acceptodds
Withdrawn from ICLR 2027

Generative Video Anonymization: Whole-Body Identity Suppression with Semantic Preservation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Modern person re-identification systems can recover identity from whole-body biometric cues such as gait, body shape, clothing configuration, and silhouette, making face-only anonymization insufficient for privacy-preserving video analysis. Existing anonymization methods either destroy semantic utility through coarse obfuscation or preserve full-body identity cues by operating only on facial regions. We introduce Generative Video Anonymization (GVA), a new task that aims to synthesize a semantically equivalent video in which whole-body biometric identity is suppressed while action dynamics, scene context, and visual realism are preserved. We first formalize GVA as a constrained identity-action decoupling problem and show that general-purpose video-to-video editing methods are structurally misaligned with this objective, failing due to inadequate subject grounding, the absence of Re-ID-aware identity objectives, and weak temporal identity consistency. To support principled evaluation, we propose a dedicated privacy-utility protocol with metrics for assessing identity retrieval resilience, feature-level identity leakage, scene preservation, temporal coherence, and behavioral action consistency. We further introduce the Belvedere Museum dataset, a naturalistic egocentric benchmark collected in a real museum environment, where visitors and bystanders must be distinguished from artistic depictions such as paintings and statues. We present MOSAIC, a task-specific framework that combines agentic perceptual grounding, counterfactual identity planning, and mask-constrained diffusion-based video generation. Experiments on OpenHumanVid and Belvedere Museum show that MOSAIC achieves the strongest whole-body identity suppression among the evaluated generative methods across all three privacy metrics, while retaining competitive, metric-dependent utility, occupying a distinct operating point on the privacy–utility trade-off.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Video Anonymization, Privacy-Preserving Computer Vision, Person Re-Identification

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_XNReOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces Generative Video Anonymization (GVA), which aims to suppress whole body identity in videos while preserving action, scene context and realism.

The authors argue that general video editing models fail at this task and propose MOSAIC, a training free pipeline that combines multimodal grounding, counterfactual identity prompts and mask constrained inpainting with a pretrained diffusion transformer. The authors also introduce an egocentric museum dataset and a privacy–utility evaluation protocol.

Experiments on OpenHumanVid and the museum dataset show that MOSAIC achieves the lowest OSNet leakage among the evaluated generative methods, with metric dependent utility.

Critical strengths
  1. The paper addresses an important and timely problem: face only anonymization leaves body appearance exposed to person retrieval models.

  2. The proposed method is training free, and querying the multimodal model once per track is a practical way to separate visitors from artworks in long egocentric videos.

  3. The ablation shows that the counterfactual prompt matters: removing it raises Rank-1 on the museum data from 41.24% to 68.40%.

Critical weaknesses
  1. The privacy evaluation does not support the whole body claim. The retrieval protocol is underspecified (lines 613–614): the number of identities, gallery size, cropping and chance level are not given, and Mask Out still reaches 20.23% Rank-1 on the museum data. MOSAIC inpaints inside the original mask and keeps the motion, so silhouette and gait survive by design, yet the only attacker is OSNet, an appearance model. A gait or clothes changing attacker is needed.

  2. The utility evidence is weak. TWE is 0.00 for the original video, which Eq. (9) cannot produce, and on the museum data Acons is 82.95 for Mask Out against 84.87 for MOSAIC. On OpenHumanVid, Gaussian blur beats MOSAIC on five of six metrics, and MOSAIC has the highest TWE (17.54). No realism metric is reported.

  3. Reproducibility and ablation. The backbone and multimodal model are not identified (three LMMs are cited, line 208), the detector and tracker appear only as "YOLO" and "SAM" in Fig. 2, and no versions, thresholds or prompt templates are given. Table 3 does not define the ablated variants, removing Stage I improves privacy, and no variance or number of runs is reported. Novelty is modest; ReGenHuman (arXiv 2606.14972) should be discussed.

