acceptodds
Under review as a conference paper at ICLR 2027

PixReenact: Pixel-Conditioned Causal Video Diffusion for Streaming Head-Avatar Reenactment

Abstract

Streaming head-avatar reenactment aims to animate a reference image according to a live driving video, requiring robust motion transfer, long-term identity stability, and low latency. Existing methods often rely on specialized identity or motion representations, which can discard useful visual information and inherit failure modes from external extractors. In addition, many recent diffusion-based reenactment methods use offline, clip-based generation, jointly processing and denoising an entire video clip before producing its output, making continuous low-latency streaming difficult. We introduce **PixReenact**, a pixel-conditioned streaming reenactment framework built on causal video diffusion. PixReenact conditions directly on VAE-encoded reference and driving frames, without specialized identity or motion representations. To separate reference identity from driver motion, we train with cross-identity pseudo supervision together with corrective objectives anchored to the original reference and driving inputs. Long self-rollouts reduce autoregressive drift, while state-aware dual-teacher distillation separately addresses cold-start and steady-state generation. Across three cross-identity benchmarks and a long-horizon streaming benchmark, PixReenact demonstrates robust cross-identity reenactment, particularly under challenging conditions such as extreme viewpoints, occlusions, and pronounced facial expressions, while maintaining the reference identity over long streams. A 4-NFE rolling student continuously emits four frames per update with a mean emission latency of 239 ms.

Primary area:
generative models
Keywords:
Head avatar reenactment, video diffusion models, autoregressive video generation, streaming video generation, portrait animation

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

The paper introduces PixReenact, a streaming portrait-animation system that transfers facial motion from a driving video to a reference identity. It directly conditions a causal video diffusion model on reference and driving pixels, avoiding keypoints or specialized motion features. To prevent the model from copying the driver’s identity, it combines Wan-Animate-generated cross-identity pseudo-labels with identity, motion, and reconstruction losses. For low-latency generation, PixReenact uses four-stage rolling denoising and separate teachers for stream initialization and steady-state generation. The resulting student continuously outputs four frames every 239 ms while maintaining identity over long sequences.

Critical strengths
  1. The method directly uses visual latents from the reference image and driving video, without relying on keypoints or specialized motion representations. This potentially preserves richer information about facial expressions, head poses, and occlusions, especially when external motion extractors are unreliable.

  2. The separation between cold-start and steady-state teachers is well motivated. The first emission must handle a potentially large transition from the static reference to the initial driving pose, whereas later emissions mainly follow incremental motion using generated history. Modeling these states separately is a reasonable design choice.

Critical weaknesses
  1. In the supplementary material, Long Video 2 exhibits severe facial deformation, with the subject’s face becoming substantially distorted. This failure appears inconsistent with the paper’s claims regarding long-horizon identity and appearance stability. The current long-video evaluation mainly emphasizes identity similarity and may not adequately capture structural facial degradation.

  2. Several central components—including four-level rolling denoising, causal attention, KV caching, DMD, and autoregressive self-rollout—are inherited from prior causal video diffusion work. The main contribution is therefore the adaptation and combination of these techniques for pixel-conditioned avatar reenactment, rather than a fundamentally new streaming-generation mechanism.

Questions for authors
  1. Could the authors provide long-horizon measurements of facial structure, motion fidelity, and temporal quality in addition to identity similarity? It would also be useful to report how often similar failures occur across the 28 long-video examples.

  2. The steady-state teacher is trained with clean target context at t=0t=0, while it is queried during student self-rollouts under potentially degraded generated histories. Although long self-rollouts address the student-side exposure bias, it remains unclear whether the teacher score is reliable under these off-distribution contexts. Recent context-matched DMD methods address this issue through Diffusion-Forcing teacher training, exact prefix scoring, and controlled prefix corruption. Does PixReenact feed the exact cached student prefix to both the real and fake score networks? Have the authors considered context corruption or student-generated-context adaptation for the steady-state teacher?

