acceptodds
Under review as a conference paper at ICLR 2027

Structure-Anchored Conditional Response Steering for Human Video Generation

Abstract

Structurally conditioned human video generators already follow trajectories strongly constrained by pose, identity, and reference appearance, making unrestricted classifier-free guidance (CFG) prone to perturbing otherwise useful structural information. We revisit guidance from a conditional-response perspective and distinguish target-specific semantic and structural responses through selective changes of the conditioning configuration while keeping the remaining context fixed. Based on these axis-wise responses, we propose a structure-anchored steering formulation in which the model's own local response geometry determines where intervention is applied, while a bounded intervention budget regulates semantic steering in direction and diffusion time before the structural response is recomposed with controlled output magnitude. On Wan-Animate, our method achieves the best numerical results among the evaluated guidance methods on CLIP, T2V, MPJPE, aggregate VBench2 Human Fidelity Score, and both VideoScore2 visual-quality and physical-consistency criteria. Cross-backbone evaluation on other DiT-based algorithms further examines the transfer of the proposed guidance principle. These results suggest that guidance for structured human video generation is better formulated as controlled conditional-response steering than as unrestricted transformation of a single conditional discrepancy.

Primary area:
generative models
Keywords:
human video generation, classifier-free guidance, pose conditioning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_XwsGOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes structure anchored conditional response steering, it’s a training free human video generation approach. The method separates semantic and structural response and experiments show improvement over existing guidance methods.

Critical strengths
  • The paper is well written and well motivated, and it is tackling an important limitation in conventional CFG.
  • The method is effective and it’s used during inference time without requiring additional training cost.
Critical weaknesses
  • The proposed method seems to be a combination of a couple existing guidance approaches, which might limit its novelty.
  • There’s insufficient ablation studies and it is difficult to evaluate the effectiveness of different components.
  • The proposed conditional differences based on semantic and structural factors are not necessarily disentangled, and removing structural conditions can also change semantic predictions, and replacing text may influence pose or appearance. It might need additional justification for semantic structural decomposition.
Questions for authors

Could the authors clarify the novelty of the proposed method relative to existing guidance approaches, provide more comprehensive component-wise ablations, and justify whether the semantic and structural conditional responses are sufficiently disentangled to support the proposed decomposition?

1 comment
AuthorsOpenReview ↗

Response to Reviewer XwsG (3/3)

W3. Disentanglement of the two responses. We agree that the two responses are not exactly disentangled: the network is nonlinear, so removing the structure can shift semantic content and changing the text can affect pose or appearance. Sec. 2.2 states that "axis-wise" refers only to selective intervention on different conditioning sources and does not assume orthogonality or independence. Each response is evaluated at the same state xtx_t and changes one condition while holding the others fixed. The method is designed for this cross-talk rather than relying on its absence. The ablations show it directly: applying semantic steering uniformly (w/o ATA^T) leaves CLIP almost unchanged (29.11 vs. 29.20) but degrades MPJPE and identity, i.e., the semantic response does leak into structure and appearance. Localizing it with ATA^T, bounding its angle relative to the anchor, and renormalizing the output to ∥vB∥\|v_B\| limits this leakage. Conversely, removing the structural term changes CLIP by only 0.04, so the structural response does not noticeably alter semantics.

We hope these responses address the reviewer's concerns, and we would be grateful if the reviewer could consider updating the score. We are happy to clarify any further questions.

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

The paper revisits training-free classifier-free guidance (CFG) for structurally conditioned human video generation, evaluated primarily on Wan-Animate. Instead of amplifying a single conditional discrepancy, the authors construct two "axis-wise" responses at the same sampling state, i.e., a semantic response and a structural response. The authors derive a token-wise allocation field from the local angular change induced by each contrast, and apply a motion-budgeted angular and temporal semantic-steering cap. On a 1,000-clip Motion-X test split, Table 1 reports the proposed method winning on multiple metrics against other CFG baselines.

Critical strengths
  • The conditional-response reframing (Eqs. 1-3, Fig. 1) cleanly organizes Vanilla CFG, APG, ADG, CFG-Ctrl and CFG++ and sharpens the research question: on a trajectory already anchored by pose/face/appearance, which conditioning response should be guided.

  • The authors disclose their one failure dimension: Human Anatomy. Table 1 shows Anatomy best = ADG 0.8902 vs Ours 0.8473, and Sec. 4.1 reports this explicitly rather than relying on the aggregate Human Fidelity score.

Critical weaknesses
  • The official CFG-Ctrl Wan-Video configuration recommends cfg_scale=5, smc_cfg_lambda=5.0, k=0.2, zero warmup, and 50 sampling steps; the authors found cfg_scale=5 "produces severe temporal flickering" and switched to cfg_scale=2 while keeping 50 steps. In contrast, the proposed method uses Flow DPM-Solver++ with 20 inference steps. Therefore, Table 1 pairs "Ours" (20 steps, 3 branches) against CFG-Ctrl (50 steps, re-tuned down to cfg=2) with two simultaneous deviations from the competitor's recommended operating point. It is difficult to tell whether CFG-Ctrl's MPJPE of 169.195 reflects the method or the re-tuned, lower-cfg, simpler-sampler configuration.

