acceptodds
Under review as a conference paper at ICLR 2027

Distilling Specialized Orders for Visual Generation

Abstract

Autoregressive (AR) image generators inherit the scalability and mature inference stack of large language models, yet most of them commit to a fixed raster-scan order. This choice precludes zero-shot inpainting, outpainting, and editing without retraining. Any-order AR models remove this restriction by learning to generate under arbitrary token permutations, but they must spread a finite capacity over factorizations and consequently trail fixed-order models in sample quality. We introduce Ordered Autoregressive (OAR) generation, a self-distillation procedure that resolves this tension. Starting from a pretrained any-order model, OAR extracts a specialized, content-dependent generation order for every training image from the model's own confidence scores, and then fine-tunes the model on these image–order pairs. Order specialization reallocates capacity from the full permutation space to a single high-confidence trajectory per image, while leaving the any-order conditionals, and hence the model's flexibility, intact. On class-conditional ImageNet , OAR improves FID from 2.39 to 2.17 over its any-order baseline under an identical evaluation protocol, and outperforms a random-to-raster annealing schedule applied to the same backbone (2.34). The same model performs zero-shot inpainting and outpainting and surpasses the bidirectional MaskGIT and MAR models on both tasks. It also retains multi-token parallel decoding. OAR generalizes to text-to-image generation with a different, scratch-trained backbone (FID on Fashion Products and on Multimodal CelebA-HQ), and human raters prefer its samples over the baseline in 64% of comparisons. On ImageNet, the procedure adds a fine-tuning stage of at most of the pretraining epochs, and it requires no architectural change or extra annotation.

Primary area:
generative models
Keywords:
Image Generation, Autoregressive Models, Causal Generation, Any Order Generation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_QHYcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes OAR, a post-training procedure for any-order autoregressive image generators. A pretrained model assigns each training image a specialized generation order by greedily selecting confident positions under teacher forcing. The model is then fine-tuned on these image–order pairs, with optional order annealing. Experiments evaluate generation quality, arbitrary-order conditioning, and text-conditioned generation.

Critical strengths
  1. The same-backbone, matched-budget comparison provides useful evidence for specialized orders. On ImageNet, OAR achieves 2.17 FID, compared with 2.34 for annealed raster specialization and 2.39 for the original random-order baseline (Table 1). This isolates the choice of target order more convincingly than comparisons across different architectures.

  2. The paper evaluates whether specialization preserves the flexibility of the original model. Inpainting and outpainting performance remains close to RandAR, while the human evaluation provides complementary evidence: OAR is preferred in 16 of 21 randomly selected image pairs. The pair-level sign test appropriately accounts for the limited number of evaluated pairs.

Critical weaknesses
  1. The inference-cost tradeoff is insufficiently characterized for the main ImageNet model. Greedy decoding queries all remaining positions at each step. Parallel evaluation and KV caching can reduce latency, but do not establish that its cost is comparable to a prescribed order. Table 9 reports latency only for the separate six-layer text-conditioned model. A matched RandAR-XL timing comparison is needed to assess the practical value of the main quality improvement.

  2. The inpainting/outpainting comparison against MAR and MaskGIT lacks important protocol details. Table 3 gives the mask-coverage range, but does not fully specify mask construction, the MAR variant/checkpoint, or task-specific sampling settings for the baselines. The results support retention relative to RandAR more clearly than they establish the broader cross-model superiority.

Questions for authors
  1. What are OAR and RandAR-XL’s latency, throughput, and peak memory under identical hardware, precision, batch size, and decoding settings? Results at 256 and 128 steps would clarify the quality–cost tradeoff.

  2. Which checkpoints and sampling configurations were used for Table 3? How were masks constructed, and was masking performed before image encoding or directly on latent tokens?

  3. In Appendix C.3, what justifies cancelling the marginal V-entropy terms between Equations (6) and (7)? Content-dependent ordering can change the marginal distributions of the selected-token random variables. Please clarify the conditioning on positions, or distinguish the measured cross-entropy difference from the claimed V-information difference.

