acceptodds
Desk rejected at ICLR 2027

ExDiff: Diffusion as an Expressive Prior for Hand Gestures and Facial Expressions

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Expressive hand gestures and facial expressions provide essential non-verbal cues for understanding human intent in image-based human-computer interac- tion. However, existing pose priors often focus on full-body motion or model the hands and face as independent components, limiting their ability to capture the dependencies between hand articulation and facial expression. We propose EXD- IFF, an expressive diffusion framework for jointly modeling bilateral hand ges- tures, jaw motion, and facial expressions within a unified expressive latent space. The proposed framework introduces part-balanced diffusion with part-aware noise schedules, cross-part residual fusion, and geometry-aware denoising supervision to model fine-grained hand articulation, facial dynamics, and their mutual depen- dencies. To support heterogeneous training, EXDIFF learns part-specific hand and face diffusion priors, then trains a zero-initialized cross-part residual fusion head on joint hand–face data, using part-availability masks and confidence-aware con- ditioning. The learned prior supports expressive generation, missing-part comple- tion, and monocular hand–face mesh recovery. Experiments show that EXDIFF consistently improves realism, coverage, and diversity on ARCTIC, FreiHAND, NOW, and BEAT2. The code will be made available to the public.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Diffusion Prior, Hand Gestures, Facial Expressions, 3D mesh

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

This paper introduces ExDiff, a diffusion prior over the expressive parts only (both hands and the face), using SMPL-X as a decoding and fitting scaffold. The method combines part-aware noise schedules with a part-balanced score loss, a zero-initialized cross-part fusion head trained on about 160k paired hand-face frames, and geometry-aware supervision through SMPL-X decoding. One model is evaluated on generation, recovery, and inverse kinematics across six datasets, and it beats DPoser-X and other recent baselines on most metrics, for example ARCTIC recovery 5.12 vs 6.53 mm and FreiHAND JPE 4.6 vs 5.0 mm. Ablations cover the noise schedules, the fusion head, and the training strategy.

Critical strengths
  1. The paper targets a real gap that existing priors model the full body or treat hands and face separately, and EXDIFF is, to my knowledge, the first diffusion prior focused exclusively on both hands and the face. A single lightweight model supports generation, missing-part completion, and monocular recovery.

  2. The part-balanced formulation with region-adaptive noise schedules is a sensible way to train hand and face parts in one diffusion space from heterogeneous hand-only, face-only, and paired data, and the zero-initialized residual head is a clean design for adding cross-part interaction without disturbing the pretrained part priors.

Critical weaknesses
  1. The paper claims EXDIFF "consistently improves realism, coverage, and diversity on ARCTIC, FreiHAND, NOW, and BEAT2", but BEAT2 only appears in the Fig. 6 ablation with recovery errors, with no generation metrics and no baseline comparison, and in Table 1 NRDF is actually higher on APD (6.74 vs 6.52) and recall (0.94 vs 0.88). The authors should either add the missing BEAT2 generation results against baselines, or scope the claim to what the tables actually show.

  2. The central hand-face dependency claim has only partial support: Table 8's joint-precision recovery is consistent with learned dependencies, but the dependencies named in Sec. 3.6 are never directly measured. The authors may need to add a matched-capacity control without cross-part attention and directly measure the named dependencies for solidity of the experiment supports.

Questions for authors

Please explain the raised weakness.

4 comments

Weakness 1

We agree on both points. We add BEAT2 baselines and scope the claim in the abstract.

A. BEAT2. Correct. BEAT2 was used only for recovery, and only among our own variants. Table R7 adds DPoser-X with existing checkpoints, both for recovery in the setting of Fig. 6 and for joint generation on BEAT2 test under the Table 1 protocol. The ranking matches ARCTIC. ExDiff-F lowers the recovery error from 9.71 to 6.52 mm (face from 7.31 to 4.52) and is best on every generation metric, with FID falling from 0.14 to 0.05, precision rising from 0.67 to 0.82 and recall from 0.75 to 0.87. DPoser-X†, trained on our data, is behind ExDiff-B on FID, recall and dNN and level with it on precision (0.72).