  4. The comparison does not isolate the contribution of the counterfactual prompt. The two generative baselines, described as task matched (line 353), do not share MOSAIC's backbone: DeepPrivacy2 works frame by frame (line 458) and UniVideo is a general editor.

Moreover, Stages I and II share a single LMM call (line 197), so removing Stage II may also change the grounding input. A baseline with the same backbone and masks but a generic caption is needed to show that the attribute opposition, rather than any appearance changing prompt, drives the privacy gain.

Questions for authors
  1. Could the authors specify how the gallery is built (number of identities, queries, crops, chance level), and explain why Mask Out keeps 20.23% Rank-1 on the museum data?

  2. How does MOSAIC perform against a gait or silhouette based attacker evaluated on the same clips?

  3. Which models and settings are used, and what does each ablation row in Table 3 run in place of the removed stage?

  4. What is OSNet Rank-1 when the gallery is drawn from a different clip of the same person?

2 comments

We thank the reviewer for the thorough review.

Gallery, chance level, Mask Out (Q1)

Evaluation uses the complete splits: 49 Belvedere clips and 256 OHV clips. A clip may contain multiple people; each detected person track is treated as a pseudo-identity. Because neither dataset provides cross-clip identity labels, tracks cannot be linked across clips. After a fixed shuffle (seed 42), the gallery contains anonymized person crops for all tracks and the queries contain the corresponding original crops. All methods use the same YOLOv8n person-crop protocol (confidence 0.5); random Rank-1 chance is the inverse of the number of gallery identities. The gallery includes observations from the same track without excluding temporally corresponding or adjacent content. Sec. 4.2 and Appendix A state the query/gallery direction backwards; we will correct them.

Mask Out retains 20.23% because detected box crops still contain clip-specific background and mask shape.

Gait or clothes-changing attacker (W1, Q2)

We added GaitBase (OpenGait, CASIA-B), CSCI (EVA-02-L, PRCC; CAL has no released weights) and TransReID (ViT-B/16 + JPM, Market-1501).

MethodBelvedere OSNetBelvedere TransReIDBelvedere CSCIBelvedere GaitOHV OSNetOHV TransReIDOHV CSCIOHV Gait
Blur61.5884.3880.84100.032.4089.6892.1273.33
Mask Out20.2318.3831.5457.140.5622.4140.2881.82
DeepPrivacy267.1335.2562.2493.3366.6023.577.0080.00
UniVideo84.1783.8982.89100.075.0079.5377.9285.71
ReGenHuman (new)61.8742.6758.2673.3375.3583.0483.0578.57
MOSAIC41.2425.7822.8960.0059.9647.8948.4466.67

<sub>Rank-1 (%); all methods are evaluated on the complete benchmark splits under the same protocol.</sub>

Under CSCI, MOSAIC reduces identity leakage on Belvedere to 22.89 Rank-1 (vs 62.24 DeepPrivacy2 and 58.26 ReGenHuman); on OHV, DeepPrivacy2 is lower under TransReID and CSCI, with lower AconsA_{\mathrm{cons}} (48.78 vs 73.32). MOSAIC has the lowest gait Rank-1 among the generative methods (60.00% Belvedere; 66.67% OHV), but both remain far above chance. We will therefore remove the "disjoint feature spaces" claim, limit the demonstrated scope to appearance-based whole-body identity, and leave gait anonymization as future work.

Utility (W2)

  • TWE: You are correct that Eq. 9 does not match the implementation. We compute the mean squared difference between the RAFT flows of consecutive original and anonymized frames. Original is therefore zero by construction, and the different dataset scales reflect resolution and motion magnitude. This measures motion fidelity to the source rather than standalone temporal smoothness; we will correct the equation, rename it optical-flow discrepancy, and soften the temporal-consistency claim on OHV.
  • AconsA_{\mathrm{cons}}: it is clip-embedding similarity rather than action accuracy; preserved scene context helps explain Mask Out's 82.95, so we treat it only as a proxy.
  • SPS: it measures unchanged background regions and should not be read as realism.
  • Blur on OHV: although strong under the original metrics, it leaks 89.68 / 92.12 Rank-1 under TransReID / CSCI and does not synthesize a realistic replacement. We did not report a direct realism metric and will state this limitation.