3 comments
AuthorsOpenReview ↗

Runtime Results, Completion Protocol, and V-Information

Thank you for the thoughtful review and for recognizing the controlled evidence and retained flexibility. Your questions help us clarify practical costs, improve reproducibility, and sharpen the theoretical interpretation.

1. Matched RandAR-XL inference cost.

We benchmarked both 775M implementations on one 80GB A100 or H100: bf16, PyTorch 2.10.0+cu128, CFG 1.75, temperature 1, top-k 0, top-p 1, seed 0, and identical classes. Both use 256 steps or 128 with the same cosine parallel schedule. Synchronized generation timing includes KV allocation and excludes VQ decoding. Timing uses random weights; trained quality is evaluated separately at CFG 3.4.

Mean batch latency in seconds:

256 steps

BatchA100H100
RandAROARRandAROAR
114.8415.456.997.48
214.8314.947.237.51
414.6815.077.137.61
814.7516.227.117.84
1614.8423.217.2311.61
3214.4640.457.2720.66

128 steps

BatchA100H100
RandAROARRandAROAR
17.457.583.653.69
27.467.603.643.70
47.387.683.653.77
87.368.573.644.21
167.4513.443.656.79
327.4224.353.6812.34

After two warmups, we used 15 repetitions through batch 16 and 10 at batch 32. SD was at most 2% of the mean, except A100 OAR at batch 1/256 steps (5%).

OAR/RandAR latency ratios are 1.01–1.16× through batch 8, 1.56–1.86× at 16, and 2.80–3.35× at 32 across both GPUs/schedules. Throughput is batch divided by latency; H100 batch-32 OAR/RandAR throughput is 1.55/4.40 images/s at 256 steps and 2.59/8.70 at 128.

Peak allocated memory, including weights, on A100 at 256 steps (GiB):

BatchRandAROAR
11.641.72
21.821.98
42.192.51
82.943.57
164.425.69
327.459.98

H100 memory differs by at most 0.02 GiB; 128-step memory is within about 0.03 GiB of the 256-step values.

At 256 steps, OAR/RandAR uses 33,407/512 token-forwards per image and an estimated 52.4/0.79 TFLOP (about 66×); at 128, 32.8/0.99 TFLOP (about 33×). Transformer forwards account for 96–98% of profiled time; confidence, selection, and sampling take 2–4%. Parallel queries use GPU parallelism: approximate profiled A100 batch-32 busy fractions are 53% for OAR and 12% for RandAR. Together with RandAR's eager-loop launch overhead, this helps explain the smaller latency ratio.

Table 4 separately reports OAR FID 2.17 at 256 steps and 2.25 at 128. We will add these cost measurements with their setup. The practical trade-off depends on batch size.

2. Completion protocol and interpretation.

Table 3 uses ImageNet 256×256 with shared random masks covering 25–75%. RandAR uses its released XL checkpoint; OAR fine-tunes that backbone. For MAR and MaskGIT, we used released ImageNet checkpoints and followed the default sampling settings in their repositories. All methods used the same images and masks. OAR conditions on visible tokens and fills masked positions greedily.

Thank you for highlighting the need for fuller protocol details. We will specify mask geometry, the encoding/masking sequence, model releases, guidance, temperature, steps, and schedules in the revision.

We agree that the controlled results support retention relative to RandAR: inpainting FID 2.10 versus 2.14 and outpainting 3.17 versus 3.21, without completion-specific training. We view these small differences as supporting evidence of preserved flexibility and will clarify the limits of the broader cross-model comparison.

3. Marginal V-entropy in Appendix C.3.

Thank you for identifying the missing conditioning. Content-dependent ordering can change the target marginal at a fixed decoding rank. We will replace the rank-wise argument with a pooled spatial-target comparison: sample J uniformly over the N positions, independently of the image and its text condition, and set Y=X_J. For each order π, define C_π as the text condition, target position J, and the ground-truth position–token prefix preceding J.