Table R7. BEAT2 test with baselines. The recovery columns report PA-MPVPE (mm) with SMPLer-X initialisation, as in Fig. 6. The generation columns follow the Table 1 protocol, 5k generated against 5k BEAT2 test samples with 3 nearest neighbours.

ModelRecovery All ↓Recovery Hands ↓Recovery Face ↓APD ↑FID ↓Prec. ↑Rec. ↑dNN ↓
DPoser-X (released)9.7110.327.315.360.140.670.751.32
DPoser-X† (same data)–––5.720.110.720.781.25
ExDiff-B9.2410.156.035.910.100.720.821.18
ExDiff-F6.527.034.526.520.050.820.870.96

B. NRDF is higher on APD and recall. Correct, and our own text said "second-highest" while the abstract said "consistently". NRDF's high APD and recall come with a precision of 0.05 and an FID of 0.78, so its samples are spread widely and mostly lie off the real manifold. Separately, the revision removes the NRDF and VPoser rows from Table 1, because neither has a verified whole-body hand–face checkpoint (Reviewer 1, Weakness 2).

C. The scoped claim. Against DPoser-X, ExDiff-F is best on every generation metric except hand FID (0.02 against 0.01) and has the lowest recovery error on ARCTIC, FreiHAND, NoW and BEAT2. The abstract now reads: "Against DPoser-X, ExDiff improves generation realism and coverage on joint, hand and face benchmarks and lowers recovery error on ARCTIC, FreiHAND, NoW and BEAT2." The words "consistently" and "diversity" are removed, Sec. 4.1 states the one exception (hand FID), and Table R7 goes into the revision.

Weakness 2

We thank the reviewer for this suggestion. We have added a matched-capacity control and directly evaluated the hand–hand and hand–face dependencies described in Section 3.6.

A. Matched-capacity control without cross-part attention. F-indep uses the same 2.1M-parameter fusion head and 160k paired training frames as ExDiff-F, but replaces cross-part attention with block-diagonal attention (Table R4, in our response to Reviewer 2). Compared with F-indep, ExDiff-F reduces recovery error from 5.52 to 5.12 mm on ARCTIC and from 7.50 to 6.52 mm on BEAT2, and improves joint generation precision from 0.80 to 0.83.

B. Direct measurement of cross-part dependencies. We evaluate the dependencies described in Section 3.6 by conditional completion on the ARCTIC and BEAT2 test sets, with 10 samples per frame: right-hand completion given the left hand, hand completion given the face, and jaw completion given the hands. We also measure the correlation between jaw opening and finger flexion. Table R8 gives the results together with the independent and shuffled-pair controls.

Both dependencies show up against the matched control. Mirrored fingers are strong: the right-hand error given the left hand drops from 32.2 to 27.1 mm on ARCTIC and from 37.8 to 31.2 mm on BEAT2, and F-shuf, which keeps true left–right pairs, obtains most of this gain. Jaw–finger co-articulation is weaker and appears mainly on BEAT2, where faces carry expression: the hand error given the face drops from 37.2 to 34.5 mm, the jaw error given the hands from 8.5 to 7.1 degrees, and the jaw–finger correlation is 0.09, against 0.11 in the real data. On ARCTIC, where the face is jaw motion only, the face–hand gain is small.

Table R8. Direct evaluation of cross-part dependencies. Conditional completion on ARCTIC and BEAT2 test, mean over 10 samples per frame. R|L and H|F: hand PA-MPJPE (mm) given the ground-truth partner part. F|H: jaw-angle error (degrees) given the ground-truth hands. ρ: correlation between jaw opening and mean finger flexion over 5,000 generated joint samples; the real value is from BEAT2 test. F-shuf randomizes face–hand pairing while keeping true left–right hand pairs.

