1. The novelty of the reference-conditioning design is not established. Kaleido [1] is cited as a multi-subject interface, but its actual mechanism is not discussed: it already concatenates reference-image and noisy-video latents, shifts reference coordinates outside the video spatial grid, and ablates H-only, W-only, and H/W shifts. This is directly relevant to Secs. 4.2 and 4.4, rather than just background on a different application.
TASS-RoPE [2] is even closer. It assigns a facial identity reference the coordinates (T, H+h, W+w). With video indices 0,...,T-1, this is the single-reference case of Eq. (3) at j=1 and δ=1. HarmoView's Jump-RoPE [3] introduces gaps between reference groups and between references and video tokens; Aura [4] uses temporal and spatial RoPE shifts to separate reference categories and instances. FLUX.1 Kontext [5] also predates this work's use of virtual temporal coordinates to distinguish concatenated context images, although its task is image generation/editing. These are not all the same layout. The potentially distinct choice here is to give multiple views a shared temporal coordinate and separate spatial bands. The paper needs to explain and isolate the benefit of that choice against the closest layouts. A comparison only with vanilla RoPE does not establish it. The supplementary standard RoPE identity and qualitative examples at 450 training steps do not fill this gap; distinct coordinates also do not, by themselves, prevent reference copying.
The masking contribution has a similar attribution problem. Video Alchemist [6], which is already cited, identifies same-video reference copying, pose/lighting leakage, and reduced motion, and uses reference augmentation to address them. TASS-RoPE also combines spatially shifted reference conditioning with augmentation. Please compare region masking with these ordinary reference augmentations under matched conditions. Omni-ID [7] is additionally relevant to aggregating complementary facial information across poses, though it is an image representation method rather than a drop-in video baseline. The issue is the unestablished technical increment, not the number of references.
The dataset construction also does not establish a new annotation method. Sec. 4.1 and Supplement Sec. 3 describe quality filtering, RetinaFace detection, shot/track filtering, head-pose thresholding, reference-frame sampling, optional Grounding DINO/SAM masks, and Qwen2.5-VL captions. This is a straightforward combination of existing tools. A useful curated dataset could still be a contribution, but the paper needs to show its value through a quality audit, pose/identity coverage statistics, and a controlled comparison with existing data. The checkbox comparison in Supplement Table 1 does not establish this. Even the selection rule is unclear: Fig. 3 uses a 60-degree pose-range threshold, whereas the text says 45 degrees. As presented, I see a data-curation effort, but no demonstrated innovation in annotation or evidence that the data contribution compensates for the limited methodological novelty.
2. The experiments do not isolate the claimed identity–motion improvement. The main comparison uses Wan-2.1-14B, but the component ablations use Wan-2.2-5B. There is no corresponding 14B baseline trained on the same 22K videos without the proposed components. The main result therefore does not tell us how much comes from the data and fine-tuning, and how much comes from RM and RD-RoPE.
Even in the 5B ablation, the full model reduces MvRC-Arc/Cur from 0.546/0.515 to 0.535/0.502, and also scores below RoPE alone at 0.552/0.522. An identity–motion trade-off is reasonable, but it needs direct measurement. MvRC in Eq. (4) simply averages face similarities across references and frames. It does not measure viewpoint coverage or temporal continuity: reordering the generated frames leaves it unchanged. NaturalScore is a general realism judgment from 16 sampled frames, without validation that this judge reliably detects the specific angle-locking and switching failures. Figures 6 and 14 show selected trajectories, but provide no aggregate measure of locking, abrupt changes, or achieved pose range. The human study compares the final systems rather than the component variants, so it does not resolve this attribution problem either.
Important reproduction details are also absent from the main paper and supplement, including the optimizer, learning rate, batch size, training duration, masking granularity, the value of the temporal offset, and inference settings. The supplement says the dataset is publicly released, but I could not find an access link in the submission or supplied files. These omissions make the comparison difficult to assess or reproduce. I do not think the current evidence is sufficient for acceptance.