Both orders then predict the same target distribution: P(J=j,Y=y)=P(X_j=y)/N. For one common predictive family V satisfying optional ignorance, the shared marginal term cancels by construction:

=H_V(Y\mid C_r)-H_V(Y\mid C_s).$$ Here s denotes specialized and r raster order. This pooled identity replaces the sum of rank-specific information terms. We will distinguish the family-optimal V-entropy from fitted-model teacher-forced cross-entropy and state the scoring model and normalization of the measured difference. This retains V-information as the motivation for usable context under finite predictive capacity, with the empirical quantity labeled precisely. Thank you again. Your suggestions have helped us characterize specialization costs, identify missing protocol details, and clarify the information-theoretic argument.

Thank you for the detailed responses and additional experiments. I appreciate the authors' efforts to address my concerns, particularly the new inference-time measurements and clarification of the theoretical analysis.

The additional latency results provide a more complete picture of the trade-off. While the overhead is relatively small at low batch sizes, it becomes substantial at larger batch sizes (up to 3.35× latency overhead). The significant increase in token-forward computations also raises concerns about the practical efficiency of OAR, especially in high-throughput settings.

The clarifications regarding the completion protocol and V-information analysis are helpful. However, I believe the inference cost remains an important limitation that needs to be considered alongside the reported quality improvements.

Overall, the rebuttal has addressed several of my concerns, but the quality-efficiency trade-off is still not entirely convincing. I therefore maintain my original rating.

Thank you for the follow-up and for recognizing the additional measurements and theoretical clarification. Your feedback has helped us make the practical trade-off more concrete.

OAR improves the same RandAR-XL backbone from FID 2.39 to 2.17 while retaining zero-shot completion. The fine-tuning stage uses at most approximately 5% of the pretraining epochs, with the architecture unchanged.

The human study directly compares OAR against its pretrained RandAR-XL baseline in a controlled A/B test. Each pair contains one OAR image and one RandAR-XL image, generated with the same class, seed, guidance, and temperature on the same architecture. The pairs are uncurated, drawn from randomly sampled classes, and presented with randomized left/right placement. Across 51 raters, OAR is preferred over RandAR-XL in 64.3% of judgments and wins the majority vote in 16 of 21 pairs (two-sided pair-level sign test, p=0.027). This provides evidence that the gain is perceptible to people as well as reflected in FID.

Table 1 provides two useful controls. With the same backbone, fine-tuning budget, and annealing schedule, specialized orders achieve FID 2.17 versus 2.34 for raster specialization. Under the same greedy inference policy, fine-tuning improves FID from 3.64 to 2.17. Together, these comparisons support the value of the specialized trajectories and show that the trained model improves under a fixed decoder.

The matched A100/H100 benchmarks identify a relevant practical regime: OAR/RandAR latency is 1.01–1.16× at batches 1–8. At batch 32, it rises to 2.80–3.35×. We agree that the additional candidate-query computation is an important limitation for high-throughput generation and will make this batch dependence explicit in the revision.

OAR also retains parallel decoding: halving the steps from 256 to 128 changes FID from 2.17 to 2.25, an increase of 0.08 (Table 4). This preserves much of the full-step quality while allowing fewer decoding rounds.

We hope these controls and retained capabilities clarify OAR's contribution: specialized orders improve an existing generator through short post-training while preserving its completion flexibility. We will incorporate the cost measurements and their practical scope into the revision. If a focused additional analysis would help assess the remaining concern, we would welcome your guidance and be happy to provide it.

Reviewer_W5SrOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Ordered Autoregressive generation (OAR), which fine-tunes an any-order image generator on specialized orders extracted from its own confidence scores. Orders are extracted once using ground-truth contexts, while inference selects positions using generated contexts. On ImageNet 256 x 256, OAR improves the same RandAR-XL backbone from FID 2.39 to 2.17 under the paper's matched evaluation. Experiments compare alternative fine-tuning orders and assess zero-shot completion, parallel decoding, human preference, and text-conditioned generation with a second architecture on two datasets.