2 comments

We thank the reviewer for recognizing the advantages of direct visual conditioning and the motivation for state-aware cold-start and steady-state teachers. We address the concerns below and provide new quantitative results and clarifications in the revised manuscript. We uploaded a revised manuscript with changes in blue (references below refer to this version).

W1 / Q1 — Long-horizon stability and facial deformation. Thank you for this remark. We evaluated all 28 five-minute sequences against PersonaLive, reporting per-minute ID/IDD, AED/APD, tLP, and FVD (Fig. 10). PixReenact maintains relatively stable identity and motion scores across five minutes and achieves consistently better FVD, while PersonaLive performs slightly better on APD and tLP.

We also quantified facial-structure consistency using SMIRK-estimated FLAME identity-shape coefficients for our method and PersonaLive. We measured deviations relative to each clip's first two seconds. Both methods exhibit stable and similar structural measures (full details in Sec. C, page 24). This is indeed a very challenging case. Our results support long-term structural consistency but some artifacts still remain, which we hope will motivate future research.

W2 — Contribution beyond existing causal diffusion. As stated in Sec. 1, our contribution lies in a novel training framework that enables effective pixel-conditioned, cross-identity streaming head-avatar reenactment. This setting remains challenging for existing causal and offline video diffusion techniques, particularly under extreme conditions and low-latency streaming requirements. Specifically, PixReenact introduces: (i) a new training strategy that makes generated data effective in a task where ground-truth cross-identity supervision is unavailable. To this end, our training strategy combines self-driving and pseudo cross-identity supervision, together with data filtering and corrective objectives, to mitigate driver-identity leakage and artifacts inherited from pseudo-generated data; (ii) long student self-rollouts that expose the model to its own generated history to mitigate training-inference mismatch and error accumulation; and (iii) state-aware dual-teacher distillation, including a dedicated cold-start teacher to handle challenging stream initialization. Together, these contributions enable robust streaming reenactment under extreme poses, expressions, and occlusions, without relying on specialized motion extractors at inference time.

Q2 — Teacher context during self-rollouts. Thank you for this remark. We have corrected the statement in Sec. 3.4 that implied the steady-state teacher always receives clean context. The teacher in fact is exposed to corrupted contexts during training (please see details in Appendix B.4). During DMD, both the frozen real-score teacher and the online fake-score network receive the exact same cached student-generated prefix from the student's self-rollout. We do not further adapt the teacher using student-generated histories, preferring to maintain a student-independent scoring target.

Dear Reviewer vEeP,

Thank you again for your time and constructive feedback. We hope that our additional experiments, clarifications, and revised manuscript address your main concerns. If any questions or concerns remain, we would be happy to discuss them further during the remaining discussion period.

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

This paper presents PixReenact, a pixel-conditioned, causal video diffusion framework for low-latency streaming head-avatar reenactment. Instead of relying on specialized identity or motion representations, the method conditions directly on VAE-encoded reference and driving frames and addresses identity leakage and streaming drift with three key ingredients: cross-identity pseudo supervision (plus real self-driving), corrective objectives anchored to the original reference/driver inputs, and long self-rollouts. To further improve online behavior, the authors introduce state-aware dual-teacher distillation (separate cold-start and steady-state score teachers) and distill a 4-step autoregressive student that continuously emits 4 frames per update with ≈239 ms latency.

Critical strengths

1.Direct pixel-level conditioning for streaming avatar reenactment is a timely and relevant design choice that avoids reliance on brittle external motion extractors; the work contributes a cohesive system that marries causal video diffusion with cross-identity learning. 2. The state-aware dual-teacher setup for causal distillation is well-motivated in streaming reenactment and appears to materially help the first emission without sacrificing steady-state behavior. 3. The restricted driving-attention window and long self-rollouts represent thoughtful architectural and training design choices to balance motion fidelity and identity preservation while mitigating autoregressive drift. 4. Provides a solid, reusable recipe (cross-ID pseudo labels + corrective objectives + long self-rollouts + state-aware teachers) that could transfer to related conditional streaming video tasks.

