- Rating
- 3(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes VPG, a training-free inference-time guidance for next-scale visual autoregressive models. Where CFG contrasts along the condition axis, VPG contrasts the prediction under the generated prefix with the prediction under its corrupted copy (same-scale full-embedding replacement), extrapolates the logits, and interprets this as prefix-axis log-ratio guidance. It needs no extra model or retraining, composes with CFG, and reports gains on VAR, Infinity, and InfinityStar. The method seems promising, yet two major concerns remain. First, the comparison with existing training-free guidance uses quoted numbers from a different environment, so whether the prefix axis actually adds something beyond existing guidance is not shown. Second, exposure bias is the stated motivation but it is never measured, and by its mechanism VPG may amplify errors in the prefix rather than prevent them.
Critical strengths
-
The practicality of VPG is strong, as it is training-free, needs no auxiliary model, applies to three public checkpoints directly, and may work harmoniously with CFG.
-
The VAR results are robust. Tab. 9 reports five-seed statistics, with selected on seed 42 (as it seems only seeds 43-46 are used to claim that the chosen is robust) and the gain holding on the four held-out seeds. Tab. 5 rules out alternative references, and Fig. 11 shows this is not just added diversity.
Critical weaknesses
-
The comparison with existing training-free guidance is concerning because of a possibly unfair comparison. The paper picks SSG as the only training-free comparison in Tab. 2, but the SSG numbers are values quoted from the SSG paper. The official VAR checkpoint is known to give worse FID than the VAR paper reports, so SSG measured its own VAR baseline, and this paper also runs its own CFG baseline, which differs again (for d24, 2.39 for SSG against 2.16 here). So the two methods sit on different baselines, and absolute FID comparison is meaningless. At d24 the paper's own CFG baseline (2.16) is already below the SSG result (2.20), so "outperforms SSG on four of five depths" is just a pipeline difference. SSG adds no forward pass, while VPG in Tab. 2 costs Transformer FLOPs, and the table has no computation cost column. The code of SSG is public, so a same-pipeline, compute-matched comparison seems feasible. IGG [1], a token-adaptive guidance for VAR, belongs in the same discussion as well. Since the novelty claim is the prefix axis, a fair comparison is required to judge whether this method actually outperforms the existing ones.
-
The motivation and the mechanism do not match. The paper motivates VPG with exposure bias, but there are two problems. First, no experiment measures exposure bias itself. There are no error propagation measurements or related experiments. The evidence is only FID, DPG, and VBench, so "exposure bias is reduced" is not something the paper shows. Instead, it reads more like an unsubstantiated explanation put on top of a better score. Second, the mechanism may go opposite to the intention of the authors, as Eq. (8) favors candidates that are more faithful to the generated prefix. If the prefix is already wrong (or contaminated), VPG pushes later scales to be more consistent with that wrong prefix. That is the error propagation of exposure bias itself, not its prevention. No experiment separates the two. The same-seed comparison in Fig. 12 has VPG on from the start, so it cannot separate them either. From my perspective, what the paper actually shows is that conditioning more sharply on the prefix improves FID.
Questions for authors
-
Can you reproduce SSG in the same pipeline, with the same seeds and matched compute? Could you also compare with IGG [1] under the same setting?
-
Given that the T2V result is single-seed with no variance, can you add a second T2I benchmark such as GenEval, and multi-seed VBench? Also, given that Tab. 13 shows all-scale VPG gives almost nothing (83.91 vs. 83.86) and only the coarse-scale schedule reaches 84.22, how was this setting chosen, and was it selected on the same VBench run reported in Tab. 4?
-
Given that the gain shrinks from 0.622 at d16 to 0.121 at d30, and VPG on d30 ends up the same as VPG on d24, is there any result on a larger or stronger backbone that addresses the scalability of the method?
-
Can you inject controlled errors at coarse scales and measure the downstream effect with and without VPG, so as to separate whether VPG prevents error propagation or amplifies it?
[1] Rethinking Visual Autoregressive Sampling with Information-Grounding Guidance (ICML 2026)
7 comments
Response to Reviewer NyFA (2/3): W2 and Q4
W2 / Q4 (motivation vs. mechanism; controlled errors).
Thank you for this point and for the proposed test. The concern is that the paper never measures exposure bias, and that VPG may amplify errors in the prefix rather than prevent them. Your test confirms the second point for an error that is already committed: VPG follows a planted error more closely than CFG. What we can now measure in support of the motivation is: the prefixes that VPG generates are closer to those of real images than CFG's. This is one aspect of dealing with exposure bias and does not show reduced error propagation.
