acceptodds
Under review as a conference paper at ICLR 2027

Expression-Diverse References for Identity-Preserving Video Generation

Abstract

Identity-preserving video generation aims to maintain a subject's identity while synthesizing realistic videos. Yet a single reference portrait captures the subject's appearance under only one facial configuration. As expressions change, facial appearance can vary in highly identity-specific ways, leaving the subject's appearance under unseen expressions underdetermined by the reference alone. This expression-dependent variation also complicates evaluation: similarity to a neutral reference may decrease under strong expressions even for real images of the same person. We investigate this limitation from both generation and evaluation perspectives. First, we quantify how face-recognition similarity varies with expression intensity using controlled photographs and MEAD videos. We then construct a compact yet expressive reference gallery that captures diverse expression-dependent facial configurations. Matching against this gallery provides a more robust measure of identity similarity under expressive motion. To further expose performance degradation with expression intensity, we report identity similarity separately for mild, intense, and extreme expressions. For generation, we extend Stand-In to condition on our expression-diverse reference sets and develop a data-curation pipeline that extracts consistent yet diverse face crops from training videos. In practical settings where only a single portrait is available, we construct the reference set by synthesizing additional expressions with a pretrained facial reenactment model. On our controlled benchmark of 10 subjects and 48 prompts per subject, both real and synthesized reference sets outperform the evaluated baselines in identity similarity across all three expression-intensity regimes, with the largest improvements for extreme expressions. In this regime, nearest-gallery CurricularFace similarity increases from 60.95 for the strongest baseline to 74.80 with real reference sets and 68.36 with synthesized reference sets. These results show that representing an identity with an expression-diverse reference gallery substantially improves identity preservation under expressive motion.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video Generation Models, Diffusion Models, Avatar, Subject-to-Video, Identity-Preserving Video Generation

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

The paper studies how strong facial expressions affect identity preservation and its evaluation in video generation. It extends Stand-In to use expression-diverse reference images and proposes gallery-based identity scoring across expression-intensity ranges. Results show improved similarity scores, although it remains unclear how much these gains reflect better identity preservation versus reduced expression intensity.

Critical strengths

The paper highlights an important evaluation issue: genuine expression changes can lower face-recognition similarity even when identity is unchanged. It supports this observation with real photographs and MEAD data, and includes useful ablations on reference count and expression diversity.

Critical weaknesses
  1. Showing that neutral-reference similarity decreases with expression intensity does not establish that taking the maximum over an expression gallery better measures identity. The same gallery also supplies the model’s conditioning images, making it difficult to distinguish identity generalization from reference matching.
  2. Table 2 shows that the proposed method produces fewer extreme-expression frames than several baselines. It is therefore unclear whether its higher identity scores reflect better identity preservation under equally challenging expressions. Table 3 considers peak intensity, which does not establish comparable expression strength throughout the video. A comparison at matched expression intensity would help resolve this concern.
  3. The proposed model is trained on 400k curated clips, whereas the baseline comparisons use official checkpoints. The existing ablations do not fully isolate the multi-reference extension from the benefits of this training.
  4. The authors do not discuss or compare with recent paper [1], which addresses closely related questions about identity preservation, expressive behaviour and the effect of reference expressions on identity evaluation. Comparisons with other recent methods, such as Slot-ID and Mv²ID, are also missing. The authors should clarify the novelty relative to these approaches and provide comparisons under compatible reference inputs, or explain why a fair comparison is not feasible.

[1] Baptiste, P., Quignon, N., Bahram, Y., Zeeshan, M. O., Dantcheva, A., & Granger, E. (2026). BEACON: Behavior and Appearance Control for Subject-Specific Video Generation. arXiv preprint arXiv:2609.13264.

Questions for authors
  1. The authors must evaluate with held-out reference photographs and test whether gallery scoring better separates genuine and impostor identities or agrees more closely with human judgments.
  2. Experiment by reporting the subset sizes in Table 3 and compare identity on matched subject–prompt pairs after controlling for the intensity distribution, rather than peak intensity alone.
  3. Provide a single-reference Stand-In control trained on the same data and budget, or clarify whether such a control already exists. For the synthetic setting, giving AnyID the same synthesized references would help isolate the benefit of the proposed conditioning mechanism.
3 comments
AuthorsOpenReview ↗

Response Part 1 of 3

Thank you for recognizing the importance of expression-aware identity evaluation and for highlighting the need to separate the effects of scoring, expression intensity, and additional training. We have added experiments and clarified the discussion to address these concerns.