Models, settings, templates (W3, Q3)

Stage III uses frozen Wan2.1-VACE-1.3B (81 frames, 480p, 50 UniPC steps, CFG 5.0, shift 16, seed 2025), and Stages I–II use Qwen2-VL-2B-Instruct. The grounding module combines SAM3 proposals, global–local multimodal verification and temporally completed masks; the planning module reasons over four keyframes. Evaluation uses YOLOv8n at confidence 0.5.

The prompts instantiate Secs. 3.1–3.2. Stage I asks the LMM to determine whether the highlighted subject is a real person, artwork or shadow using both the full scene and local crop. Stage II separates identity attributes from action/scene semantics and instructs: "Rewrite the description with the counterfactual person while keeping the action and scene exactly unchanged." For example, "a young woman in a red jacket walking through a museum" becomes "an elderly man in a dark jacket walking through the same museum." Complete templates and category sets will be added to the appendix.

Ablations (W3, Q3)

w/o Stage III keeps the generator, masks and sampling settings fixed but feeds the raw video instead of the identity-erased masked video as visual conditioning. As reported in Table 3, TWE rises from 183.35 to 402.15 and AconsA_{\mathrm{cons}} drops from 84.87 to 72.40.

Removing Stage I lowers Rank-1 slightly because every person-like region, artworks included, is regenerated, but SPS drops to 0.542. All results are single runs (seed 2025).

Isolating the counterfactual prompt (W4)

The Table 3 control fixes masks, generator and sampling settings (68.40% vs 41.24% Rank-1), isolating counterfactual prompt construction.

Novelty and ReGenHuman (W3)

ReGenHuman (arXiv 2606.14972) is concurrent work; we ran its official code (Wan2.1-VACE-1.3B + released LoRA, SAM3 masks) under our protocol.

DatasetMethodR-1 ↓\downarrowSPS ↑\uparrowTWE ↓\downarrowAconsA_{\mathrm{cons}} ↑\uparrow
BelvedereReGenHuman61.870.768355.7384.64
BelvedereMOSAIC41.240.815183.3584.87
OHVReGenHuman75.350.87955.2272.28
OHVMOSAIC59.960.87817.5473.32

On the challenging museum data, ReGenHuman shows higher identity leakage under all appearance attackers and higher TWE. Qualitatively, regenerated visitors often retain source clothing colours, hairstyle and accessories (e.g. glasses), while depicted people may also be edited. We will add these examples to the appendix and cite and discuss ReGenHuman.

We see MOSAIC's contribution in grounding real visitors versus depicted humans, in the Re-ID-aware counterfactual prompt, and in the Belvedere benchmark.

Different clip of the same person (Q4)

This requires cross-clip identity labels, which neither dataset has; we list it as a limitation.

Ethics

The study had university ethics approval; camera wearers gave written consent, and museum visitors were notified by exhibition signage. We will provide the committee details in the camera-ready version. The masks and masked videos described in Sec. 4.1 are the intended release; release remains subject to museum approval and a specified access policy, and synchronized gaze data are not part of the current release. We will clarify this and add an Ethics Statement.

Reviewer_3HBNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies video anonymization that changes a person's whole-body appearance while retaining actions and scene context. MOSAIC combines track-level human/artwork filtering, counterfactual descriptions of identity-related attributes, and a pretrained video-inpainting model. It also introduces the Belvedere Museum recordings and evaluates privacy and utility on these recordings and OpenHumanVid. MOSAIC reduces retrieval performance for a frozen OSNet relative to the two generative baselines, although its utility varies considerably across metrics. I find the application worthwhile, especially the need to distinguish museum visitors from depicted people, but the evaluation does not yet establish the broader identity-suppression claim.