  • Although the proposed method wins 8 of 9 Table-1 metrics, but the T2V margin over CFG-Ctrl is 0.7142 vs 0.7139 = 0.0003 (0.04%); the CLIP margin over APG is 29.1972 vs 29.0128 = 0.184. These margins are thin. No standard error, confidence interval, or repeated-seed result is reported.

  • Phrases such as "Bifurcation of Stability," "Manifold Inflation," "total manifold collapse," "stability tax," "rectifies the latent trajectory," and "deep semantic penetration" are difficult to understand.

  • The results presented in the supplementary video appear blurry. The method should be evaluated at a higher resolution to better showcase fine-grained visual details. Otherwise, it is difficult for reviewers to assess the potential effects of the proposed method in terms of visual quality.

Questions for authors

None

3 comments
AuthorsOpenReview ↗

Response to Reviewer GuYq (1/3)

We thank the reviewer for the constructive review. All changes below are included in the revised PDF and supplementary material.

W1. CFG-Ctrl configuration.

To align with Table 1, all methods, including CFG-Ctrl, use the same sampler and 20 steps; "50 steps" in Appendix B.4 was a typo and has been corrected. Apart from this shared step budget, the only deviation from the official CFG-Ctrl setting is cfg_scale.

All baseline scales were selected on a fixed 200-clip validation subset (disjoint from the test set; a full validation sweep costs ≈300 H200 GPU-hours per configuration). For CFG-Ctrl:

SettingNFECLIP↑T2V↑MPJPE↓Anat.↑Ident.↑Cloth.↑HF↑VQ↑Phys.↑
cfg 1, 20 steps4030.250.746167.030.7350.9370.9820.8993.5903.267
cfg 2, 20 steps (used)4030.180.801165.400.7890.8910.9800.9033.5193.335
cfg 3, 20 steps4029.900.792165.160.7780.6620.9780.8243.3492.965
cfg 4, 20 steps4029.730.744181.910.8340.7450.9870.8603.0013.135
cfg 5, 20 steps4029.470.750186.540.8420.6940.9810.8753.0993.082
cfg 5, 50 steps (official)10030.030.774164.920.8620.7790.9850.8803.5003.220

At 20 steps, cfg = 5 degrades MPJPE (+21.14 mm) and VQ (3.099 vs. 3.519), consistent with the flickering we observed. cfg = 2 is the best 20-step setting and performs on par with the official setting (MPJPE 165.40 vs. 164.92, CLIP 30.18 vs. 30.03, VQ 3.519 vs. 3.500), so CFG-Ctrl is compared close to its official performance level. Scale sweeps for all other baselines are in the revised supplementary material.

AuthorsOpenReview ↗

Response to Reviewer GuYq (2/3)

W2. Small margins and statistical significance.

We re-analyzed the same Table 1 results (no new generation) with a paired bootstrap over the 1,000 test clips (10,000 resamples). Below is Ours vs. the runner-up on each metric; the full comparison against all six baselines is added to the "Quantitative Results on Wan-Animate" subsection of the revised paper.

Metric (runner-up)Δ\Delta = Ours − runner-up [95% CI]
CLIP↑ (APG)+0.184 [+0.086, +0.281]*
T2V↑ (CFG-Ctrl)+0.0003 [−0.0087, +0.0094]
MPJPE↓ (CFG-Ctrl)−9.87 [−15.09, −4.58]*
Cloth.↑ (APG)+0.0037 [−0.0069, +0.0143]
Ident.↑ (CFG-Ctrl)+0.0306 [+0.0081, +0.0546]*
Anat.↑ (ADG)−0.0429 [−0.0621, −0.0242]*
HF↑ (APG)+0.0103 [−0.0007, +0.0218]
VQ↑ (APG)+0.157 [+0.088, +0.226]*
Phys.↑ (Vanilla)+0.102 [+0.027, +0.180]*

* 95% CI excludes 0.

Gains on MPJPE, CLIP, Identity, VQ and Phys. are significant against every baseline. We agree that the T2V (VQAScore) margin is significant only against Vanilla and LTX, and now describe T2V as comparable. CLIP, however, improves significantly over all six baselines (including the +0.184 over APG raised by the reviewer), supporting improved video–text semantic alignment.