Validation of Nearest-Gallery Identity Scoring (W1, Q1)

We agree that taking the maximum over a gallery requires validation beyond an increase in genuine similarity. Table 2 adds a leave-one-out evaluation on real photographs, excluding each query's expression from every subject's gallery. Nearest-gallery scoring reduces the mild-to-extreme gap in genuine similarity while slightly widening the margins against both mean and hardest-impostor scores.

Table 2. Leave-one-out metric validation.

RecognizerGalleryImpostor meanImpostor hardestGenuine meanMargin meanMargin hardestGenuine mildGenuine intenseGenuine extremeDifference ↓
ArcFaceSingle4.0214.4177.0673.0462.6583.8976.3871.3612.53
ArcFaceNearest8.1918.3281.6973.5163.3784.7381.1879.675.06
CurricularFaceSingle4.9214.0479.4574.5465.4186.7578.1874.6412.10
CurricularFaceNearest8.6217.5783.5474.9165.9787.1082.7481.605.50

Table 3 adds a study with six human evaluators. Nearest-gallery rankings agree with 79.3% of individual judgements and 88.6% of pairwise majorities, compared with 62.7% and 68.2% for single-reference rankings. These results provide complementary evidence from generated frames.

Table 3. Agreement with human identity judgements.

GroupPairsVotesVotes: SingleVotes: Nearest-galleryMajorities: SingleMajorities: Nearest-gallery
Flip1810832.4%67.6%26.7%73.3%
All5030062.7%79.3%68.2%88.6%

We also copy the held-out-expression evaluation from the appendix below. The query photographs depict four expressions excluded from the eight-image gallery. Nearest-gallery scores vary less across these expressions than neutral-reference scores, showing that the metric's robustness extends to expressions outside the gallery.

Table 11. Identity similarity on held-out real expressions.

Held-out expressionArcFace: NeutralArcFace: Nearest galleryCurricularFace: NeutralCurricularFace: Nearest gallery
H178.05 ± 3.3580.51 ± 3.0578.70 ± 3.7681.03 ± 3.32
H275.35 ± 3.5380.94 ± 3.5976.89 ± 3.4382.33 ± 3.34
H377.85 ± 2.7981.05 ± 2.5079.96 ± 2.1782.78 ± 2.32
H479.17 ± 2.8081.21 ± 2.2580.42 ± 2.3082.02 ± 2.13
AuthorsOpenReview ↗

Response Part 2 of 3

Comparison at Matched Expression Intensity (W2, Q2)

We agree that matching peak intensity alone does not control the intensity distribution throughout a video. Table 6 (Table 3 in the original submission) now reports the proportions of video and frame pairs included in the higher-intensity subsets, together with the additional frame-level comparisons.

To control for the intensity distribution more directly, we added Table 13 in Appendix B.2. Within each subject–prompt pair, we divide frames into expression-intensity bins of width 0.5. For each bin reached by both models, we compare their mean nearest-gallery similarity. We then average the differences, weighting each bin by the smaller of the two frame counts. Thus, each comparison holds subject, prompt, and expression-intensity bin fixed.

Table 13. Identity difference at matched expression intensity.

Ours (Real) minus each baseline:

BaselineAll: ΔArcAll: ΔCurWR Arc %WR Cur %Mild: ΔArcMild: ΔCurIntense: ΔArcIntense: ΔCurExtreme: ΔArcExtreme: ΔCur
Concat-ID+19.64+22.0299%100%+15.62+17.62+20.36+22.88+27.47+29.37
Phantom+8.02+8.4392%92%+4.48+5.93+7.82+8.24+12.18+11.63
ConsisID+41.21+42.06100%100%+34.77+36.32+40.69+41.53+47.37+47.91
Stand-In+18.55+19.6697%97%+6.88+7.31+20.13+21.44+26.50+27.46
AnyID+13.39+14.2898%99%+11.04+11.78+13.67+14.56+13.61+14.69
Gen-4.5+21.08+21.9999%100%+12.47+12.96+22.22+23.25+30.87+31.82