ModelR|L ARCTIC ↓R|L BEAT2 ↓H|F ARCTIC ↓H|F BEAT2 ↓F|H ARCTIC ↓F|H BEAT2 ↓ρ (jaw, fingers)
ExDiff-B34.439.433.839.33.38.70.00
F-indep32.237.832.437.23.38.50.00
F-shuf28.732.532.136.13.38.30.01
ExDiff-F27.131.231.334.53.17.10.09
Real BEAT2 test data––––––0.11

C. Manuscript revisions. We have revised Section 3.6 to clarify the supported cross-part dependencies, added the completion protocol and Table R8 to Appendix A.1, and updated Section 4.2 to discuss the direct measurements and their limitations.

We would like to thank you again for your careful review and constructive suggestions. We hope our responses have addressed your concerns regarding the paper. Please let us know if you have any further questions or if any additional information is required.

Thanks for pointing out both issues. The BEAT2 baselines, the scoped abstract, and especially the matched-capacity control with the direct completion evaluation address the majority of my concerns, and I am keeping my weak acceptance.

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

The paper proposes to train a human motion diffusion model, but explicitly focused only to represent hand and face motion in wrist- and neck-centric space. The authors choose this part-based representation instead of full-body diffusion, in order to reduce the computational complexity.

The training is done in two stages: first the hand and face parts are trained separately on domain specific datasets, then a zero-initialized fusion mechanism is added to allow training on full body datasets. The authors take special care of normalizing the losses to balance out the gradient signal, which is important due to the non-uniform space in which diffusion takes place.

The paper compares against several state-of-the-art methods in generation and mesh reconstruction. For the latter, the paper uses the learned diffusion model as a prior in an optimization loop which is in itnitialized with a feed-forward predictor.

Critical strengths
  • The paper shows impressive numbers
  • The authors achieved stable training across both domains jointly, and the authors ablated that a shared noise schdule performs worse.
Critical weaknesses
  • The experimental validation of the main claims are flawed. In particular:
    • It is not clear whether joint training is better bc. the joint stage uses additional data, whether the extra model capacity helps, or whether the two domains actually benefit from each other. The latter is of course the central premise of the paper. Ultimately, the reader cannot know whether the full model outperforms the individual parts bc. of joint benefits, or bc. of the additional training data of the joint stage. This needs to be properly ablated by using the same data. E.g. Table 6 would suggest that full model achieves better face shape reconstruction. But since the hands are seldom visible in the NoW Benchmark and are only weakly correlated with the detailed head shape, this would point to improvements due to more data.
    • Similarly, it is not properly ablated, why the step of removing the other body parts is required. By simply finding a proper noise schdule for those parameters, the model could be extended to understand the whole body, and the data that is used would support such an approach. To me, the decision to model floating hands and the face without the relation of the hands to each other nor the face seems very arbitrary and is also far away from the actual applications, where the whole body is usaully of interest. Simply resorting to a model capacity argument here seems a bit slim.
  • The paper does not show any qualitative comparisons against baselines. The result visualizations do not show the input image, makeing it hard to judge whether the actual expression/hand pose are accurately captured. The face alignmed seems poor.
Questions for authors
  • The "random" noise schdule ablation uses the same noise range as for the hands. On the ARCTIC benchmarks ExDiff-B (part-aware) achieves 6.01, while the random variant achieves 9.57. I am wondering where this difference is coming from, since the "random" scheudle is the "hand" schedule and ExDiff-B should treat the parts separate from each other, if I am not mistaken. What am I missing here?
  • What was your strategy to choose the noise schdule?
7 comments

Weakness 1

We agree, and your reading of Table 6 is right. With matched data, about 60% of the gain from ExDiff-B to ExDiff-F is data and capacity and about 40% is the hand–face link. On NoW all of it is data.

