acceptodds
Under review as a conference paper at ICLR 2027

STAR: Selective Timestep Anchor Reward for Identity-Preserving Video Generation

Abstract

Identity-preserving image-to-video post-training must strengthen facial identity without unnecessarily suppressing motion. We introduce STAR, which addresses this trade-off by selectively placing reward gradients along the denoising trajectory. It calibrates an anchor window from the joint response of identity similarity and facial dynamics to timestep-wise perturbations. Within a single-leap backpropagation framework, it samples anchors from this window while an identity-forward connector evaluates the reward on the fully generated video. The resulting update separates timestep coverage from backpropagation depth. Our analysis characterizes terminal scoring, endpoint attenuation of raw reward gradients, and the local identity–motion trade-off of anchor selection. We build HQ4000, a 4,000-clip dataset with per-subject pose and expression annotations, and adapt an identity verifier to support training. We evaluate STAR on four I2V settings: HunyuanVideo-1.5 with 50 and 12 sampling steps, Wan2.2 with 50 steps, and MiniMax-H3. STAR improves test500 identity similarity by 9.8%–35.6% over the base models, with facial dynamics well preserved and no visible reward hacking. Our project page is available at https://anonymous.4open.science/w/STAR-2027/.

Primary area:
generative models
Keywords:
video generation, identity preservation, reward post-training, diffusion transformer, image-to-video, anchor window calibration, differentiable identity reward, facial dynamics, reinforcement learning, fine-tuning

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_3znHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes STAR, a method for fine-tuning image-to-video generators to preserve facial identity while limiting the loss of facial motion. It selects a window of denoising timesteps and applies identity-reward gradients through one sampled step per update. The results across several generators and human evaluations provide encouraging evidence of identity improvement. However, my main concern is whether the proposed timestep selection offers a better balance between identity and expression preservation than simpler alternatives. The window ablation reports only identity similarity, while the complete method reduces facial-motion amplitude by 11%–12%. This leaves unclear how much of the improvement comes from better identity preservation versus restricting facial variation.

Critical strengths

The paper addresses a relevant problem and provides reasonable evidence of identity improvements across several generators. The evaluation includes both automatic metrics and human judgments. The authors also acknowledge the remaining reduction in facial motion.

Critical weaknesses
  1. Figure 1 does not convincingly separate identity improvement from changes in expression and pose. In the second row, STAR mainly appears to reduce the smile, making the face look closer to the reference expression. Enlarged face crops and expression trajectories would help clarify the difference. The caption should also define “Base” and identify the generator used in each row.
  2. Table 2 reports only identity similarity, although the proposed window aims to balance identity and motion. Reporting both metrics for nearby windows, fixed anchors and the existing sampling alternatives would better establish the benefit of calibration.
  3. In Table 4, the author shows that changing the identity reward contributes substantially to the improvement. A joint ablation of verifier choice and timestep selection would help establish how much each component contributes.
  4. The paper reports an 11%–12% reduction in facial-motion amplitude. Although the human evaluation is useful, stronger evidence is needed to show that requested expressions remain intact rather than becoming weaker. Targeted expression judgments and examples of failure cases would help.
Questions for authors

See the critical Weaknesses

Reviewer_Z8GqOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses facial identity drift in image-to-video generation while aiming to preserve natural motion. It proposes STAR, a post-training framework that selects a timestep window based on identity and facial-dynamics responses, then backpropagates rewards evaluated on the final generated video through a single sampled anchor. The framework also incorporates pose-aware reference pools and an adapted identity verifier ensemble. Experiments across multiple video generation models show improved identity consistency while largely preserving facial dynamics and overall visual quality, with human evaluations supporting the identity improvements.

Critical strengths
  • The paper is clearly presented and proposes a well-motivated approach to improving identity preservation while limiting motion degradation.
  • Experiments across multiple backbones and human evaluations support the framework’s effectiveness.
  • HQ4000 provides useful single- and multi-person training data with pose and expression annotations.
Critical weaknesses
  • More complete evaluation of the identity–motion trade-off. Table 2 compares timestep windows using FaceSim alone, although preserving facial dynamics is central to the proposed selection strategy. Reporting Dyn× alongside FaceSim for these windows would better demonstrate the benefits of joint identity–dynamics calibration. A small comparison across regularization strengths would also help distinguish the contribution of window selection from that of the teacher regularizer.
  • Qualitative comparisons for key ablations. The quantitative ablations in Tables 2 and 4 could be complemented by visual comparisons across different anchor windows and between models optimized with the original ArcFace reward and the adapted min3 reward. Side-by-side videos or matched frame sequences using the same inputs and random seeds would help illustrate how these choices affect identity preservation, facial expressions, and motion.
  • Clearer window calibration details. Section 3.2 and Appendix B describe the calibration principle and report the selected windows, but the precise criteria used to determine their boundaries could be explained more explicitly. Providing the actual identity and dynamics thresholds, together with a small sensitivity analysis, would improve reproducibility and clarify how reliably the procedure transfers across backbones.
Questions for authors
  1. Could you report Dyn× alongside FaceSim for the windows in Table 2, and clarify whether their relative performance changes with the teacher regularization strength?

  2. Could you provide matched visual examples comparing different anchor windows and models trained with the original ArcFace versus adapted min3 rewards?

  3. What exact criteria determine the window boundaries, and how sensitive are the selected windows and final results to these criteria across backbones?

Reviewer_NZzYOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents STAR (Selective Timestep Anchor Reward), an efficient reward-based post-training framework for improving identity consistency in image-to-video generation without suppressing motion. The key insight is that different denoising timesteps exhibit different sensitivities to identity preservation and facial dynamics during inference, making selective gradient placement more effective than uniformly optimizing across the sampling trajectory.

Its key contributions include: (1) a probe-guided timestep selection strategy that balances identity improvement and motion preservation; (2) a single-anchor optimization framework that evaluates rewards on the final generated video while backpropagating through only one denoising step; and (3) the HQ4000 dataset and a pose-aware identity verifier for robust reward computation.

Experiments across different backbones demonstrate improvements in identity similarity, with comparable video quality and moderate degradation in facial motion.

Critical strengths
  • The paper is clearly written and well organized.
  • The paper addresses an important challenge (tradeoff between identity and motion) in video generation post-training by introducing probe-guided timestep selection and single-anchor backpropagation. This enables identity optimization without backpropagating through the entire denoising trajectory, offering a practical approach for training large video diffusion models.
  • The method demonstrates consistent identity improvements across several video generation models, supported by ablation studies and human evaluations, highlighting its practical effectiveness and generalizability.
Critical weaknesses
  • The calibrated timestep windows are fixed for each backbone, but their robustness across sampling schedules, data distributions, and training stages remains unclear. For example, MiniMax-H3 is evaluated only with 50 sampling steps, whereas practical inference may use fewer steps. It is unclear whether recalibration is necessary and whether the selected windows remain effective under these settings.
  • The calibrated window provides a relatively modest FaceSim improvement over simple late-stage sampling (0.7249 vs. 0.7110), whereas replacing the original ArcFace reward with the adapted min3 verifier yields a larger gain (0.6574 to 0.7249). A factorial ablation across timestep selection strategies and reward models would help clarify whether the reported improvements primarily arise from the proposed calibration mechanism or the stronger reward model.
  • Although preserving facial dynamics is a central motivation, STAR still reduces facial-motion amplitude by approximately 11–12% as shown in Table5.
Questions for authors
  • Given a different inference step (20 steps vs 50 steps), is it necessary to recalibrate for the same backbone?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.