Ours (Synthetic) minus each baseline:

BaselineAll: ΔArcAll: ΔCurWR Arc %WR Cur %Mild: ΔArcMild: ΔCurIntense: ΔArcIntense: ΔCurExtreme: ΔArcExtreme: ΔCur
Concat-ID+17.59+18.8898%98%+15.40+15.62+17.84+19.37+17.84+18.49
Phantom+4.89+4.6079%77%+1.71+2.49+4.10+4.04+8.27+6.99
ConsisID+37.58+38.21100%100%+32.39+32.23+37.12+37.71+40.08+40.97
Stand-In+16.41+16.7395%94%+8.02+7.08+16.35+16.99+19.77+19.51
AnyID+9.14+9.4093%93%+9.89+9.66+9.33+9.65+7.95+8.08
Gen-4.5+19.90+20.3499%100%+13.33+12.27+19.77+20.42+24.55+25.15

Both reference settings retain positive identity-similarity differences against every evaluated baseline in all three intensity regimes under both recognizers. These results support an identity-fidelity advantage after controlling for expression intensity under this protocol.

AuthorsOpenReview ↗

Response Part 3 of 3

Single-Reference Stand-In and Matched-Reference AnyID Controls (W3, Q3)

We have added the requested single-reference training control. We finetune the original Stand-In model on our curated OpenHumanVid data using the same training recipe as our multi-reference model, except that each training clip uses only its sharpest reference image.

Table 5 (last three rows). Single-reference training control.

ModelAll: Arc ↑All: Cur ↑Mild: %Mild: Arc ↑Mild: Cur ↑Intense: %Intense: Arc ↑Intense: Cur ↑Extreme: %Extreme: Arc ↑Extreme: Cur ↑
Finetuned Stand-In64.6565.0629.483.4784.7939.364.7564.8931.346.8546.74
Ours (Synthetic)76.3576.5414.983.1684.3171.076.4876.5314.168.4868.36
Ours (Real)80.7881.6829.083.6885.0363.580.2380.957.474.2774.80

Finetuning improves Stand-In's overall identity similarity, but our multi-reference model retains higher scores overall and in the intense and extreme regimes under both real and synthetic reference settings. This indicates that the gains cannot be attributed solely to additional training on our curated data.

We also compare our model with AnyID using identical references in both settings: five real images (5), or one real neutral portrait and four LivePortrait-generated expressions (1R+4S).

Table 8 (first four rows). Matched five-reference comparison.

ModelRefs.All: Arc ↑All: Cur ↑Mild: %Mild: Arc ↑Mild: Cur ↑Intense: %Intense: Arc ↑Intense: Cur ↑Extreme: %Extreme: Arc ↑Extreme: Cur ↑
AnyID562.8362.747.873.8673.7754.466.3666.0237.855.4855.77
AnyID1R+4S60.9560.825.974.8074.5252.366.1365.3741.852.5353.18
Ours577.2277.7724.983.9285.0560.677.0377.3914.566.5166.82
Ours1R+4S74.4274.4915.482.8783.2767.874.8074.8316.865.1165.06

Our model retains higher identity similarity in both settings, supporting its benefit when the reference inputs are held fixed.

Comparison with Related Work (W4)

We have expanded the related-work section to discuss Mv$^2$ID, Slot-ID, and BEACON and clarify how their reference inputs and objectives relate to ours:

  • Mv$^2$ID uses multiview portraits to improve robustness to pose variation. Viewpoint changes describe how the same facial geometry is observed from different directions; expressions change the facial geometry itself through combinations of local deformations with varying intensity. Their coverage is consequently less intuitive: several visibly different expressions may provide similar information about one facial region while leaving other deformations unrepresented. A set of reference angles therefore does not directly translate into a criterion for selecting complementary expressions. This motivates our standardized eight-image expression gallery, which makes the choice of reference deformations explicit and reproducible. We also draw on pretrained facial reenactment models to synthesize the expressive references from a single portrait, using the prior knowledge of facial deformation encoded in LivePortrait.