Critical strengths
  1. The museum setting gives the problem a concrete and useful motivation. Treating a painted person and a real visitor differently is an important requirement that a purely category-based mask can miss. The Stage I ablation is informative here: removing it reduces SPS from 0.815 to 0.542, suggesting that semantic filtering matters for preserving the scene.
  2. The results show a meaningful improvement over the evaluated generative baselines on the stated OSNet protocol. On Belvedere, Rank-1 drops from 67.13% for DeepPrivacy2 to 41.24%, and mAP from 26.95% to 17.41%. I also appreciate that the paper acknowledges the competing utility rankings rather than claiming to win every metric. The Stage II ablation, where Rank-1 rises to 68.40%, makes counterfactual prompting worth investigating further.
Critical weaknesses
  1. The privacy evidence is narrower than the problem statement. Rank-1, mAP, and cosine similarity all come from the same frozen OSNet, so they are three views of one representation rather than independent tests of whole-body privacy. Appendix A describes anonymized frames as queries and original frames as the gallery, but does not specify identity counts, person-crop construction, or exclusion of corresponding and adjacent frames. These details matter particularly for two continuous museum recordings. Even within this protocol, Rank-1 remains 59.96% on OpenHumanVid and 41.24% on Belvedere. This supports reduced retrieval success, but is not persuasive evidence of strong anonymization. The claim in Section 2 that identity and utility occupy disjoint feature spaces is also too strong: the paper itself identifies gait as an identity cue while aiming to preserve motion. A gait-sensitive or otherwise independent attacker is needed to test this tension. Section 5 acknowledges adaptive attacks and gait as future work, but these limitations already affect the central claim.
  2. The method and utility evaluation leave too much unresolved to interpret the gains reliably. Section 3.3 does not identify the actual pretrained video checkpoint or sampling configuration; the LMM version, evaluation subset sizes, and the replacement used for “w/o Stage III” are also unspecified. More concretely, Eq. (9) defines TWE as a within-video warping residual, yet both Original rows report exactly zero. An original video need not have zero warping error under that definition. If the implementation measures excess error relative to the source, the formula should say so. MOSAIC also has the highest TWE in Table 1 (17.54), despite temporal consistency being a central motivation. R(2+1)D feature similarity alone does not establish preserved action recognition, and no direct realism evaluation is reported. These are consequential gaps because the method's value depends on what utility is retained in exchange for the residual privacy risk.
Questions for authors
  1. How many evaluated clips and distinct identities are there in each dataset, and exactly how are person crops, queries, and galleries constructed? Do the galleries contain the original frame corresponding to each query?
  2. What TWE computation produced the zero Original entries and the very different numerical scales across datasets? Please reconcile the implementation with Eq. (9). Are action-label accuracy or subject-level temporal-consistency results available to support the feature-similarity measurements?
  3. Which LMM and video-generation checkpoints were used, with what prompts and inference settings? What replaces Stage III in Table 3, and are the masks and generation settings otherwise held fixed?
2 comments

e thank the reviewer for the detailed and balanced review.

Three views of one representation; "disjoint feature spaces" (W1)

We added TransReID (ViT-B/16 + JPM, Market-1501), the clothes-changing CSCI (EVA-02-L, PRCC; CAL has no released weights) and GaitBase (OpenGait, CASIA-B).