A. Same-data ablation. Two controls keep ExDiff-F's head, data and schedule and change one thing. F-indep uses block-diagonal attention, so it has no link. F-shuf pairs the face of one frame with the hands of a random other frame, so it has no true pairing. They recover 1.6 to 1.7 mm of the 2.7 mm gain on BEAT2 and 0.6 of the 1.0 mm gain on ARCTIC. The remaining 1.0 to 1.1 mm and 0.4 to 0.5 mm need the true pairs (Table R4).

Table R4. Same-data controls (same 2.1M head and 160k paired frames). Left, PA-MPVPE (mm, All) with SMPLer-X initialisation. Right, 5k generated against 5k ARCTIC test samples, 3 nearest neighbours. Gap is the mean marginal precision minus the joint precision.

VariantCross-part attentionTrue pairingARCTIC All ↓BEAT2 All ↓Joint Prec. ↑Hand-marg. Prec. ↑Face-marg. Prec. ↑Gap ↓
ExDiff-B✗–6.159.240.750.820.840.080
F-indep✗ (block-diag.)✓5.527.500.800.850.860.055
F-shuf✓✗ (shuffled)5.587.650.790.850.860.065
ExDiff-F✓✓5.126.520.830.870.880.045

B. NoW points to more data. Hands are absent on NoW, and the fusion head does not see shape. The NoW gain therefore comes from extra face data, not from the link. F-indep, with the same data and no link, matches ExDiff-F (Table R5). Table 6 gets the F-indep row.

Table R5. NoW (Table 6) with F-indep. Median, mean and standard deviation over all views, in the unit of Table 6.

MethodFrom scratchWith MICA
ExDiff-B11.2 / 14.0 / 11.88.2 / 10.1 / 8.8
F-indep (same data, no link)9.8 / 13.0 / 10.56.4 / 8.6 / 7.6
ExDiff-F10.1 / 12.9 / 10.66.7 / 8.7 / 7.7

C. The revised claim. The hand–face benefit is claimed only where both parts are observed, on ARCTIC and BEAT2. On NoW and FreiHAND the gain is described as in-domain refinement. Sec. 4.2, the related work and the captions are revised accordingly.


Edited on 19 Nov 2026 to correct the DPoser-X description and numbers, as explained in our latest reply.

Weakness 2

We agree that the choice was argued but not tested. We now test it and drop the capacity argument.

A. Ablation. DPoser-X is the whole-body alternative you describe. It uses one sub-VP schedule for all parts, is built from frozen part priors with a fusion module, and is trained on ARCTIC, BEAT2, GRAB, EgoBody and part-only sets. Yet it is behind ExDiff-B on hands and face (6.51 against 6.15 mm on ARCTIC, and joint precision 0.70 against 0.75). Because this comparison mixes scope, data and architecture, we add ExDiff-F with body. It is our recipe with a fourth, frozen body token with its own noise schedule, and its head is retrained on the same paired frames (Table R6). On ARCTIC the error rises from 5.12 to 5.20 mm, on BEAT2 from 6.52 to 6.65 mm, and joint precision falls from 0.83 to 0.82. Removing the body is therefore not required. Our recipe extends to the whole body at a cost of about 0.1 mm and stays ahead of DPoser-X (6.51, trained on different data) and of the same-data controls (5.52 and 5.58).

Table R6. With and without a body token, same recipe and paired data. PA-MPVPE (mm) with SMPLer-X initialisation, and joint precision on the hand and face dimensions as in Table 1. The DPoser-X row is our own run, as in Table R9.