  • Slot-ID and BEACON use video references to capture expressive variation, but do not specify explicit criteria for the expression coverage of those videos or a principled procedure for constructing suitable reference videos when only one portrait is available. The reference video's expression content and quality can therefore introduce uncontrolled differences into the comparison. We therefore cannot make a reliable, fair comparison under our reference-input setting without introducing additional assumptions about reference-video selection or synthesis. One of our contributions is to make both the reference content and its construction explicit: we define a standardized eight-image expression gallery, provide a LivePortrait-based procedure for generating it from a single portrait, and demonstrate identity-fidelity improvements using these synthetic galleries. Table 8 further compares our model with AnyID using exactly the same five-image synthetic gallery, controlling for the reference images supplied to both models. Our method thus provides a reproducible protocol for both captured and synthetic references, without requiring video-reference encoding or an additional VLM.

Thank you for considering the additional experiments and clarifications.

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

This paper studies identity-preserving video generation under strong facial expressions. The authors first observe that a single reference image provides limited information about identity-specific facial deformations, and that FR similarity is sensitive to expression change. To address these issues, the authors propose an expression-diverse reference gallery, a data-curation pipeline to select or synthesize diverse references, and extend Stand-In to support multi-reference conditioning. Experiments show improved identity similarity over existing methods, especially under strong expressions.

Critical strengths
  1. The paper is well motivated: The authors identify an important problem in identity-preserving video generation, i.e., a single portrait may not fully capture how a person’s facial appearance changes under different expressions. I particularly appreciate the analysis in Fig. 2/3 and Tab. 1, which provides strong evidence for the problem. This makes the motivation clear and convincing.
  2. The proposed pipeline is intuitive and practically effective: The authors extend an existing video generation model with multi-reference conditioning and provide a synthetic solution for the single-reference setting. The experimental results show substantial improvements.
Critical weaknesses
  1. While I appreciate the practical effectiveness of the proposed method, its novelty seems modest: The key paradigm of using multiple references appears already well established (e.g., AnyID, WildActor). I would appreciate a clearer discussion of what the proposed conditioning mechanism adds beyond these methods.
  2. Expression diversity is central to the paper, but I am not fully convinced that the data-curation pipeline reliably produces expression-diverse training references. In Sec. 4, the authors use ArcFace dissimilarity and temporal distance to select reference images. Given the observed expression sensitivity of ArcFace, this seems like a reasonable proxy, but it may also capture differences in pose, lighting, or image quality. While Table 5 provides useful evidence for the benefit of expression-diverse references at inference, some direct evidence on the expression diversity of the curated training images would strengthen this part of the paper.
Questions for authors
  1. The reference selection strategy in Sec. 4 is simple and reasonable. However, the authors could further provide some visualizations or quantitative evidence showing the expression diversity of the selected OpenHumanVid references. For example, measuring their coverage in SMIRK expression space and comparing it with uniform or random sampling.
  2. The authors compared the proposed method with AnyID under 5 reference images, which is very much appreciated. Could the authors clarify whether both models receive exactly the same 5 references? It would also be interesting to compare the two models using the same 5-image subset of the LP-generated references. This would help us understand how much the synthetic-reference strategy benefits an existing multi-reference model, and better identify the gains of the proposed method.
3 comments

Thank you for your positive assessment of the motivation and pipeline, and for identifying where the contribution and evidence needed clarification. We have expanded the related-work discussion and added the requested expression-diversity and matched-reference evaluations.

Comparison with Related Work (W1)

Our contribution centers on the choice and evaluation of expression-diverse references. WildActor and Mv$^2$ID use multiview portraits to improve robustness to pose variation. Facial deformation introduces another source of variation: a person's appearance changes with expression, and the set of images to coverage the majority of changes is not as clear as the set to cover changes in the view angle. We specify a standardized eight-image gallery to capture this variation and evaluate identity fidelity across expression-intensity regimes.

Methods such as Slot-ID and BEACON capture expressive variation using video references. Our protocol makes the reference expressions explicit and reproducible. It also supports cases where only one portrait is available: a pretrained facial reenactment model, LivePortrait, synthesizes the additional expressions. The generator extends Stand-In to accept multiple images without requiring video-reference encoding or an additional VLM. We have clarified these distinctions in the related-work section.