MethodBelvedere OSNetBelvedere TransReIDBelvedere CSCIBelvedere GaitOHV OSNetOHV TransReIDOHV CSCIOHV Gait
Blur61.5884.3880.84100.032.4089.6892.1273.33
Mask Out20.2318.3831.5457.140.5622.4140.2881.82
DeepPrivacy267.1335.2562.2493.3366.6023.577.0080.00
UniVideo84.1783.8982.89100.075.0079.5377.9285.71
ReGenHuman (new)61.8742.6758.2673.3375.3583.0483.0578.57
MOSAIC41.2425.7822.8960.0059.9647.8948.4466.67

<sub>Rank-1 (%); all methods are evaluated on the complete benchmark splits under the same protocol.</sub>

On Belvedere, MOSAIC consistently reduces identity leakage relative to the other generative methods under all three appearance attackers; on OHV, DeepPrivacy2 is lower under TransReID and CSCI, at a much lower AconsA_{\mathrm{cons}} (48.78 vs 73.32). As you anticipated: MOSAIC has the lowest gait Rank-1 among the generative methods (60.00% Belvedere; 66.67% OHV), but both remain far above chance. We will therefore remove the "disjoint feature spaces" claim, limit the demonstrated scope to appearance-based whole-body identity, and leave gait anonymization as future work.

Protocol (Q1)

Evaluation uses the complete splits: 49 Belvedere clips and 256 OHV clips. A clip may contain multiple people; each detected person track is treated as a pseudo-identity. Because neither dataset provides cross-clip identity labels, tracks cannot be linked across clips. After a fixed shuffle (seed 42), the gallery contains anonymized person crops for all tracks and the queries contain the corresponding original crops. All methods use the same YOLOv8n person-crop protocol (confidence 0.5); random Rank-1 chance is the inverse of the number of gallery identities. The gallery includes observations from the same track without excluding temporally corresponding or adjacent content. Sec. 4.2 and Appendix A state the query/gallery direction backwards; we will correct them.

Belvedere clips come from two continuous recordings, so one visitor may occur in several clips; without cross-clip labels, tracks from different clips cannot be linked and are treated as distinct pseudo-identities. We make no cross-clip identity claim.

TWE (Q2)

You are correct that Eq. 9 does not match the implementation. We compute the mean squared difference between the RAFT flows of consecutive original and anonymized frames. Original is therefore zero by construction, and the different dataset scales reflect resolution and motion magnitude. This measures motion fidelity to the source rather than standalone temporal smoothness; we will correct the equation, rename it optical-flow discrepancy, and soften the temporal-consistency claim on OHV.

Action accuracy, realism (Q2)

Action-label accuracy is unavailable because the clips have no action annotations. AconsA_{\mathrm{cons}} measures R(2+1)D clip-embedding similarity and is therefore only a proxy for action preservation. SPS measures background preservation, not realism; we report neither a direct realism metric nor subject-level temporal consistency and will state these limitations.

Checkpoints, settings, templates (Q3)

Stage III uses frozen Wan2.1-VACE-1.3B (81 frames, 480p, 50 UniPC steps, CFG 5.0, shift 16, seed 2025), and Stages I–II use Qwen2-VL-2B-Instruct. The grounding module combines SAM3 proposals, global–local multimodal verification and temporally completed masks; the planning module reasons over four keyframes. Evaluation uses YOLOv8n at confidence 0.5.

The prompts instantiate Secs. 3.1–3.2. Stage I asks the LMM to determine whether the highlighted subject is a real person, artwork or shadow using both the full scene and local crop. Stage II separates identity attributes from action/scene semantics and instructs: "Rewrite the description with the counterfactual person while keeping the action and scene exactly unchanged." For example, "a young woman in a red jacket walking through a museum" becomes "an elderly man in a dark jacket walking through the same museum." Complete templates and category sets will be added to the appendix.

Table 3 settings (Q3)

w/o Stage II keeps the generator, masks and sampling settings fixed and replaces the counterfactual prompt with a generic caption. As reported in Table 3, Rank-1 rises from 41.24% to 68.40%, while SPS (0.810 vs. 0.815) and AconsA_{\mathrm{cons}} (85.10 vs. 84.87) remain similar.

w/o Stage III keeps the generator, masks and sampling settings fixed but feeds the raw video instead of the identity-erased masked video as visual conditioning. As reported in Table 3, TWE rises from 183.35 to 402.15 and AconsA_{\mathrm{cons}} drops from 84.87 to 72.40.