Exposure bias is the mismatch between the ground-truth (teacher-forced) histories seen in training and the generated histories seen at inference. The two experiments below measure different parts of it: the distributional analysis measures how far the prefixes VPG generates drift from teacher-forced ones, and the controlled-error test measures how VPG responds to an error that is already in the prefix.
(i) Distributional analysis. We decode the latent after each scale and compare the generated partial images with the same partial decodes of real ImageNet images encoded by the VAR tokenizer, i.e., with teacher-forced prefixes (FD-DINOv2 [2] (lower is closer), mean ± s.d. over seeds 42–44 with 10k samples each, per token map):
| FD ↓ per token map | 2×2 | 3×3 | 4×4 | 5×5 | 6×6 | 10×10 | 16×16 (final) |
|---|---|---|---|---|---|---|---|
| CFG | 32.3 ± 0.6 | 33.2 ± 0.1 | 38.1 ± 0.5 | 45.4 ± 0.1 | 57.9 ± 0.9 | 186.9 ± 0.7 | 474.5 ± 1.0 |
| CFG, T = 1.1 | 34.9 ± 1.0 | 35.0 ± 0.2 | 39.7 ± 0.4 | 46.3 ± 0.6 | 55.8 ± 1.1 | 182.0 ± 1.3 | 466.6 ± 1.7 |
| CFG, scale 1.25 | 33.0 ± 1.1 | 33.4 ± 0.3 | 38.4 ± 0.7 | 46.0 ± 1.0 | 57.5 ± 0.6 | 180.8 ± 1.2 | 461.1 ± 1.9 |
| SSG (T = 1.15) | 34.0 ± 0.5 | 32.3 ± 0.3 | 36.5 ± 0.4 | 41.8 ± 0.3 | 52.5 ± 0.9 | 180.6 ± 1.0 | 469.7 ± 2.4 |
| VPG, all scales | 26.3 ± 0.4 | 26.3 ± 0.4 | 31.7 ± 0.6 | 41.2 ± 0.1 | 57.5 ± 0.4 | 183.3 ± 1.2 | 447.4 ± 2.6 |
| VPG, coarse-scale schedule | 25.5 ± 0.4 | 25.3 ± 0.4 | 31.1 ± 0.9 | 40.3 ± 0.4 | 55.5 ± 0.3 | 178.4 ± 0.3 | 442.1 ± 2.0 |
On the 2×2–4×4 maps, only VPG moves the prefixes clearly toward teacher-forced ones. Both VPG settings lower FD there by 6–8, and the coarse-scale schedule (with larger λ) also gives the closest final images. The three controls also add diversity and lower the final FID, yet on these maps they stay within 2.6 of CFG. A higher temperature on the coarse maps moves them further from teacher-forced ones (4×4: 39.7 ± 0.4 at T = 1.1 against 38.1 for CFG).
Most of the FID gain comes from the coarse scales (VPG switched on at chosen scales, paired random numbers, CFG reference 3.40):
| VPG applied on maps | all | 2×2–8×8 | 2×2–4×4 | 10×10–16×16 |
|---|---|---|---|---|
| FID ↓ | 2.73 | 2.88 | 3.10 | 3.29 |
The 2×2–8×8 maps hold 23% of the tokens but give 77% of the gain, and the 2×2–4×4 maps give 45% of the gain. Across depths, the reduction in coarse-map drift is largest on d16 and absent on d30 (mean over the 2×2–4×4 maps, CFG − VPG, three seeds):
| d16 | d20 | d24 | d30 | |
|---|---|---|---|---|
| Drift reduction | 6.4 | 2.7 | 2.9 | −0.6 |
| FID gain (Tab. 9) | 0.62 | 0.45 | 0.32 | 0.12 |
Larger models already produce coarse prefixes closer to teacher-forced ones (CFG FD on the 2×2 map: 32 on d16, 28 on d30). On d30, VPG still lowers FID by 0.12 without reducing coarse drift, so drift does not account for the whole gain.
(ii) The controlled test. It isolates an error that is already committed: VPG is switched on only afterwards. We sample with CFG up to the 4×4 (or 8×8) scale, replace 30% of its codes with random codes, and continue the same prefix with CFG or with VPG.
| Injected at | Continuation | FID ↓ clean → injected | Distance to clean twin |
|---|---|---|---|
| 4×4 | CFG | 3.40 → 3.26 | 0.281 |
| 4×4 | VPG (scales 5–10) | 3.17 → 3.60 | 0.311 |
| 8×8 | CFG | 3.40 → 3.43 | 0.182 |
| 8×8 | VPG (scales 8–10) | 3.29 → 3.66 | 0.209 |
Distance: DINOv2 cosine distance (10k pairs). Paired difference VPG − CFG: +0.030 [0.027, 0.033] and +0.027 [0.025, 0.029]. We use the paired distance, which compares each sample with its uninjected twin. VPG follows the planted error more closely than CFG, and its FID ends up above CFG's. Eq. 8 implies this: it favors what the prefix supports and cannot tell that a planted code is wrong.