Critical strengths
  1. The method is simple and requires no architectural change. Table 1 provides a useful controlled comparison with random-to-raster fine-tuning on the same backbone.
  2. The paper evaluates generation, completion, and text-conditioned synthesis. The improvement over RandAR is also supported by the human preference study.
Critical weaknesses
  1. Choice of confidence score: Figures 1 and 4 show content-dependent orders, and Table 1 compares random, raster, and greedy sampling. However, these results do not isolate the choice of confidence score in Eq. (3). Why rank positions by maximum token probability rather than lowest predictive Shannon entropy? A brief explanation or a small comparison using the same candidate predictions at shared ground-truth prefixes would be sufficient.
  2. Inference cost: Section 4.4 describes parallel querying and KV caching, but greedy selection still evaluates the remaining positions. Table 4 reports decoding steps rather than runtime. What is the inference overhead relative to random-order RandAR on the 775M ImageNet backbone? Please report any existing timing comparison under the same hardware, batch size, and sampling settings. A short computation breakdown would also be useful.
  3. Scope of generalization: Section 6 evaluates separately trained text-conditioned models, and Section 7 acknowledges the restricted data distributions. These results support applying the procedure to other datasets, but do not show how a fixed checkpoint handles a domain shift. Could the authors discuss this distinction? If suitable, class-compatible examples are already available, a few paired completions from frozen OAR and RandAR, using identical inputs, masks, and sampling settings, could provide a qualitative check.
  4. Relation to learned-order models: Section 2 already cites LO-ARM [1], which learns a probabilistic order policy with a variational objective. OAR instead extracts greedy orders once and fine-tunes on them. Could the authors explain how the two training objectives differ, and what is gained or lost by fixing the extracted training orders? Please distinguish the cached training orders from the adaptive greedy order used at inference.

[1] Learning-Order Autoregressive Models with Application to Molecular Graph Generation, ICML 2025.

Questions for authors

Please refer to weaknesses.

3 comments
AuthorsOpenReview ↗

Confidence Ablation, Inference Costs, and Method Clarifications

Thank you for the thoughtful review and for recognizing our controlled comparisons and broad evaluation. We will add these results and clarifications to the revision.

1. Confidence criterion.

Following your suggestion, we reran entropy-based order extraction with matched annealing and budgets:

Extraction criterionFIDISPrecisionRecall
Maximum probability2.17258.370.800.60
Minimum Shannon entropy2.27248.120.780.61

Maximum probability gives 0.10 lower FID and higher IS/precision; entropy gives 0.01 higher recall. The max score measures peak confidence; entropy measures full-distribution uncertainty. Both are O(V) in vocabulary size. At V=16,384, direct entropy sums 16,384 p log p terms; max reduces shared probabilities. This is an operation-count comparison, without entropy-scoring timings. We will add the rationale and ablation.

2. Inference cost.

We measured OAR and random-order RandAR on the same 775M architecture with one 80GB A100 or H100: bf16, PyTorch 2.10.0+cu128, CFG 1.75, temperature 1, top-k 0, top-p 1, seed 0, and the same classes. Both use 256 decoding steps or 128 with the same cosine parallel schedule. Synchronized generation timing includes KV allocation and excludes VQ decoding. Timing uses random weights; submitted FID uses trained weights and CFG 3.4.

Mean batch latency in seconds:

256 steps

BatchA100H100
RandAROARRandAROAR
114.8415.456.997.48
214.8314.947.237.51
414.6815.077.137.61
814.7516.227.117.84
1614.8423.217.2311.61
3214.4640.457.2720.66

128 steps

BatchA100H100
RandAROARRandAROAR
17.457.583.653.69
27.467.603.643.70
47.387.683.653.77
87.368.573.644.21
167.4513.443.656.79
327.4224.353.6812.34

After two warmups, we used 15 repetitions through batch 16 and 10 at batch 32. SD was at most 2% of the mean, except A100 OAR at batch 1/256 steps (5%).

