset. On 3D Spatial, global soft IoU between noun attention maps fell by 16.1%, and the category’s image benchmark score rose by about 14% in the attention audit. The lower overlap shows that VISTA separates the noun attention maps more strongly. (Appendix I, Table 13).

BoN, TTS-VAR, and matched compute

We rewrote the BoN evaluation pipeline and reran the eight-candidate baseline, generating eight complete trajectories and selecting the image with the highest ImageReward score. The corrected BoN result is 0.514 Overall, up from the Base score of 0.470 (Table 1). TTS-VAR at (N=1) and (N=2) now uses four seeds on the same Infinity-2B backbone, prompts, and evaluator as VISTA and Base. It reaches 0.494 and 0.502, respectively, versus 0.546 for VISTA at the reported settings (Table 1 and Appendix M, Table 19).

At the settings profiled in Table 20, VISTA with ScaleKV took 6.68 s, compared with 19.01 s for TTS-VAR at N=1 and 36.37 s at N=2. In the benchmark comparison, VISTA achieved 0.546 Overall, versus 0.494 and 0.502 for the two TTS-VAR settings (Table 1). The approaches could be combined by using search to retain promising coarse prefixes and applying VISTA before sampling subsequent scales. We leave that combination for future evaluation.

Generality, attention requirements, and constraint coverage. Switti and Infinity share the same core VAR process of generating discrete token grids from coarse to fine. Switti uses a different tokenizer and current-scale self-attention while retaining explicit text cross-attention. VISTA raises Switti’s seven-category CompBench mean by 13.3%, including Numeracy, and its three-category GenEval mean by 36.4% (Tables 1–2 and Appendix G). The uncompressed Infinity‑2B reference scores 0.2% higher than the ScaleKV baseline, showing that the cache-compressed baseline starts at a comparable level (Appendix F).

The current implementation requires differentiable text cross-attention. For extracting constraints, we use a simple dependency parser. Fixed templates would suffice for many benchmark prompts. Parsing lets us use the same extraction rules for prompts phrased outside those templates (Section 5).

The CountCluster extension (Lee et al., 2025, arXiv:2508.10710) raises Infinity‑2B Numeracy by 6.6%. With counting disabled, full-category Numeracy and Non-Spatial scores are nearly unchanged. Few mixed prompts received binding updates, and a dedicated non-spatial objective remains unevaluated (Appendix F, Tables 9–10). The Complex gain is modest (3.6%), consistent with the greater difficulty of satisfying several constraints in the same image.

Adaptive versus fixed scheduling on Infinity‑2B. On 1,800 T2I-CompBench validation prompts with two seeds, a fixed five-step budget reaches 0.5712, within 0.0021 of adaptive scheduling (0.5733), without using a gradient threshold. Adaptive scheduling uses 3.3% fewer updates and 1.8% less generation time than the five-step policy. The second fixed policy also scores 0.5712 and matches the adaptive run’s 52,145 optimizer updates, controlling for the total number of updates (Appendix L.2, Table 17).

Hyperparameter sensitivity and the Infinity‑8B scheduler. We added a sweep over five learning rates and four step budgets on 400 prompts to show how composition and image quality vary across settings beyond the defaults (Appendix L.1, Figure 6). We selected learning rates and step budgets on T2I-CompBench training subsets and ran the component ablation on validation prompts.

The scheduler checks gradient norms before per-token normalization. On Infinity‑8B, the raw norms fell below the Infinity‑2B threshold of 10⁻³, so we lowered it to 10⁻¹² to allow updates. At that setting, Infinity‑8B effectively took all five steps per eligible scale. The reported efficiency gains are for Infinity‑2B (Appendix L). Our compute budget precluded a threshold sweep.

Thank you again for your review. We added experiments addressing your concerns about baseline comparisons, adaptive versus fixed scheduling, and transfer beyond Infinity (Generality). With discussion closing soon, we’d appreciate your reading our response and reconsidering your recommendation in light of the new evidence.