So the answer to Q4 is that VPG does not prevent the propagation of an error already in the prefix. It follows it more closely, and correcting such errors is what training-time remedies (scheduled sampling, noisy-context training) address. What the measurements support is: in normal sampling, VPG's coarse prefixes are closer to those of real images; i.e., on these maps VPG narrows a mismatch between generated and teacher-forced prefixes that relates to exposure bias.
The Introduction and Limitations now state this scope: VPG reduces prefix drift during sampling and does not reduce an error already present in the prefix. App. D.5 reports both experiments.
Response to Reviewer NyFA (3/3): Q2 and Q3
Q2 (multi-seed VBench; schedule selection; GenEval).
The questions are whether the video gain holds beyond a single seed, how the schedule was chosen and whether it was chosen on the reported run, and whether the text-to-image gain carries over to GenEval [3]. The outcome is mixed: the video gain is positive in all four seeds (+0.18 ± 0.13 on the three seeds not used to choose the schedule). On the three held-out seeds, the chosen schedule has the highest mean gain. GenEval does not improve, while the DPG-Bench gain holds in every seed and per-prompt diversity rises on every measure we used.
(i) Multi-seed VBench. On the three seeds not used to select the schedule, VPG raises VBench Overall over InfinityStar's original sampler without VPG by +0.18 ± 0.13, and the gain is positive in all four seeds (+0.22 ± 0.14 including the selection seed; Tab. 10, bottom). Sec. 5.2 now reports the four-seed result.
| InfinityStar-8B, four seeds | Quality ↑ | Semantic ↑ | Overall ↑ (mean ± s.d.) |
|---|---|---|---|
| Original sampler | 84.23 | 82.50 | 83.88 ± 0.10 |
| + VPG | 84.47 | 82.65 | 84.10 ± 0.22 |
(ii) Schedule selection. Yes, the schedule was selected on the seed-0 run of Tab. 4. On the three seeds not used for that choice, we ran all four combinations of guided scales and guided repeats. The chosen schedule has the highest mean gain (+0.18 ± 0.13, against +0.07 to +0.09) and is the cheapest (1.07× FLOPs), but the differences between the means (at most 0.11) are smaller than each schedule's s.d. (0.13 to 0.24), so three seeds do not show that one schedule is consistently best. Its gain comes mainly from dynamic degree and scene content, while temporal-consistency metrics change little (Sec. 5.2). On these seeds, the gain lies in the Quality score, and the seed-0 Semantic gain of Tab. 4 does not recur.
| Guided scales | Guided repeats | Transformer PFLOPs per clip ↓ | ΔOverall, mean ± s.d. over 3 seeds ↑ |
|---|---|---|---|
| first 12 of 15 (ours) | first | 21.76 (1.07×) | +0.18 ± 0.13 |
| all 15 | first | 35.66 (1.76×) | +0.09 ± 0.24 |
| first 12 of 15 | all | 24.66 (1.21×) | +0.09 ± 0.17 |
| all 15 | all | 40.61 (2.00×) | +0.07 ± 0.20 |
(iii) GenEval. VPG does not improve GenEval: Overall is 0.707 ± 0.007 with VPG, against 0.707 ± 0.006 without it (mean ± s.d. over five seeds; Infinity-2B with the λ and n_p of DPG-Bench and Infinity's prompt rewriter, 553 prompts, four images each). What VPG does change in text-to-image is sample diversity, consistent with its higher recall on ImageNet: images for the same prompt are more varied on every measure we used (CLIP distance [4] +6% on DPG-Bench prompts; DreamSim [5] +6% on OneIG-Bench [6]; LPIPS [7] +2.6%, higher on 98% of prompts). App. D.6 reports the GenEval categories (Tab. 20) and these diversity measures (Tab. 21).
Q3 (scalability).