Across both GPUs, latency ratios are 1.01–1.16× through batch 8, 1.56–1.86× at 16, and 2.80–3.35× at 32. H100 batch-32 throughput (OAR/RandAR) is 1.55/4.40 images/s at 256 steps and 2.59/8.70 at 128. A100 batch-32 peak allocated memory is 9.98/7.45 GiB, including weights.

For OAR/RandAR at 256 steps, we count 33,407/512 token-forwards per image and estimate 52.4/0.79 TFLOP (about 66×); at 128 steps, 32.8/0.99 TFLOP (about 33×). Transformer forwards account for 96–98% of profiled time, and confidence, selection, and sampling for 2–4%.

OAR uses modern GPU parallelism by querying all remaining positions together. At batch 32, approximate profiled A100 GPU busy fractions are 53% for OAR and 12% for RandAR. Together with RandAR's launch overhead, this helps explain why the latency ratio is smaller than the FLOP ratio. We will report this batch-dependent trade-off.

3. Generalization scope.

We appreciate this distinction. Section 6 demonstrates applying OAR to separately trained models on other datasets; robustness of a frozen checkpoint under domain shift remains to be evaluated. Table 3's RandAR/OAR completions, using shared 25–75% masks within ImageNet, support retention without completion-specific training. We will clarify this scope and identify paired domain-shift evaluation as future work in Sections 6–7.

Future task-specific orders could incorporate expert knowledge of semantic or causal dependencies, offering control over generation [2] in healthcare or geospatial models. Current OAR uses confidence.

4. Relation to LO-ARM.

LO-ARM [1] jointly learns generation and ordering through an ELBO, combining token log likelihood with the log ratio of generation and variational order probabilities. The generation policy uses the prefix; the variational distribution uses the full example. Discrete paths use REINFORCE leave-one-out.

OAR caches one teacher-forced greedy order per image, then anneals fine-tuning from random toward these orders using next-token cross-entropy. This reuses supervision without discrete policy-gradient estimation, but the cached orders do not adapt to the improving generator. OAR does not learn an explicit stochastic order policy.

At inference, confidence is recomputed after each generation step using generated contexts, allowing the order to adapt. We will make the objectives and this training/inference distinction clearer in Sections 2 and 4.

Thank you again. We believe your suggestions have helped us strengthen the confidence comparison, quantify inference costs, and clarify OAR's scope and training objective. If accepted, we look forward to sharing these findings and discussing these extensions with the community.

[1] Wang et al., Learning-Order Autoregressive Models with Application to Molecular Graph Generation, ICML 2025.

[2] Zhang et al., Adding Conditional Control to Text-to-Image Diffusion Models, ICCV 2023.

Halton-MaskGIT[1] is closely related to this work in terms of generation order. That study found that using a Halton sequence can outperform confidence-based sampling—a result that appears to contradict your logic of "prioritizing high-confidence tokens." Table 1 currently lacks results for OAR using a random sampling order within an annealing strategy. Could you please conduct an experiment—using the same final annealing checkpoint and keeping the number of decoding steps and other evaluation settings constant—to compare the performance of a random order and a Halton sequence against your proposed greedy order? If such a controlled experiment demonstrates that your proposed order outperforms these alternatives, I would be willing to increase my score.

[1] Halton Scheduler for Masked Generative Image Transformer. International Conference on Learning Representations (ICLR), 2025.

Thank you for highlighting Halton-MaskGIT and proposing this useful controlled comparison. Your suggestion has helped us strengthen the evidence for the inference-order choice. On the same final annealed OAR checkpoint, greedy decoding achieves FID 2.17, compared with 2.32 for Halton and 2.41 for random order.

