acceptodds
Under review as a conference paper at ICLR 2027

CinePersona: Camera-Aware Dynamic Persona Encoding for Identity-Preserving Human Video Generation

Abstract

Cinematic human video generation must jointly control who appears in frame, what the subject does, and how the camera moves, yet the literature has split into single-axis specialists whose serial composition compounds error on every axis. We argue that the right conditioning object is a chronologically ordered space-time lattice of micro-clips cut from a reference video, and present CinePersona, a camera-aware reference encoder for a frozen 14B video diffusion transformer. Its Chrono-Mosaic Lattice packs nine time-ordered micro-clips into one trunk-compatible tensor, and we show that this ordering is what keeps camera direction decodable: time-reversed trajectories are provably indistinguishable under any shuffled interface. A dual-pathway encoder pairs a diffusion-synchronized implicit stream with explicit persona, action and camera branches, Cine-RoPE writes the target trajectory into the trunk's rotary phases, and a tri-axis sampler exposes one guidance dial per control, all while training 1.5% of parameters. On CineHuman-Bench and OpenS2V-Eval, CinePersona natively controls all three axes and achieves state-of-the-art persona similarity, action alignment and camera fidelity, the latter verified with an estimator decoupled from training, on rendered scenes with ground truth, and by 41 human raters.

Primary area:
generative models
Keywords:
human video generation, identity preservation, camera control, video diffusion, reference encoding, multimodal conditioning

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

This paper introduces CINEPERSONA, which addresses the separation problem of identity, action, and camera control in character video generation, which often leads to mutual damage when combined. It proposes to cut the reference video into 3x3 short segments sorted by time and use dual-path encoding: the implicit pathway retains the overall visual clues, while the explicit pathway extracts character identity, action posture, and camera. Subsequently, CINE-ROPE injects the target camera trajectory into the generation model and uses three-axis guidance to adjust text, identity, and camera/action conditions. The results report improved identity, action, and camera metrics on the newly established CINEHUMAN-BENCH.

Critical strengths
  1. The method has a clear motivation and targeted ablation experiments. CINEPERSONA integrate time-ordered reference fragments, implicit/explicit dual pathways, and CINE-ROPE mechanisms; they retain rich clues in the reference video, provide structural conditions, and inject target camera geometry information into the backbone network. The results of ablation experiments show that when the reference segment order is disrupted or the CINE-ROPE mechanism is closed, the rotation error will increase, while the identity consistency index remains relatively stable.

  2. The paper constructs CINEHUMAN-BENCH, a dataset comprising 50 identities and 400 video clips, while simultaneously assessing identity, motion, camera shots, and video quality. Among them, camera metrics are computed using MegaSaM and validated using ground-truth trajectories from synthetic scenes as well as human evaluations. This tests the core arguments of the paper more effectively than just reporting on overall perceptions or individual indicators of identity, making the paper's results more convincing.

Critical weaknesses
  1. The paper does not discuss the ActCam, 3STC-HIA, and Tri-Prompting, which have previously involved character identity consistency and combinations of action and camera movement. It is necessary to supplement the qualitative and quantitative comparisons of these work.

ActCam: Zero-Shot Joint Camera and 3D Motion Control for Video Generation. El Khalifi, O. et al. (2026).

3D Scene-Adaptive Trajectory-Controllable Human Image Animation with Camera Movement. Liu, D. et al. (2026).

Tri-Prompting: Controllable Video Generation with Scene, Subject, and Motion Prompts. Zhou, Z. et al. (2026).

Questions for authors
  1. Can the reference video and target trajectory be different? When inference, the model extracts the camera pose of the original video from the reference video while receiving the user-specified target camera trajectory. Please explain: Can the target trajectory differ from the camera movement in the reference video? How do the reference video and target trajectory usually pair during training? How many examples of “the same reference video, different target trajectory” has the model seen?

  2. Why does shuffling cells lead to a significant increase in camera error? In Table 4, after shuffling cells, the rotation error of the generated video relative to the target trajectory increased. Since the target trajectory is input separately, why would the order of the reference segments have such a great impact on the ability to follow them? Please explain whether the target trajectory is same as the camera in the reference video, and analyze whether the increase in error indicates that the model depends on the camera movement clues in the reference video. It is best to report the results of the same reference video paired with multiple different target trajectories to test whether the model can truly switch camera.

  3. Can the action also be controlled independently of the reference video? The paper extracts action poses from the reference video but does not explicitly demonstrate a target action input separate from the reference video. The guidance during inference also places “camera/action” on the same axis. Please clarify whether users can specify new actions independently, whether action control and camera control can be adjusted separately, and which experiments have verified this independent control.