Critical weaknesses

1.Heavy reliance on pseudo labels from offline reenactors (Wan-Animate or PersonaLive) risks baking in their biases/failure modes, despite filtering and corrective losses; a stronger discussion or experiments quantifying this dependency would help. 2.The ±1 driving attention window includes a one-latent lookahead; while the authors argue this aligns with the rolling schedule, it still weakly relaxes strict causality and may complicate fair comparison with strictly causal baselines. 3.Corrective objectives depend on external estimators (identity, pose, expression, emotion), potentially introducing supervision noise and hidden biases; robustness to estimator errors is not deeply analyzed.

Questions for authors

1.Can you report first-frame latency separately from steady-state emission latency, and break down where the cold-start teacher most helps in terms of latency and quality metrics? 2.How strict is causality in your inference-time attention? The ±1 driving lookahead is argued to be available “for free” in the rolling schedule, could you clarify timing alignment and whether any future driver pixels are used before their corresponding outputs are emitted? 3.How sensitive is performance to the identity/motion estimators used in corrective objectives (e.g., swapping to different face ID or expression models)? Any ablations on estimator noise or bias? 4. Could you include a quantitative comparison to Wan-Animate-2 Lite (or another pixel-conditioned streaming reenactor) to better position your contributions relative to concurrent, similarly scoped work? 5. To what extent does the model overfit to the pseudo-label generator’s style? Did training with PL-labels vs. Wan-labels (Appendix) show notable qualitative behavioral differences beyond the reported counts/thresholds?

2 comments

We thank the reviewer for recognizing our contribution of direct visual conditioning, state-aware dual-teacher distillation, and our training framework. We address the raised three weaknesses first, followed by answers to the additional questions. We uploaded a revised manuscript with changes in blue (references below refer to this version) that address these concerns.

W1 / Q5 — Dependence on pseudo-label generators. Pseudo supervision is effective but drives flaws without further training design. We mitigate its flaws through pseudo-target filtering, real self-driving reconstruction, and corrective objectives (ablations: Tabs. 3, 16). We believe that this design demonstrates the value of pseudo-labeled data for such a task that has no direct cross-identity supervision data. PixReenact surpasses its pseudo-label generator (Wan-Animate) on 8/10 public/NeRSemble metrics (Tab. 2), while the variant with PersonaLive pseudo-labels yields comparable performance (Tab. 9). We discuss the trade-offs between training with these different pseudo-label sources in Sec. C (page 22).

To examine dependence on pseudo-label generator style, we measured VGG19 Gram-matrix distances on matched outputs (Tab. 10). The two pseudo-label generators differ substantially (24.19), whereas their corresponding PixReenact variants are considerably more similar (11.23). Each PixReenact variant is stylistically closer to its own pseudo-label generator than to the alternative in only 31% / 46% of pairs. Six qualitative comparisons (Fig. 7) further illustrate similar PixReenact appearances despite different reenactors, including corrections of identity-related artifacts. These results suggest reduced dependence on pseudo-label generator style.

W2 / Q2 — Driving lookahead and causality. Thank you. We agree that our [n−1,n+1] window introduces a four-frame lookahead that does not follow the strict causality definition. However, in practice, future frames are accessed only after capturing and encoding. They become available by the second denoising pass without additional delay (See Fig. 5). In a matched 5K-step continuation, the past only [n−2,n] window improves ID (0.707→0.754) but degrades AED (0.816→0.856), and slightly APD (0.071→0.077), and FVD (389→452; see Sec. C under Causal driving attention). PersonaLive uses an image-based architecture, so an equivalent process is not directly applicable, but both methods are evaluated on identical reference-driver pairs and under a common live-streaming protocol (Tab. 6). Tab. 2 also shows competitive results against offline methods that access future driving frames.