Following your suggestion, we held the final checkpoint and evaluation settings fixed: ImageNet 256×256, 256 decoding steps, CFG 3.4, temperature 1.0, and 50,000 generated samples. We additionally evaluated the final non-annealed checkpoint:

Final OAR checkpointGreedy FID ↓Halton FID ↓Random FID ↓
With annealing2.172.322.41
Without annealing2.222.392.44

Within each row, the model weights and evaluation settings are fixed; only the decoding order changes.

Greedy achieves the lowest FID on both evaluated checkpoints. On the final annealed checkpoint you requested, it improves FID by 0.15 over Halton and 0.24 over random order. Halton itself improves over random on both checkpoints, making your suggested comparison a useful stronger test of our decoding choice.

We appreciate the connection to Halton-MaskGIT's dependency analysis. The shared motivation is to account for visual-token dependencies when choosing an order. Nearby tokens often have strong local dependencies. Halton spreads updates spatially because tokens sampled together from the same context do not condition on one another's sampled values. OAR's 256-step causal AR decoder conditions each prediction on the tokens already committed, allowing nearby or semantically related content to inform subsequent predictions. Intuitively, confidence can favor positions well constrained by this context. This motivates confidence as a proxy for conditional predictability within the causal AR factorization.

OAR also preserves the practical benefits of its causal decoder: it caches the committed context and processes new queries at each step, while MaskGIT uses full-sequence bidirectional passes during iterative masked decoding. This retains KV caching and compatibility with established LLM decoder inference optimizations, including mixed precision and optimized attention kernels.

The results support the intended alignment between specializing on confidence-guided trajectories and using the same criterion at inference. The annealed checkpoint's random-order FID of 2.41 also remains close to the pretrained RandAR-XL baseline of 2.39. Together, these findings strengthen OAR's contribution: improved fidelity through order specialization while retaining arbitrary-order flexibility and the causal decoder architecture.

We will incorporate this comparison and the Halton-MaskGIT discussion into the revision. Thank you again for suggesting a direct test that makes the empirical case clearer and helps us better explain how the decoding order and model specialization work together.

Reviewer_A5BUOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a confidence-guided greedy self-distillation procedure that improves the fidelity of generated samples while preserving zero-shot editing behavior.

Critical strengths
  1. The proposed greedy-finetuning approach is cheap (<5% finetuning cost) and leads to improvement in FID values, without a significant degradation in the random-order performance and preserving the zero-shot editing behavior.
  2. The authors demonstrate that the proposed approach generalises to text-to-image generation.
Critical weaknesses
  1. The authors hypothesize that the fidelity issue primarily stems from model capacity allocation. However, the reported results show diminishing gains with increasing model size. The authors should evaluate the proposed approach on larger models, including RandAR-XXL, to understand whether these gains persist at scale. This is particularly important when scaling to text-to-image generation, where larger models are typically used, and the diminishing improvements raise concerns about the scalability of the proposed approach.
  2. (Minor) The paper is missing comparisons with recent works in AR Image generation [1,2,3]

[1] Srivastava, D., Mehra, A., Maneriker, P., Sanyal, D., Raj, V., Kamarshi, V., ... & Kimball, J. (2026). Dpar: Dynamic patchification for efficient autoregressive visual generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 23215-23226).

[2] Li, T., Sun, Q., Fan, L., & He, K. (2025). Fractal generative models. arXiv preprint arXiv:2502.17437.

[3] Pang, Y., Jin, P., Yang, S., Zhu, B., Lin, B., Feng, C., ... & Yuan, L. (2026, March). Next patch prediction for autoregressive visual generation. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 40, No. 10, pp. 8260-8268).

Questions for authors

My primary concern is the scalability of the proposed approach in real-world text-to-image generation setting, as the reported gains diminish with increasing model size. The authors should consider demonstrating that the gains hold with increasing model size using existing baselines like RandAR-XXL.

1 comment
AuthorsOpenReview ↗

Scaling Evidence and Recent AR Comparisons

We sincerely thank you for the thoughtful assessment and for recognizing OAR's fidelity gains and preservation of zero-shot editing. Your suggestions have helped us strengthen the empirical evidence and clarify the next steps for evaluating OAR at scale. We report new matched-budget experiments and recent-work comparisons below, which we will incorporate into the revision.