Multi-seed. We re-ran Ours and the main competitors with seeds {0, 1, 42} on a fixed 200-clip test subset, other settings as in Table 1 (mean ± std):

MethodCLIP↑T2V↑MPJPE↓Anat.↑Ident.↑HF↑VQ↑Phys.↑
ADG28.26±0.230.684±0.008184.9±3.90.883±0.0010.867±0.0080.910±0.0013.17±0.012.84±0.02
APG28.57±0.350.687±0.003178.3±1.10.838±0.0050.903±0.0170.913±0.0023.35±0.033.01±0.05
CFG-Ctrl28.62±0.070.685±0.004157.5±1.90.834±0.0110.867±0.0240.915±0.0023.33±0.033.08±0.03
Ours28.84±0.060.704±0.004149.7±4.20.834±0.0060.902±0.0210.920±0.0043.40±0.013.15±0.02

Ours ranks first on every seed for CLIP, T2V, MPJPE, HF, VQ and Phys., with margins exceeding the seed-to-seed variation; Identity is comparable to APG and Anatomy remains highest for ADG, consistent with Table 1.

AuthorsOpenReview ↗

Response to Reviewer GuYq (3/3)

W3. Terminology.

We agree. These phrases were intended as high-level, intuitive descriptions of the observed behavior rather than formal claims, and they are loosely defined relative to the formulation in Sec. 2–3. We have removed all such expressions in the revised version and replaced them with direct, measurable descriptions:

OriginalRevised
Bifurcation of StabilityMPJPE of Vanilla and CFG++ increases steadily beyond λ=15\lambda=15
total manifold collapseADG shows severe deformation (MPJPE > 580 mm at λ=30\lambda=30)
Manifold Inflationnorm growth and over-amplification from unbounded linear extrapolation
rectifies the latent trajectorythe angular bound and magnitude restoration constrain the guided prediction
deep semantic penetration / stability taxstronger semantic guidance without loss in structural accuracy

The corresponding appendix section has been rewritten accordingly.

W4. Blurry supplementary videos.

The blurriness came from our layout: multiple methods were placed side by side within a fixed export resolution, so each video occupied only a small region. In the revised supplementary material, we keep all original cases but enlarge each method's video so that it is displayed with substantially more pixels. We also add zoomed-in key-frame crops of fine-grained regions, including interacting objects, human limbs and faces, which correspond to the semantic, structural and identity aspects evaluated in Table 1.

We hope these responses address the reviewer's concerns, and we would be grateful if the reviewer could consider updating the score. We are happy to clarify any further questions.

Reviewer_5ZFrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces a training-free guidance framework for structurally conditioned human video generation. Authors reformulate guidance from a conditional-response perspective, and separate target-specific semantic and structural responses through selective condition contrasts. Experiments on Wan-Animate and PoseAnything demonstrate that the proposed method shows boosted performance on different metrics.

Critical strengths
  1. The benchmark is comprehensive and covers a broad spectrum of metrics, including semantic scores, pose errors, etc.
  2. The method is validated on the primary Wan-Animate backbone, and also works on PoseAnything, showing good transferability across different DiT-based architectures.
Critical weaknesses

This work shows weakness in following aspects:

  1. The proposed method evaluates multiple conditioning branches, e.g., fully conditioned, generic-semantic, and structure-ablated, to construct the axis-wise responses within the sampling loop. This strategy introduces latency costs. The computation cost should be discussed to provide more details.
  2. While aggregate Human Fidelity and individual components like Clothes and Identity improve, the method slightly lags behind competitors such as ADG or TV-CFG on the Human Anatomy of VBench2. 3 The proposed framework introduces several coupled hyperparameters, e.g., sharpened exponent, angular cap θsafe\theta_{safe}, temporal half-width profiling, that require careful tuning.
Questions for authors
  1. Given the added computational expense of computing multiple conditional branches, have the authors explored approximate caching strategies, feature reuse, or selective step-skipping to bridge the latency gap with standard CFG without sacrificing the benefits of response steering?
  2. Why does the proposed method could underperform competitors on the Human Anatomy VBench2 dimension while dominating Human Clothes, Human Identity, and overall aggregate fidelity? Convincing reasons and discussions are required.
  3. How does the response-induced allocation field behave when scaling from single-person character animation to complex multi-person interactions where overlapping poses and spatial occlusions occur frequently?
  4. Reasonable hyperparameter tuning strategies should be discussed.

Comments on the paper

2 comments
Area_Chair_X8LDOpenReview ↗

Reminder for Discussion

Dear Reviewers,

We’re about halfway through the discussion period. This is a friendly reminder to check all of your assigned papers. If the authors have answered questions or responded to critiques in reviews, please leave a comment for the authors!

Thank you for your contribution to the community!

Your AC

Area_Chair_X8LDOpenReview ↗

Author-Reviewer Discussion Period is About to Finish

Dear reviewers,

Thank you for your reviews and constructive feedback.

The author-reviewer discussion period will be finishing soon. Please take a moment to review it and consider whether your concerns have been adequately addressed. It would be helpful if you could briefly indicate:

  • Whether the rebuttal addresses your main concerns;

  • Whether any significant issues remain unresolved;

  • Whether your overall recommendation has changed.

If appropriate, please also update your review and recommendation in the system.

Thank you for your time and valuable contributions.

Best regards,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.