Reviewer_hBCBOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes VISTA, a training-free framework for improving compositional alignment in visual autoregressive generation. Built on Infinity, VISTA optimizes the continuous visual inputs at selected generation scales using gradients from cross-attention-based objectives for attribute binding, 2D spatial relations, and depth ordering, without updating model parameters. Experiments on Infinity-2B and Infinity-8B demonstrate improvements on T2I-CompBench and GenEval, particularly for spatial relations. Human evaluations further support the improvements in compositional alignment, with no statistically significant difference in perceived image quality.

Critical strengths

1. Simple and intuitive approach. The paper presents a straightforward test-time optimization framework that uses cross-attention maps and compositional objectives to optimize continuous visual inputs at selected generation scales. The overall method is intuitive and easy to understand.

2. Consistent empirical improvements. Experiments on Infinity-2B and Infinity-8B demonstrate consistent improvements across the targeted compositional categories, with substantial gains over the original backbones, particularly for spatial relations. Human evaluations further support these improvements in compositional alignment.

Critical weaknesses

1. Limited methodological novelty. The core pipeline of optimizing visual representations using cross-attention-based objectives closely resembles existing diffusion guidance methods. Most of the objectives are adapted from prior work, although the paper introduces a new depth objective. The main contribution appears to be adapting these techniques to Infinity, while components such as cache isolation are primarily engineering details rather than substantial algorithmic innovations. Overall, I find the methodological novelty limited.

2. Insufficient ablation studies. VISTA introduces several optimization strategies tailored to Infinity, including per-token gradient normalization, scale-aware step sizes, and adaptive step allocation. However, the paper lacks direct ablations to isolate their individual contributions, making it difficult to assess the necessity and effectiveness of these designs.

Questions for authors
  1. Generality. Could the authors further validate VISTA on other visual autoregressive architectures, such as HART, to demonstrate its generality beyond Infinity?

  2. Ablation studies. Could the authors provide ablations to validate the effectiveness of the proposed optimization strategies, particularly per-token gradient normalization, scale-aware step sizes, and adaptive step allocation?

  3. Comparison with diffusion guidance. Although Table 1 includes results from diffusion models, the paper lacks direct comparisons with existing diffusion-based compositional guidance methods, such as CONFORM. Could the authors provide comparisons with these methods under the same evaluation benchmarks, including the performance gains relative to their respective backbones?

  4. Untargeted constraints. Numeracy and Non-Spatial relations are not explicitly optimized in VISTA. Could the authors investigate whether the framework can be extended to these constraints by introducing additional objectives? Also, does optimizing the existing constraints affect these abilities when multiple constraint types appear in the same prompt?

I am open to updating my assessment if the authors can satisfactorily address these concerns.

4 comments

Thank you for the concrete suggestions. We added controlled optimization ablations, a second backbone, a diffusion-guidance comparison, and experiments on counting and mixed constraints.

Methodological contribution. Our work addresses a technically challenging problem because the discrete, multi-resolution, and stateful nature of visual autoregressive generation makes directly transferring existing alignment methods particularly difficult. To our knowledge, VISTA is the first to transfer compositional cross-attention losses originally defined for diffusion models to discrete next-scale visual autoregressive generation. We demonstrate this with CONFORM for attribute binding, PSG for planar relations, and CountCluster for numeracy. Each scale predicts a larger discrete token grid conditioned on earlier scales. VISTA differentiates through the frozen transformer to update the continuous input before sampling that grid, so the intervention can change the tokens that condition all finer scales. In Infinity, auxiliary forwards can also modify the cache that later scales read. Cache isolation preserves the state required for ordinary sampling to resume. Sections 3.1–3.4 explain this procedure, and Appendices E–F show that the same optimization machinery supports alternative objectives and counting guidance.