Expression Diversity of the Data-Curation Pipeline (W2, Q1)

We have added a quantitative evaluation in SMIRK expression space. Table 4 shows that our selected references have greater mean pairwise expression distance than references selected uniformly or randomly. The additional Effective Dimensions statistic is also higher for our selection than for either sampling baseline.

Table 4. Expression diversity of selected frames.

MethoddΨd_\Psi ↑Effective Dimensions ↑
Ours0.49042.76
Uniform0.44912.59
Random0.43132.45
Whole Dataset0.42373.23

Additional Comparisons with AnyID (Q2)

Yes, both models receive exactly the same five reference images in Table 8. We have also added the requested comparison using the same five-image LivePortrait gallery: one real neutral portrait and four synthetic expressions (1R+4S). The rows labelled 5 use five real references.

Table 8 (first four rows). Matched five-reference comparison.

ModelRefs.All: Arc ↑All: Cur ↑Mild: %Mild: Arc ↑Mild: Cur ↑Intense: %Intense: Arc ↑Intense: Cur ↑Extreme: %Extreme: Arc ↑Extreme: Cur ↑
AnyID562.8362.747.873.8673.7754.466.3666.0237.855.4855.77
AnyID1R+4S60.9560.825.974.8074.5252.366.1365.3741.852.5353.18
Ours577.2277.7724.983.9285.0560.677.0377.3914.566.5166.82
Ours1R+4S74.4274.4915.482.8783.2767.874.8074.8316.865.1165.06

Synthetic references yield slightly lower overall similarity than real references for both models. Our model retains higher identity similarity than AnyID in every reported intensity regime under both reference settings. This comparison controls for the reference images supplied to the two models.

Thank you for considering these clarifications and additional experiments.

Thank you for the detailed rebuttal and the additional experiments evaluating expression diversity and comparing both methods with matched references. I appreciate these efforts to further strengthen the paper’s experimental evaluation. The SMIRK analysis and the comparisons with AnyID address my main concerns.

Thank you for the quick follow-up and for acknowledging that the SMIRK analysis and the matched-reference comparisons with AnyID addressed your main concerns.

We noticed that the official rating remains at 2. Are there any remaining concerns underlying that assessment that we could possibly clarify? If your overall assessment has changed in light of the rebuttal, we would be grateful if you could update the rating to reflect your current view.

Thank you for your consideration.

Reviewer_BGR7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates identity-preserving video generation under large facial-expression changes and argues that a single reference image is insufficient for both generation and evaluation. It introduces expression-diverse reference galleries for multi-reference conditioning and gallery-based identity evaluation, showing substantial gains in identity preservation, especially for intense and extreme expressions.

Critical strengths
  1. The paper identifies an important and somewhat overlooked evaluation issue.
  2. The proposed evaluation protocol is simple but meaningful.
Critical weaknesses
  1. The nearest-gallery metric takes the maximum similarity over multiple references, which may naturally increase as the gallery size grows. Although the current ablations partially address this issue, stronger validation using genuine/impostor identity discrimination or metrics such as ROC-AUC, EER, and TAR at fixed FAR would make the metric more convincing.
  2. The quantitative evaluation mainly relies on ArcFace and CurricularFace similarity. Since the proposed conditioning strategy may also affect motion quality, temporal consistency, prompt alignment, and overall visual quality, additional video-level metrics and/or human evaluation would provide a more comprehensive assessment.
Questions for authors
  1. Could the authors provide additional evaluation on video quality, temporal consistency, and text-prompt alignment?
  2. Have the authors evaluated the proposed nearest-gallery similarity using genuine/impostor pairs? In particular, does the gallery-based metric improve robustness to expression while maintaining or improving identity discrimination across different subjects?
2 comments
AuthorsOpenReview ↗

Response Part 1 of 2

Thank you for recognizing the importance of the evaluation issue and for suggesting additional validation. We have expanded the experiments to address both identity discrimination and the broader quality of the generated videos.