W3 / Q3 — Sensitivity to corrective estimators. We compared matched 5K-step continuations using ArcFace/DECA or AdaFace/SMIRK for corrective training, evaluating both variants with both estimator sets (Tab. 17). AdaFace/SMIRK improves motion, temporal consistency, and FVD, while ArcFace/DECA improves reference identity. These trends persist under cross-estimator evaluation, suggesting that the improvements are not merely due to optimizing against the same evaluation models. Nevertheless, estimator choice affects the identity-motion trade-off. ArcFace/AdaFace achieve approximately 99.8% LFW verification accuracy, while DECA/SMIRK achieve NeRSemble Chamfer-L1 errors of 2.385/2.276 mm, respectively. These results support the reliability of these estimators on established benchmarks.

Q1 — First-output latency and cold-start quality. The originally reported 238.9 ms measures one generation/decoding pass, not an individual latent's four-pass residence. On one H200 under the same simulated 13.8-FPS live-camera stream, warm time to first aligned output is 1253 ms (ours) vs. 1369 ms (PersonaLive), with different initial output sizes (one vs. four frames). Capture-to-output latency is 1266 vs. 1191 ms (see Tab. 6 and Fig. 5 for full details). The cold-start teacher operates during training and does not affect the inference stage. For large reference-driver pose gaps, it improves first-chunk APD from 0.265→0.249 (public) and 0.382→0.371 (see Tab. 18), based on matched 5K-step continuations.

Q4 — Wan-Animate-2 Lite. We discuss Wan-Animate-2 Lite in Related Work. Since its checkpoint is unavailable, a reproducible comparison was not possible. We compare against streaming PersonaLive and pixel-conditioned Wan-Animate-2 Base and X-Portrait. Our contribution is direct visual conditioning for continuous cross-identity head-avatar reenactment, together with corrective training and state-aware distillation (as described in page 2), rather than a new causal-diffusion backbone.

Dear Reviewer PwBF,

Thank you again for your time and constructive feedback. We hope that our additional experiments, clarifications, and revised manuscript address your main concerns. If any questions or concerns remain, we would be happy to discuss them further during the remaining discussion period.

Reviewer_GVP5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents PixReenact, a streaming framework that animates a reference portrait from a driving video across identities. It adapts Wan2.1-T2V-1.3B to causal video diffusion with four-step rolling denoising and directly conditions on VAE-encoded reference and driving frames. To address driver-appearance leakage, autoregressive drift, and large initial pose changes, it combines cross-identity pseudo-target supervision, corrective objectives anchored to the real reference and driver, self-rollouts over the training clip, and separate cold-start and steady-state distillation teachers.

Evaluation includes three cross-identity benchmarks, 28 five-minute streams, self-reenactment, component ablations, and a pseudo-label-source sensitivity study. The clearest result is improved identity preservation under extreme poses and occlusions: ID increases from 0.472 for PersonaLive to 0.727. Standard benchmarks also show competitive identity and video-quality metrics, with trade-offs in some pose, temporal, and speed measures. The reported runtime is 238.9 ms per four frames, or 16.74 FPS, on one H200; its relationship to end-to-end latency requires clarification.

Critical strengths

S1. Substantial identity-preservation gains on difficult inputs are supported across multiple benchmarks.

On the 68-pair extreme-condition benchmark, Table 1 reports ID of 0.727 versus 0.472 for PersonaLive, and Table 13 shows consistent gains across all four subgroups. The static examples in Figure 2 also show clearer, more reference-consistent outputs on several difficult cases. The benefit is not limited to this author-constructed benchmark: Table 2 improves ID from 0.662/0.734 to 0.709/0.753 on the public/NeRSemble benchmarks, while also improving AED, tLP, and FVD on NeRSemble. This is consequential for streaming applications that prioritize identity stability, even though the method does not dominate every metric.

S2. The training design targets relevant failure modes, and the evaluation examines several important confounds.