Optimization ablations. We ran six variants on 1,800 CompBench validation prompts with two seeds (Appendix L.2, Table 17). Replacing per-token normalization with global L2 normalization or raw gradients reduces the targeted mean from 0.5733 to 0.5318 or 0.5283, respectively. This is the clearest evidence that preserving the strength of individual token updates matters. To test scale-aware steps, we gave the constant-step variant a calibrated step size based on the full method. It scores 0.5698, indicating a modest benefit from varying the step size across scales. Per-token normalization has the largest effect on composition. Both fixed-budget policies score 0.5712, close to Full’s 0.5733 without using the gradient threshold. This includes the policy matched to Full’s 52,145 updates. Adaptive scheduling offers a smaller efficiency gain, using 3.3% fewer updates and 1.8% less generation time than taking five updates at every eligible scale.

Transfer beyond Infinity. We chose Switti to test whether VISTA transfers across architectures that share the core VAR generation process. Both Switti and Infinity predict discrete token grids from coarse to fine, but they use different tokenizers and self-attention designs. Switti uses current-scale self-attention and retains explicit image-to-text cross-attention, which provides the same kind of steering signal used by compositional guidance in text-to-image diffusion models.

HART is a hybrid model. It generates discrete tokens autoregressively and uses a lightweight diffusion module to model continuous residuals. Its autoregressive transformer uses text tokens as context in self-attention. Extending VISTA to HART would therefore require adapting the objectives to an attention interface where text and image tokens share the same attention operation. HART remains untested in this work.

With three steered scales, VISTA raises Switti’s seven-category T2I-CompBench mean by 13.3%, including Numeracy. Its three-category GenEval mean rises by 36.4%, driven by gains in Position and Color Attribution, while Colors declines (Tables 1–2 and Appendix G). The transfer shows that the gains extend beyond Infinity’s tokenizer and cache design when the backbone exposes explicit text cross-attention.

Diffusion-guidance comparison. We ran CONFORM on two diffusion backbones and report each guidance method’s gain over its own unguided backbone on the same four CompBench categories: Color, Texture, Shape, and Complex (Appendix H, Table 12). CONFORM improves SD 1.5 from 0.3808 to 0.5629 and SD 2.1 from 0.4353 to 0.5946. VISTA improves Infinity‑2B from 0.5600 to 0.6385. Reporting both endpoints makes the different starting scores visible. The comparison shows that attention-based guidance improves composition within each tested backbone and puts VISTA’s gain in context under the respective generation protocols.

Numeracy extension. For numeracy and non-spatial relationships, attention-based loss functions remain relatively unexplored in the literature and there are no widely adopted loss formulations to capture these types of relationships. However, we added CountCluster (Lee et al., 2025, arXiv:2508.10710) through the existing objective interface and tested it on Infinity‑2B and Switti. At three steered scales, Infinity‑2B Numeracy improves by 6.6% relative to the backbone over four seeds. Switti shows a smaller relative gain of 0.7% (Table 1 and Appendices F–G). The Infinity‑2B result shows that extending VISTA to counting can yield a measurable gain using the same steering machinery.

Joint binding and planar steering. For prompts combining two color–object bindings with a planar relation, joint steering saves 15.4–16.1% generation time relative to the two sequential orders. Joint success is 40.8%, versus 42.1% and 45.8% sequentially. Both paired confidence intervals include zero and allow for a success-rate cost (Appendix L.3, Table 18).

Thank you for the detailed rebuttal and extensive additional experiments. The new ablation studies have substantially addressed my concerns regarding empirical validation and generality. I also appreciate the additional analysis of numeracy and mixed constraints. While I still have some reservations about the methodological novelty, I acknowledge the technical challenges of adapting compositional guidance to visual autoregressive generation and the value of the proposed framework. Overall, the rebuttal has improved my assessment of the paper, and I am willing to raise my rating to Weak Accept.

Thank you for taking the time to read the revision and update your assessment. We’re glad the new evidence helped address your concerns, and we appreciate the feedback that guided the revision.

Reviewer_M7MoOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes VISTA, a training-free test-time alignment framework for Visual Autoregressive (VAR) models to improve the compostional alignment in image generation. This is achieved by optimizing intermediate representations through gradient-based cross-attention guidance, which consists of 3 compositional losses, cache isolation and loss stablizing mechanism. Experiments on several benchmarks demonstrate significant improvements, particularly in spatial relationships.