References mentioned in this review
[1] Kaleido: Open-Sourced Multi-Subject Reference Video Generation Model. Secs. 4.2/5.4, Eq. (4), Table 2. https://arxiv.org/html/2510.18573v1
[2] Spatial-Temporal Decoupled Reference Conditioning for Identity-Preserving Text-to-Video Generation (ST-DRC). Sec. 3.2, Eq. (4). https://arxiv.org/html/2606.02441v1
[3] HarmoView: Harmonizing Multi-View Constraints for Identity-Consistent Video Generation. Sec. 3.1, Jump-RoPE. https://arxiv.org/html/2606.10839v1
[4] Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment. Sec. 3.2, Eq. (2). https://arxiv.org/html/2607.04311v1
[5] FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space. Token sequence construction, Eq. (2). https://arxiv.org/html/2506.15742
[6] Multi-subject Open-set Personalization in Video Generation (Video Alchemist). Sec. 3.3. https://arxiv.org/html/2501.06187v2
[7] Omni-ID: Holistic Identity Representation Designed for Generative Tasks. https://arxiv.org/html/2412.09694
Response to Reviewer t4i4
We thank the reviewer for the constructive suggestions on copy-paste evaluation, metrics, and the dataset.
1. Quantitative evaluation of reference preference.
Thank you for recommending WithAnyone. Its reference-versus-target comparison provides a valuable perspective on reference-copying behavior. We will cite this work in the revision.
We adapt its normalized angular-distance formulation to our setting with three conditioning references and ten held-out same-identity gallery images. The gallery excludes the conditioning references. For each valid generated-frame embedding g, we select the closest conditioning reference:
r∗(g)=argmaxrj∈Rsim(g,rj).
We compute the score against each gallery embedding hk:
mk(g)=max{d(hk,r∗(g)),ε}d(g,hk)−d(g,r∗(g)).
M(g)=101∑k=110mk(g).
d(a,b)=arccos(sim(a,b)).
Here, ε prevents division by zero. Higher scores indicate stronger preference for the closest conditioning reference. Scores are averaged over valid frames within each video and then across test cases. This is a gallery-based adaptation: the held-out images provide alternative same-identity appearances rather than prompt-corresponding target GT.
Increasing masking from 40% to 60% reduces reference preference under this evaluation. We report this supplementary measure alongside identity consistency and naturalness.
2. MvRC, NaturalScore, and the identity–naturalness trade-off.
MvRC measures identity similarity between generated faces and a held-out multi-view gallery. The supplementary measure above evaluates relative preference for conditioning references. NaturalScore follows the OpenS2V-Nexus protocol and provides a complementary VLM-based assessment of facial-motion naturalness.
The component ablation in Table 2 illustrates the identity–naturalness trade-off:
Our objective is to improve naturalness while retaining strong identity consistency, rather than solely maximizing reference-face similarity.
3. Dataset contribution and data ablation.
Our dataset contains 22,516 video clips and is publicly available on Hugging Face. Its construction emphasizes substantial within-clip facial-pose variation for multi-view identity learning.
We agree that the dataset contribution merits separate quantification. Table 2 isolates model components using the same training data. A data-only ablation instead varies the training data while fixing the backbone, training budget, and components. We will clarify this distinction; a controlled data-only ablation is not included in the current results.
4. MvRC gallery and failed detections.
The test set contains 30 identities and 150 cases, with five prompts per identity. MvRC uses ten held-out images per identity, totaling 300 images. Test identities are disjoint from training, and the three inference-time conditioning references are excluded from the gallery.
Images are selected for quality and diversity, including sharpness, contrast, brightness, and diversity across clips, shots, and frame positions, rather than a fixed yaw quota. We measured the gallery distribution using InsightFace (buffalo_l):
Additionally, 16.7% of images have absolute yaw of at least 45 degrees. All 30 identities have yaw ranges above 45 degrees; 29 exceed 60 degrees, and 22 include both left and right views. These statistics describe static-gallery coverage.
Generated frames without a valid face detection or identity embedding are excluded from the MvRC average. The generated-frame failure rate is not included in the current results. We will clarify gallery composition, separation from conditioning references, and invalid-frame handling in the revision.