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

The paper introduces a training-free pipeline for generative video anonymization using whole-body identity replacement. A large multimodal model distinguishes people from artworks depicting people and constructs counterfactual appearance prompts to edit identity. A frozen video diffusion transformer generates replacements conditioned on the masked video and prompt. The paper also introduces a museum recording dataset to highlight the people vs. artwork problem not covered by prior works. The paper evaluates privacy and utility on OpenHumanVid and this new dataset. MOSAIC reduces re-id retrieval relative to existing generative anonymization methods, showing reasonable privacy-utility trade-offs.

Critical strengths
  • Important privacy problem, unique problem setting: Whole-body identity leakage is an important problem that face-only editing does not resolve. The museum setting gives a concrete reason to distinguish actual people from paintings and statues. The ablation supports the benefit of filtering for scene preservation, rather than treating all person-like detections as valid privacy targets.

  • Promising empirical trade-off among the evaluated generative methods: The proposed method achieves the lowest identity leakage while maintaining reasonable utility scores.

Critical weaknesses
  • The privacy evaluation is not convincing: It relies on one re-identification model, and identity retrieval remains relatively high, even higher than simple blurring. The paper motivates protecting against gait-based identification but does not test a gait- or video-based attacker. It also leaves key details about the query-gallery re-id setup unclear, making it hard to tell how much the reported results demonstrate genuine privacy protection.

  • Comparison to closely related work is missing: ReGenHuman [1] already studies full-body video anonymization with video diffusion, but the paper neither cites nor compares against it. Key implementation details such as the DiT model/conditioning specifics and the LMM used are also missing, making it difficult to assess the contributions or reproduce the system.

Citations:

[1] Sun, Adam, et al. "ReGenHuman: Re-Generating Human Appearances for Realistic Full-Body Video Anonymization." arXiv preprint arXiv:2606.14972 (2026).

Questions for authors
  1. What are the exact evaluation video and identity counts on each dataset? How are queries and galleries constructed? What pretrained generator and LMM checkpoints implement MOSAIC, with which prompts and sampling settings?

  2. Can the authors compare with ReGenHuman [1] and isolate the contribution of counterfactual prompts from a generic identity prompt under the same generator?

  3. Do the claimed privacy and semantic benefits hold against a gait/video Re-ID attacker and on an actual labeled downstream task?

2 comments

We thank the reviewer for the constructive review.

More Re-ID models; blur; gait (W1, Q3)

We added TransReID (ViT-B/16 + JPM, Market-1501), the clothes-changing CSCI (EVA-02-L, PRCC; CAL has no released weights) and GaitBase (OpenGait, CASIA-B, YOLOv8n-seg silhouettes).

MethodBelvedere OSNetBelvedere TransReIDBelvedere CSCIBelvedere GaitOHV OSNetOHV TransReIDOHV CSCIOHV Gait
Blur61.5884.3880.84100.032.4089.6892.1273.33
Mask Out20.2318.3831.5457.140.5622.4140.2881.82
DeepPrivacy267.1335.2562.2493.3366.6023.577.0080.00
UniVideo84.1783.8982.89100.075.0079.5377.9285.71
ReGenHuman (new)61.8742.6758.2673.3375.3583.0483.0578.57
MOSAIC41.2425.7822.8960.0059.9647.8948.4466.67

<sub>Rank-1 (%); all methods are evaluated on the complete benchmark splits under the same protocol.</sub>

On Belvedere, MOSAIC consistently reduces identity leakage relative to the other generative methods under all three appearance attackers, so the improvement is not specific to OSNet. Blur beating MOSAIC on OHV is specific to OSNet (TransReID / CSCI: 89.68 / 92.12 for blur vs 47.89 / 48.44). On OHV, DeepPrivacy2 is lower under TransReID and CSCI, at a much lower AconsA_{\mathrm{cons}} (48.78 vs 73.32).