Critical strengths
  • Compositional alignment is an important and well-motivated problem, and addressing it can significantly improve the performance and reliability of image generation models.
  • The authors provide insightful analysis of compositional issues from the perspective of cross-attention, offering a meaningful approach to improving text-image alignment.
  • The authors introduce a series of effective techniques to stabilize gradient-based optimization with multiple compositional losses in the multi-scale autoregressive generation process.
Critical weaknesses
  • The compositional losses adopted in this paper are largely based on existing methods. Therefore, the novelty and contribution of this part are relatively limited.
  • The paper lacks a detailed analysis of how compositional errors are reflected in cross-attention maps and whether the proposed method actually corrects these attention patterns as intended. Visualizations of attention maps alongside generated examples would help validate the assumptions and mechanisms described in the method section.
  • Although the proposed method demonstrates improvements on compositional benchmarks, some visual quality metrics, such as LAION Aesthetic Score, show a slight decline, suggesting a potential trade-off between compositional alignment and visual fidelity. More detailed analysis and qualitative examples would help explain how and why the proposed optimization affects visual quality.
  • It remains unclear whether the proposed method is fundamentally limited to next-scale VAR models or can be generalized to other autoregressive image generation architectures. If the framework is not inherently restricted to next-scale generation, experiments on a broader range of VAR-based models would be necessary to demonstrate its generalizability.
Questions for authors

Please refer to the weaknesses part.

6 comments

Thank you for the constructive review. We added quantitative and qualitative attention evidence, a clearer account of the quality trade-off, and experiments on Switti.

Contribution and objective provenance. Our work addresses a technically challenging problem because the discrete, multi-resolution, and stateful nature of visual autoregressive generation makes directly transferring existing alignment methods particularly difficult. To our knowledge, VISTA is the first to transfer compositional cross-attention losses originally defined for diffusion models to discrete next-scale visual autoregressive generation. We demonstrate this transfer with CONFORM for attribute binding, PSG for planar relations, and CountCluster (Lee et al., 2025, arXiv:2508.10710) for numeracy. Each scale predicts a larger token grid conditioned on earlier scales. VISTA differentiates through the frozen transformer to update that scale’s continuous input before sampling the discrete tokens. In Infinity, the auxiliary optimization passes preserve the cache that later scales read. We clarify this in Sections 3.1–3.4 and test the stabilization choices in Appendix L.2, Table 17. Appendix E substitutes alternative losses under the same optimization machinery, and Appendix F adds a CountCluster-based counting objective. These experiments show how the interface can be used to test new objectives on next-scale generators.

Attention errors and corrections. Figure 1 shows a matched example where the crow and flower attention maps overlap in Base and separate under VISTA. The Base image contains two birds, while the VISTA image contains both requested objects. Figure 5 adds six examples with their attention maps.

We also measured cross-attention during the final sampling forward pass on 600 matched Base–VISTA prompts (Appendix I, Table 13). At the unsteered 16×16 grid, cosine similarity between different nouns’ attention maps fell by 2.8%, 2.5%, and 4.7% for Color, Shape, and Texture, respectively, relative to Base. Two maps have high cosine similarity when they place attention on the same image locations in similar proportions, even if one noun receives more attention overall. The binding margin measures how much more closely an attribute’s attention map matches the noun it describes than the other nouns in the prompt. For example, in “a red apple and a green pear,” the margin for “red” is its cosine similarity to the “apple” map minus its cosine similarity to the “pear” map. When there are several other nouns, we average their similarities. The binding margin increased by 35.6%, 70.6%, and 53.6% for Color, Shape, and Texture, respectively. These changes indicate better separation between nouns and stronger alignment of attributes with their intended nouns.