Corrective losses use the real reference and driver rather than relying entirely on pseudo targets. Identity and motion training use ArcFace/DECA, whereas evaluation uses AdaFace/SMIRK, reducing concern about optimizing exclusively against the same evaluator (B.6, B.11). Table 3 demonstrates that apparent motion improvements can accompany identity leakage; Table 9 shows competitive results with an alternative pseudo-label generator; and the single-window evaluation in Table 12 reduces the confound of long-video stitching for offline methods. The five-minute evaluation also provides evidence of identity stability beyond four-second training clips. Together, these choices increase confidence in the system-level results.

Critical weaknesses

W1. The central low-latency claim lacks a consistent timing definition and complete online evidence.

Section 3.2 states that a latent enters at iteration k and is emitted at k+3 after four forward passes. However, the 238.9 ms in Table 6 includes only one DiT forward pass (155.1 ms), one decoding operation (82.9 ms), and minor overheads. This appears to measure the steady-state interval between successive emitted chunks, whereas B.8 defines emission latency as the time from admitting a particular chunk's noise latent until its frames are decoded. Pipeline throughput and an individual chunk's residence time are different quantities; the reported number cannot support both definitions without further explanation. Timing also uses precomputed driver latents, excluding 45.5 ms of driver-VAE encoding and startup, while PersonaLive's driving preprocessing is included. Stating that encoding can overlap generation does not establish the cost under actual concurrency on one GPU.

Because low latency is a central contribution, the paper should separately report the steady-state emission interval, time to first output, and capture-to-output latency for corresponding driving frames under a common protocol for both methods. The availability of neighboring driving latents and any lookahead waiting also need to be specified. At 16.74 FPS, processing every frame of a 25/30 FPS input without dropping or resampling would accumulate a queue. This does not invalidate streaming generation, but it affects interactive applicability. Simply multiplying 238.9 ms by four or adding encoding time would not constitute a measured end-to-end latency.

W2. Attribution of the proposed training contributions remains incomplete, particularly for dual teachers and long self-rollouts.

The main ablations in Table 3 continue an already trained full model for 5K steps. This tests sensitivity to removing a component, but does not adequately establish what happens when training omits it from the outset. In particular, the cold-start teacher is supported mainly by modest changes in clip-averaged AED/APD/FVD, without targeted first-chunk or large-pose-gap results; the initialized student may retain capabilities acquired during earlier distillation. Furthermore, Figure 3 primarily measures long-term identity similarity, and Tables 3/14 do not directly compare rollout lengths. Consequently, long-term stability cannot be clearly attributed to long-rollout training, and stable identity alone does not establish stable motion quality.

This matters for assessing the technical advance over established causal diffusion and pixel-driven animation. At least one budget-matched single-/dual-teacher comparison from a common pre-distillation initialization, with first-chunk metrics, would strengthen the claim. A short-/long-rollout comparison measuring both identity and motion errors over time would address the second attribution gap. Table 14 does include a useful 20K-step DMD comparison initialized from the same dense teacher, but that experiment does not resolve these two specific questions.

Conditions for raising the score. These concerns are the main reasons for my current weak-reject recommendation, rather than irreparable flaws. If the rebuttal clearly reconciles the latency definition and pipeline timing, provides online measurements under a consistent protocol, and supplies targeted evidence clarifying the contributions of the cold-start teacher and long-rollout training, I would be willing to reconsider and raise my score. Any increase would depend on whether the response substantively resolves these concerns.

Questions for authors
  1. [Clarification and measurement; W1] Is 238.9 ms the interval between successive emitted chunks or the residence time of the same latent from admission to decoding? Please provide a timeline covering four denoising passes, driving-input availability, the ±1-latent attention window, and decoding, together with first-output latency, end-to-end latency, and sustained throughput under live input. Has encoding/generation overlap on the same H200 been measured? Can PersonaLive be measured with the same timing boundaries?

  2. [Targeted experiment request; W2] Can single- and dual-teacher training be compared from a common pre-distillation checkpoint, reporting ID/AED/APD specifically for the first four-frame chunk and large reference/driver pose gaps? Is there a budget-matched rollout-length comparison with long-horizon motion metrics? These results would directly affect my assessment of the claimed training contributions, rather than merely suggesting future work.