MOSAIC has the lowest gait Rank-1 among the generative methods (60.00% Belvedere; 66.67% OHV), but both remain far above chance. We will therefore remove the "disjoint feature spaces" claim, limit the demonstrated scope to appearance-based whole-body identity, and leave gait anonymization as future work.

Protocol (Q1)

Evaluation uses the complete splits: 49 Belvedere clips and 256 OHV clips. A clip may contain multiple people; each detected person track is treated as a pseudo-identity. Because neither dataset provides cross-clip identity labels, tracks cannot be linked across clips. After a fixed shuffle (seed 42), the gallery contains anonymized person crops for all tracks and the queries contain the corresponding original crops. All methods use the same YOLOv8n person-crop protocol (confidence 0.5); random Rank-1 chance is the inverse of the number of gallery identities. The gallery includes observations from the same track without excluding temporally corresponding or adjacent content. Sec. 4.2 and Appendix A state the query/gallery direction backwards; we will correct them.

Checkpoints, settings, templates (Q1)

Stage III uses frozen Wan2.1-VACE-1.3B (81 frames, 480p, 50 UniPC steps, CFG 5.0, shift 16, seed 2025), and Stages I–II use Qwen2-VL-2B-Instruct. The grounding module combines SAM3 proposals, global–local multimodal verification and temporally completed masks; the planning module reasons over four keyframes. Evaluation uses YOLOv8n at confidence 0.5.

The prompts instantiate Secs. 3.1–3.2. Stage I asks the LMM to determine whether the highlighted subject is a real person, artwork or shadow using both the full scene and local crop. Stage II separates identity attributes from action/scene semantics and instructs: "Rewrite the description with the counterfactual person while keeping the action and scene exactly unchanged." For example, "a young woman in a red jacket walking through a museum" becomes "an elderly man in a dark jacket walking through the same museum." Complete templates and category sets will be added to the appendix.

ReGenHuman (Q2)

ReGenHuman (arXiv 2606.14972) is concurrent work; we ran its official code (Wan2.1-VACE-1.3B + released LoRA, SAM3 masks) under our protocol.

DatasetMethodR-1 ↓\downarrowSPS ↑\uparrowTWE ↓\downarrowAconsA_{\mathrm{cons}} ↑\uparrow
BelvedereReGenHuman61.870.768355.7384.64
BelvedereMOSAIC41.240.815183.3584.87
OHVReGenHuman75.350.87955.2272.28
OHVMOSAIC59.960.87817.5473.32

On the challenging museum data, ReGenHuman shows higher identity leakage under all appearance attackers and higher TWE. Qualitatively, regenerated visitors often retain source clothing colours, hairstyle and accessories (e.g. glasses), while depicted people may also be edited. We will add these examples to the appendix and cite and discuss ReGenHuman.

Counterfactual vs. generic prompt (Q2)

w/o Stage II keeps the generator, masks and sampling settings fixed and replaces the counterfactual prompt with a generic caption. As reported in Table 3, Rank-1 rises from 41.24% to 68.40%, while SPS (0.810 vs. 0.815) and AconsA_{\mathrm{cons}} (85.10 vs. 84.87) remain similar.

Labeled downstream task (Q3)

The clips have no action labels, so AconsA_{\mathrm{cons}} is the cosine similarity of R(2+1)D clip embeddings, not action-label accuracy. We use it only as a proxy for retained action semantics and will state this scope explicitly.

Ethics

The study had university ethics approval; camera wearers gave written consent, and museum visitors were notified by exhibition signage. We will provide the committee details in the camera-ready version. The masks and masked videos described in Sec. 4.1 are the intended release; release remains subject to museum approval and a specified access policy, and synchronized gaze data are not part of the current release. We will clarify this and add an Ethics Statement.

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

Related papers

Discussion (0)

Sign in to comment.