Video Quality, Temporal Consistency, and Prompt Adherence (W2, Q1)

Our focus is on identity following, but it does not come at the cost of losing the video quality. We have added evaluations of video quality, temporal consistency, and prompt adherence. Table 7, copied below, shows that our video quality (flicker, motion smoothness, imaging quality), consistency (subject and background), and prompt following scores (VLM, CLIP) remain comparable to those of the evaluated baselines. These results complement the identity-fidelity evaluation and indicate that the identity gains are compatible with maintaining video quality and prompt adherence.

Table 7. Video quality and prompt adherence.

ModelSubject consistency ↑Background consistency ↑Flicker score ↑Motion smoothness ↑Imaging ↑VLM ↑CLIP ↑
Concat-ID98.4895.9099.3999.4774.153.6719.86
Phantom94.9791.9397.0697.6470.833.9420.95
ConsisID89.4390.9298.2899.1272.583.7821.48
Stand-In95.4093.6498.9899.2667.523.9719.45
AnyID96.1694.5898.9999.2668.003.9519.82
Gen-4.586.9288.6496.7397.6271.223.4720.84
Ours (Real)97.0795.2798.9899.2070.213.8619.59

The first five scores are VBench metrics. For prompt adherence, we sample eight frames per video: Qwen2.5-VL-7B-Instruct assigns a rating from 1 to 5, and CLIP ViT-L/14 measures frame-to-text cosine similarity.

AuthorsOpenReview ↗

Response Part 2 of 2

Validation of Nearest-Gallery Identity Scoring (W1, Q2)

The current face similarity metrics are already sufficient in telling the real photos of genuine and impostors apart, and we propose the nearest-gallery identity similarity metric to mitigate the fact that stronger emotions in generated images or videos are penalized in terms of face similarity. Still, we agree that higher similarity alone does not validate nearest-gallery scoring, since taking a maximum can increase impostor scores as well as genuine scores. We therefore evaluate whether it reduces sensitivity to expression changes while preserving genuine–impostor separation.

Table 2 reports a leave-one-out evaluation on 70 non-neutral photographs of 10 subjects. For each query, we exclude its expression from every subject's gallery, leaving seven reference images. We compare the query with its own subject's gallery and with the galleries of other subjects.

Table 2. Leave-one-out metric validation.

RecognizerGalleryImpostor meanImpostor hardestGenuine meanMargin meanMargin hardestGenuine mildGenuine intenseGenuine extremeDifference ↓
ArcFaceSingle4.0214.4177.0673.0462.6583.8976.3871.3612.53
ArcFaceNearest8.1918.3281.6973.5163.3784.7381.1879.675.06
CurricularFaceSingle4.9214.0479.4574.5465.4186.7578.1874.6412.10
CurricularFaceNearest8.6217.5783.5474.9165.9787.1082.7481.605.50

Nearest-gallery scoring reduces the mild-to-extreme gap from 12.53 to 5.06 for ArcFace and from 12.10 to 5.50 for CurricularFace. Although impostor scores also increase, the genuine–impostor margins become slightly larger for both recognizers. This supports improved expression robustness without reduced separation under this protocol.

We also conducted a study with six human evaluators. As shown in Table 3, nearest-gallery rankings agree with 79.3% of individual judgements, compared with 62.7% for single-reference rankings.

Table 3. Agreement with human identity judgements.

GroupPairsVotesVotes: SingleVotes: Nearest-galleryMajorities: SingleMajorities: Nearest-gallery
Flip1810832.4%67.6%26.7%73.3%
All5030062.7%79.3%68.2%88.6%

Finally, Table 11 evaluates real photographs of four expressions excluded from the eight-image gallery. Nearest-gallery scoring also reduces variation across these held-out expressions, supporting robustness beyond the expressions represented in the gallery.

Table 11. Identity similarity on held-out real expressions.

Held-out expressionArcFace: NeutralArcFace: Nearest galleryCurricularFace: NeutralCurricularFace: Nearest gallery
H178.05 ± 3.3580.51 ± 3.0578.70 ± 3.7681.03 ± 3.32
H275.35 ± 3.5380.94 ± 3.5976.89 ± 3.4382.33 ± 3.34
H377.85 ± 2.7981.05 ± 2.5079.96 ± 2.1782.78 ± 2.32
H479.17 ± 2.8081.21 ± 2.2580.42 ± 2.3082.02 ± 2.13

Thank you for considering these additional results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.