  3. [Evaluation clarification] B.12 acknowledges that extreme poses and occlusions reduce the reliability of SMIRK metrics, and B.11 also relies on face detection and alignment. What are the per-method detection-failure rates, how are failed frames scored, and are valid-frame sets matched? Could a small blinded evaluation or manually checked subset validate the joint identity/motion trade-off? This affects interpretation of the difficult-condition results.

2 comments

We thank the reviewer for highlighting substantial identity-preservation gains across extreme and standard benchmarks, long-horizon identity stability, and systematic ablations and evaluations supporting our system-level results. We address W1–W2 with new live-streaming measurements, targeted ablations, and matched-frame evaluation. We uploaded a revised manuscript with changes in blue (references below refer to this version).

W1 / Q1 — Online latency and streaming protocol. Thank you for this remark. Following this remark we noticed that the originally reported 238.9 ms corresponded to one generation/decoding pass. For the per-update processing time, we added the 45.5 ms of driver VAE encoding, so the total steady-state processing time is 284.4 ms per four-frame update (14.07 FPS processing capacity). We thank the reviewer again for this comment and we have updated the revised version accordingly.

To report end-to-end latency, we measured both methods on one H200 under the same simulated 13.8-FPS live-camera stream, including driving preprocessing and no sustained queues. Table 6 and the new Figure 5 include the full details and a new four-pass timeline. PixReenact/PersonaLive achieve close 1112/1083 ms per-latent residence, 1253/1369 ms time to first driver-aligned output, and 1266/1191 ms capture-to-output latency. Their maximum sustainable camera rates are 13.85/19.04 FPS. The initial aligned output contains one frame for PixReenact vs. four for PersonaLive. The ±1 driving window attends to the future driving latent only after its frames are captured and encoded, starting at the second pass, with no additional waiting. At 13.8 FPS, outputs are emitted every 289.9 ms. Sustaining 25/30-FPS input requires temporal subsampling. We have corrected the latency terminology in the abstract, Sec. 4.1, and Appendix B.8.

W2 / Q2 — Cold-start teacher and long self-rollouts. We assess the impact of the cold-start supervision through targeted first-chunk experiments. For large reference-driver pose gaps, the dual-teacher variant, with the cold-start teacher, reduces APD from 0.265 to 0.249 (public) and 0.382 to 0.371 (extreme). Improvements occur in 92% / 91% of large-gap pairs (95% bootstrap CIs; see Tab. 18, Fig. 13). The improvements persist over the first 24 frames. Effects on ID are small, while AED increases slightly. Thus, cold-start supervision primarily improves first-emission head-pose alignment. We compared budget-matched 5K-step continuations with and without specialized cold-start supervision, starting from the same PixReenact checkpoint. These continuations imply targeted evidence, for the cold-start contribution. Table 18 and Appendix C (page 29) include additional details of our analysis and the reported results.

Long self-rollouts. Following established practices in autoregressive video generation (e.g., Self Forcing), we train the student on its own generated history over the full 4-second training clip (≈28 emissions), comparable to the 5-second video horizon used in prior work. This is intended to reduce training-inference mismatch and accumulated errors. Our additional five-minute evaluation of ID, AED/APD, tLP, and FVD (Fig. 10) demonstrates the complete system's long-horizon stability.

We additionally attempted budget-matched training from a common pre-distillation checkpoint for 20K steps, using different rollout lengths of 8/16/28 emissions. However, all variants exhibited poor reenactment quality, preventing a meaningful comparison. We believe that isolating the contribution of the rollout length requires repeating the full multi-stage training procedure, which is infeasible within the rebuttal period.

Q3 — Reliability under extreme conditions. Thank you for this insightful comment. We recomputed metrics on common-valid frames and our conclusions hold: PixReenact vs. PersonaLive yields ID 0.728/0.473, AED 1.138/1.295, and APD 0.170/0.133. Representative matched examples (Fig. 11) reveal that PersonaLive can sometimes obtain lower (better) AED and APD despite severely degraded or faceless output.