For planar relations, we checked whether the two noun maps’ attention centroids followed the requested left/right or above/below order. Accuracy increased by 61.4% relative to Base on 72 directional prompts. The signed centroid-gap change correlated with the image benchmark score change (Spearman =0.31), and 11 of 27 prompts with improved centroid order showed no score gain. On 3D Spatial, global soft IoU between noun attention maps fell by 16.1%, and the category’s image benchmark score rose by about 14% in the attention audit. The lower overlap shows that VISTA separates the noun attention maps more strongly. The separate depth-loss comparison found that boundary containment scored highest among the five objectives tested (Appendix E, Table 8, and Appendix I, Table 13).

Visual quality and the choice of scales. At the default three steered scales, both metrics that take the prompt into account improved: ImageReward by 20.1% and CLIPScore by 1.1%. The two LAION aesthetic predictors, which score the image without the prompt, declined by 1.3–2.2% (Appendix K). In the 200-pair human study, VISTA won 83.3% of decisive compositional comparisons. For perceived visual quality, we observed 52 VISTA wins, 45 Base wins, and 103 ties, with no significant preference ((p=0.543), Table 4 and Appendix J). Figure 3 and Table 3 show why we stop at three scales: adding a fourth increased runtime by 15.0%, improved Infinity‑2B’s targeted mean by 0.7%, and reduced Infinity‑8B’s by 1.2% at the tested settings.

Generality across autoregressive architectures. Switti and Infinity share the core VAR generation process. Both generate discrete token maps from coarse to fine, with each scale conditioned on coarser content. They differ in architecture: Switti uses a different tokenizer and restricts self-attention to the current scale, while Infinity also retains keys and values across scales (Section 2).

With three steered scales, VISTA raises Switti’s seven-category T2I-CompBench mean by 13.3% across four seeds, including Numeracy with counting guidance enabled (Table 1 and Appendix G). Its three-category GenEval mean rises by 36.4%, with gains in Position and Color Attribution and a decline in Colors (Table 2). This demonstrates transfer across two architectures that share next-scale generation. Other autoregressive generation schemes remain untested.

Thank you again for your review. We added experiments and analyses addressing your concerns about cross-attention patterns, visual quality, and transfer beyond Infinity (Generality). With discussion closing soon, we’d appreciate your reading our response and reconsidering your recommendation in light of the new evidence.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Manuscript Update

Thank you for the feedback. We revised the manuscript and summarize the main changes below.

Transfer and constraint coverage. We added Switti results on T2I‑CompBench and GenEval. With three steered scales, VISTA improves Switti’s seven-category CompBench mean by 13.3% and its three-category GenEval mean by 36.4%. We also added CountCluster-based numeracy guidance for Infinity‑2B and Switti, plus a separate audit of Numeracy and Non-Spatial performance with counting guidance disabled (Tables 1–2 and Appendices F–G).

Attention analysis. We added six paired image and attention-map examples and a 600-prompt audit. The audit measures noun-map cosine similarity, the attribute-binding margin, and centroid ordering at the unsteered 16×16 scale. It also reports cases where better centroid ordering did not improve the image benchmark score (Appendix I, Figure 5 and Table 13).

Optimization and mixed constraints. We added controlled ablations of per-token normalization, scale-aware step sizes, and adaptive versus fixed step budgets on Infinity‑2B. We also compared joint and sequential binding–spatial steering, reporting generation time and joint success (Appendix L.2–L.3, Tables 17–18).

Baseline comparisons. Table 1 reports the corrected eight-candidate BoN run and four-seed TTS‑VAR results. The appendices add within-backbone CONFORM comparisons on two diffusion models, a four-seed SSG reproduction on Infinity‑2B, and VPG’s published GenEval results under its own evaluation protocol (Appendices H, M, and N).

Method scope, tuning, and quality. We clarified how VISTA adapts compositional cross-attention losses to discrete next-scale generation and documented the role of the parser and cross-attention. We added a learning-rate and step-budget sweep and visualized how the number of steered scales affects compositionality, prompt-conditioned quality, aesthetics, and runtime (Figures 3 and 6).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.