Thank you for raising this point. The question is whether VPG's gain survives at larger scale, given that it shrinks from 0.62 FID on d16 to 0.12 on d30. At 256² the gain does shrink with depth, but it stays positive at every depth (no paired 95% interval includes zero; Tab. 9). As you note, VPG on d30 reaches the same FID as VPG on d24 (1.84 for both with all-scale VPG, and 1.83 on d30 against 1.76 on d24 with the coarse-scale schedule). VPG thus narrows the gap between the two depths (CFG: 2.16 on d24, 1.97 on d30) rather than preserving it, since its gain is smaller where the CFG baseline is already stronger. On the largest checkpoints we could test, the gain is clear on VAR-d36 (0.48 FID, 5/5 seeds) and positive in all four seeds on InfinityStar-8B.
On the larger backbones, VAR-d36 (2.3B) at 512² improves from 2.65 to 2.17 FID (Tab. 9) and is ahead of SSG (2.33 with the coarse-scale schedule and 2.17 on all scales, against 2.44). On InfinityStar-8B, the Overall gain is +0.18 ± 0.13 on the three held-out seeds.
At 256², with CFG alone, the coarse prefixes of larger models are already closer to teacher-forced ones (FD-DINOv2 [2] averaged over the 2×2–4×4 maps: 34.5 on d16, 30.5 on d30). VPG's reduction of this drift shrinks as well, from 6.4 on d16 to about 3 on d20–d24 and none on d30 (reply to W2/Q4). The correspondence is not exact. The drift reduction matches the FID gain at the two ends of the range but not between d20 and d24 (reductions 2.7 and 2.9, FID gains 0.45 and 0.32). On d30 VPG still lowers FID by 0.12 without reducing coarse drift, so drift does not account for the whole gain. References ([1, 2] in part 1/3)
[3] Ghosh et al. GenEval: An Object-Focused Framework for Evaluating Text-to-Image Alignment. NeurIPS 2023.
[4] Radford et al. Learning Transferable Visual Models From Natural Language Supervision. ICML 2021.
[5] Fu et al. DreamSim: Learning New Dimensions of Human Visual Similarity using Synthetic Data. NeurIPS 2023.
[6] Chang et al. OneIG-Bench: Omni-dimensional Nuanced Evaluation for Image Generation. arXiv:2506.07977, 2025.
[7] Zhang et al. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. CVPR 2018.
Response to Reviewer NyFA (1/3): W1 and Q1
Thank you for the technically substantive feedback. Our response is split into three comments: W1 and Q1 here, W2 and Q4 in (2/3), Q2 and Q3 in (3/3).
Thank you for the constructive feedback. The concern is that our advantage over SSG could reflect different pipelines and baselines rather than the method. In a same-pipeline comparison, which replaces the quoted SSG numbers, the advantage persists on four of the five models, and SSG is ahead on d30. The coarse-scale schedule costs 22% more Transformer FLOPs than SSG, and the revised Tab. 2 now reports this cost. SSG and IGG [1] were re-run in our pipeline with the same official checkpoints of VAR and 50k-sample ADM protocol, the same seeds (42–46), and the official VAR sampler (CFG 1.5 linear, top-k 900, top-p 0.96). SSG uses the official update rule from the released code. IGG follows the released code and Eq. 6 of its paper.
Each baseline was tuned on seed 42. VPG keeps the paper's settings: all-scale VPG uses the λ of Tab. 7, and the coarse-scale schedule the λ of Tab. 12. Tab. 12 has no d36 entry, so that λ (1.8) was chosen on seed 42, like the baselines. Tuning grids:
No selected setting lies at the edge of its grid.
Five-seed results (FID ↓, mean ± s.d.; Transformer FLOPs relative to CFG; the coarse-scale schedule costs 1.13× on d36, whose finest maps are larger). CFG and all-scale VPG are the five-seed means already in the paper (Tab. 9).
With every method tuned and run on the same seeds, VPG with the coarse-scale schedule lowers FID by 0.40–0.63 on d16–d24, where SSG lowers it by 0.24–0.36 and CFG with a tuned temperature by at most 0.24. VPG is ahead of these two baselines in each of the five seeds at each of these depths. On d36 VPG is best at both costs, and on d30 SSG is best. IGG stays within 0.06 of CFG, in line with the IGG paper's own VAR results (d30: 1.93 → 1.92; d36: 2.61 → 2.56).
Revised Tab. 2 replaces the quoted SSG column with these rows and adds the FLOPs row; App. C.3 gives the protocol and grids. Sec. 2 now discusses IGG.
References (for parts 1/3 and 2/3)
[1] Nguyen et al. Rethinking Visual Autoregressive Sampling with Information-Grounding Guidance. ICML 2026.
[2] Stein et al. Exposing Flaws of Generative Model Evaluation Metrics and Their Unfair Treatment of Diffusion Models. NeurIPS 2023.