In the original protocol, ID averages only detected generated faces, whereas AED/APD scores every frame by applying the driver's face box to the generated image before SMIRK, even when no recognizable face is present. For the common-valid-frame evaluation, ID requires face detection in both generated outputs (96.7% retained), while motion requires FAN detections in the driver and both generated outputs (94.3% retained), retaining the original driver-box-based motion scoring. Detection failure-rates are provided in Table 15.

Dear Reviewer GVP5,

Thank you again for your time and constructive feedback. We hope that our additional experiments, clarifications, and revised manuscript address your main concerns. If any questions or concerns remain, we would be happy to discuss them further during the remaining discussion period.

Comments on the paper

2 comments
Area_Chair_rNHeOpenReview ↗

Please read the response from authors and discuss

Dear reviewer fellows,

The author rebuttal has been posted. The rebuttal period is a critical part of our review process, allowing authors to clarify misunderstandings and address specific concerns.

Please take a moment to read the author response, engage in the discussion by replying to the authors' responses, and update your review if the rebuttal changes your assessment.

Your active participation is vital to ensuring a fair and thorough evaluation of the submissions. A constructive discussion helps us make the most informed decisions.

Thank you for your time and valuable contributions to the review process.

Best regards,

Your AC

AuthorsOpenReview ↗

General Response

We thank all reviewers for their careful and constructive feedback. As the discussion period comes to an end, we would like to summarize our main responses and hope that they address the reviewers' concerns.

The main updates are:

  1. Online latency and streaming protocol (GVP5, PwBF): We clarified the latency definitions and conducted new end-to-end measurements under a common live-camera protocol, reporting processing throughput, time to first output, and capture-to-output latency (Tab. 6, Fig. 5). PixReenact achieves faster first-aligned output, comparable capture-to-output latency and per-chunk residence, while PersonaLive offers higher throughput.

  2. Training contributions (GVP5, PwBF): New targeted cold-start experiments demonstrate improved initial head-pose alignment, with improvements in 92% / 91% of challenging pairs with large reference-driver pose gaps across two benchmarks (Tab. 18, Fig. 13). Our long self-rollout training follows established practices (e.g., Self Forcing) to mitigate accumulated errors, and our five-minute evaluation demonstrates the complete system's long-horizon stability.

  3. Long-horizon stability (vEeP): Our extended evaluation across 28 five-minute sequences demonstrates stable identity and motion metrics, and improved video quality (FVD) (Fig. 10, Appendix C).

  4. Pseudo-label supervision (PwBF): Our experiments demonstrate that PixReenact does not simply imitate the visual style of its pseudo-label generator. Training with different generators yields comparable performance, while the resulting PixReenact variants are substantially more similar in visual style than the generators themselves (style distance 11.23 vs. 24.19). Moreover, PixReenact outperforms Wan-Animate on 8/10 metrics. These experiments support the effectiveness of combining pseudo-label supervision with real self-driving and corrective objectives for the challenging cross-identity streaming reenactment task (Tabs. 2, 9-10; Fig. 7).

  5. Additional ablations and clarifications:

    • Driving lookahead (GVP5, PwBF): We clarified that our bounded driving lookahead introduces no additional waiting and evaluated a past-only alternative, demonstrating the trade-offs between identity preservation and motion fidelity (Fig. 5, Appendix C).

    • Evaluation clarification (GVP5): Our conclusions hold under the common-valid-frame evaluation protocol, confirming PixReenact's advantages in identity preservation and AED under extreme conditions (Tab. 15, Fig. 11).

    • Corrective estimators (PwBF): Alternative estimators demonstrate consistent trends across evaluation protocols, with an identity-motion trade-off depending on estimator choice (Tab. 17).

    • Teacher context (vEeP): We clarified that the steady-state teacher is also trained with corrupted contexts and that both DMD score networks receive the same student-generated history (Appendix B.4).

Detailed results and explanations are provided in the revised manuscript and our individual responses to each reviewer.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.