1. New matched-budget evidence across model sizes.

Following your suggestion, we applied OAR under matched comparison budgets to the public any-order LlamaGen-L/XL models from Unlocking Any-Order Generation in Pretrained Autoregressive Image Models [4] (accepted at NeurIPS 2026; released on arXiv on 29 October 2026):

BackboneParametersBaseline FIDOAR FIDAbsolute reductionRelative reduction
LlamaGen-L343M4.244.030.214.95%
LlamaGen-XL775M3.913.630.287.16%

XL gains 0.28 FID (7.16%), compared with 0.21 (4.95%) for L: the larger tested configuration improves more in both absolute and relative terms. Together with the submitted RandAR-XL gain (2.39 to 2.17), these results support specialized orders at both tested sizes and in another pretrained any-order framework.

Table 1 provides a controlled comparison with the same backbone, fine-tuning budget, and annealing schedule: OAR achieves FID 2.17 versus 2.34 for raster specialization. This supports the value of content-dependent trajectories beyond simply reducing the number of training orders. Our interpretation is that specialization focuses learning on useful conditionals. Capacity reallocation remains a hypothesis consistent with these controls; we will make its evidential scope explicit in the revision.

For text-conditioned generation, the submitted experiments improve FID from 4.58 to 2.56 on Fashion Products and 1.94 to 1.41 on Multimodal CelebA-HQ. These support applying the procedure in two controlled text-to-image domains.

We agree that RandAR-XXL and large, open-domain text-to-image models are valuable next tests. The official RandAR release lists L and XL checkpoints, with no XXL checkpoint [5], so we evaluated the released any-order L/XL models [4]. Our largest evaluated model remains 775M. The released L and XL configurations differ in native resolution and token-grid size; these results establish benefits across configurations, without isolating parameter count. We will include the new results and make these scope and checkpoint limitations explicit in the revision.


2. Positioning OAR within recent autoregressive work.

Thank you for these relevant and helpful suggestions. Autoregressive visual generation is advancing rapidly, with opportunities to reuse mature LLM inference infrastructure [4]. We will add DPAR [1], Fractal Generative Models [2], and Next Patch Prediction [3] to the final manuscript's discussion and comparison table. These additions will strengthen the paper's positioning and account of recent progress. Representative published ImageNet 256 results are:

MethodParametersReported FID
DPAR-XL, native 256, CFG 2.0 [1]789M2.67
FractalMAR-H [2]848M6.15
NPP-XL [3]775M2.65
OAR, submitted result775M2.17

These works address complementary modeling choices. DPAR uses entropy-guided dynamic patchification to reduce sequence length; Fractal Generative Models recursively compose generators across hierarchical levels; and NPP uses a coarse-to-fine patch-grouping curriculum during training while preserving the original inference procedure. OAR specializes content-dependent training orders within a pretrained any-order decoder, retaining its architecture and completion flexibility.

These published results provide context and use different backbones, representations, guidance, and training protocols. We will state these differences and keep the same-backbone comparisons central to the evidence for order specialization.

Thank you again. We believe your suggestions have helped us strengthen the evidence across model configurations, clarify the scaling questions that remain, and improve OAR's positioning within recent work. If accepted, we look forward to sharing these findings and discussing extensions to larger text-to-image models with the community.

[1] Srivastava et al., DPAR: Dynamic Patchification for Efficient Autoregressive Visual Generation.

[2] Li et al., Fractal Generative Models.

[3] Pang et al., Next Patch Prediction for Autoregressive Visual Generation.

[4] Pramanik et al., Unlocking Any-Order Generation in Pretrained Autoregressive Image Models, NeurIPS 2026; arXiv:2610.39293.

[5] RandAR official repository, ziqipang/RandAR, "Pre-trained Models" and linked Hugging Face release.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.