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

In this paper, the authors propose a unified framework for controlling human identity, human motion, and camera motion in video generation models. To this end, they introduce a conditioning scheme (CinePersona encoder), which combines an entangled encoding of the raw input conditions (implicit path) with separate encodings of each disentangled condition (explicit path). These conditions are then fused within the self-attention and cross-attention blocks of a pretrained video backbone. Additionally, the authors encode camera trajectories using a geometric positional embedding (Cine-RoPE), with a learnable scaling gate that allows the model to recover the original positional embedding of the pretrained video backbone. At inference time, they introduce a three-component velocity guidance strategy (Tri-Axis Guidance), that separately controls text, identity, and camera-action conditions.

Finally, the authors compare their method against state-of-the-art approaches and conduct ablation studies to assess the contribution of each component. They also evaluate camera control quality using an external model that was not used during training.

Critical strengths
  • The unified control of these three conditions is highly relevant and represents a valuable contribution to video generation.

  • The evaluation protocol is well designed and conducted, strengthening confidence in the reported results.

  • The extensive experiments demonstrate the effectiveness of the proposed method and provide strong justification for the design choices.

Critical weaknesses
  • The paper lacks sufficient qualitative results. In particular, I could not locate the full video clips referenced in Appendix M. Furthermore, the qualitative results appear to lack diversity. Including a broader and more diverse set of generated examples would strengthen the evaluation and provide a more comprehensive assessment of the method's capabilities.

  • Overall, the paper is easy to follow; however, some parts are unnecessarily complex (e.g., lines 183–187; 210–212; ...), and experimental results are sometimes embedded directly into the method section to justify design choices (e.g., lines 191–192; 212; ...), breaking the flow. The paper would benefit from further polishing, particularly by more clearly separating the method and experiments sections. This would allow the experimental results to provide a more analytical discussion and deeper insights, while keeping the method section focused on a clear and concise description of the proposed approach. Additionally, some notations are either introduced without sufficient explanation or used without being recalled (e.g., lines 209l; 248; ...), which may hinder readability.

Questions for authors
  • Could the authors provide additional qualitative results to better illustrate the capabilities and limitations of their method?

  • Several camera-aware RoPE embeddings have been proposed in the literature, including PRoPE (Yi et al., NeurIPS 2025) and RayRoPE (Wu et al., ECCV 2026), among others. Could the authors better position Cine-RoPE against these approaches and clarify the specific motivations behind its design? In particular, why was a new positional embedding necessary for their framework, and why existing approaches could not be directly adopted or adapted to meet their requirements?

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

This paper formulates reference conditioning as a chronologically ordered space-time lattice, with Cine-RoPE modulating the split-axis rotary phases from Plucker geometry. The formulation is applied on the frozen 14B Wan-2.1 video DiT, leaving only 1.5% of the backbone, about 205M parameters for training. The authors introduce CINEHUMAN-BENCH (400 cases over 50 identities, 10 trajectories), and report SOTA across nine metrics on their benchmark. Unfortunately, the video demonstrations promised in the paper is not present, making reviewers difficult to judge the qualitative claims.

Critical strengths
  • Estimator-decoupled evaluation. All camera numbers use MegaSaM, which shares no weights with training (App. B.7)。 Blender synthetic-GT floors are reported (0.11 deg MegaSaM, 0.19 deg VGGT, Table 10); persona/action backbones are re-scored with AdaFace / MAE-ViT / ViTPose (Table 4, supp p.9), preserving margins within 0.009. The authors openly disclose their own VGGT self-coupling bias (-0.07 deg; the only system better under VGGT; Table 11).
Critical weaknesses
  • No video demonstrations were actually submitted. The supplementary §M (“Video Demonstrations and Prompts”) states that “The supplementary videos provide the complete four-model comparison in the main paper and five further CINEPERSONA examples” and that “the complete clips accompany this PDF”; §M.2 describes a videos/comparison/, videos/cinepersona/, prompts/ package with media_manifest.csv. In the submission bundle, however, the only supplementary file for this paper is the 42-page PDF; no video files are present. For a method whose headline claims are one-take camera motion (dolly-in, 180° orbit) and identity retention over time, the qualitative claims (Fig. 3, human-preference study) cannot be independently verified by reviewers. For this reason, I tend to reject this submission.

  • Single-axis baselines are scored on axes they never attempt. Identity-only systems receive the trajectory as text via a fixed "camera-copy prompt" (App. H, supp p.30), yielding RotErr 5.6-6.4 deg; camera-only systems inherit identity from one shared first frame, yielding Face-Sim ~0.40. CINESCORE folds these near-floor columns into its mean.

  • Protocol asymmetry across baselines. Table 20: CamI2V at 320x512 x 16 frames and ConsisID at 480x720 x 49 frames, vs 480x832 x 81 here. FVD is set-level and resolution-sensitive, and RotErr on a 16-frame trajectory is not directly comparable to 81 frames; the cross-resolution FVD gap (248.6 vs CamI2V 309.5) conflates resolution with quality.

Questions for authors

None

Comments on the paper

1 comment
Area_Chair_dzCzOpenReview ↗

Auhor-Reviewer Discussion Peroid 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.