ModelPartsARCTIC All ↓Hands ↓Face ↓BEAT2 All ↓Joint Prec. ↑
DPoser-X (released)body, hands, face6.517.302.769.710.70
ExDiff-Fhands, face5.126.011.476.520.83
ExDiff-F with bodybody, hands, face5.206.101.516.650.82

B. Floating hands and face. The parts are related. The fusion head models both hands and the face jointly, which is worth 1.0 to 1.1 mm on BEAT2 and 0.4 to 0.5 mm on ARCTIC in single runs (Table R4). What the hands-and-face model leaves out is where the hands are relative to the head, and ExDiff-F with body covers this through the body token. The revision adds it as the whole-body variant in Appendix A.1.

C. The capacity argument. We agree that it is weak, and we drop it. Data is not a reason either, because DPoser-X also trains frozen part priors on part-only data and joins them with a fusion module. The hands-and-face scope is our choice of focus, not a requirement, and Table R6 shows that the body can be added at a small cost. The revision says this in Sec. 1 and the related work.

Weakness 3

All three points are fair, and the revision changes the figures.

A. Qualitative comparison with baselines. Figs. 3 to 5 compare generated samples with baselines, but we did not show a recovery comparison. The closest baseline is DPoser-X, which models hands and face only as part of a whole-body prior, and its paper shows no results for hands and face together. For a fair comparison, we modify the released DPoser-X code so that it fits only the hands, jaw and expression, with the body fixed. The revision adds a recovery figure with the input image, the fit with this DPoser-X, and the fit with ExDiff-F. Both fits use the same inputs and the same first-stage body fit, so only the prior is different.

B. Input images. In the revision, the new comparison figure shows the input image, and so does every qualitative recovery result in the appendix.

C. Face alignment. The in-the-wild fits start from the mean SMPL-X shape with a camera set from the face width, so the outline often misses the face, and the opaque mesh makes this look worse. The revision initialises the shape from MICA, uses a semi-transparent overlay and gives the landmark reprojection error under each example. The measured face error is in Tables 2 and 6.


Weakness 4

Q1. Why do the hands change? You are not missing anything, and our description was incomplete. The matched comparison is 9.57 against 7.18, while 6.01 is ExDiff-F. The hand range and the hand prior are the same in both rows, and on FreiHAND alone the prior gives the same 4.8 mm. The difference comes from fitting. With one shared range the face is perturbed at hand-scale noise, and the face error rises from 2.06 to 4.85. Since hands and face share the camera and global orientation, the face misfit raises the hand error too. We are fitting the hands alone under both schedules to confirm this and will add the result to Appendix A.1. Table 7 renames the "random" variant to "single shared schedule".

Q2. How was the schedule chosen? We use the variance-exploding schedule of Song et al. with one range per part. The hands use its default range, from 0.01 to 50. The face range, from 0.005 to 20, was set by hand and is narrower because the jaw and expression code has a smaller spread. Neither range was grid-searched, and Table 7 is our only test of the choice. Sec. 3 now says so.

We would like to thank you again for your careful review and constructive suggestions. We hope our responses have addressed your concerns regarding the paper. Please let us know if you have any further questions or if any additional information is required.

Reviewer_a9EDOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a diffusion prior over bilateral hand articulation, jaw pose, and facial expression. It combines part-specific noise schedules, dimension-normalized score matching, cross-part residual fusion, and geometry-aware supervision. Independent part priors are pretrained and then frozen while a fusion module is trained on paired hand-face data. The resulting prior supports generation, missing-part completion, and monocular recovery.

Critical strengths
  1. The focus on expressive regions is well motivated, and the distinction between local articulation and placement inherited from the body scaffold is clearly acknowledged.
  2. The experiments cover several generation and recovery settings and report promising improvements. For example, with SMPLer-X initialization, Table 2 reduces overall PA-MPVPE from 6.53 mm with DPoser-X to 5.12 mm with the proposed fused prior.
  3. Figure 6 provides a useful factorial ablation of fusion and geometry supervision. I also appreciate the shared fitting procedure and iteration budget described in Appendix A.2.
Critical weaknesses
  1. The benefit of cross-part dependency modeling is not sufficiently isolated. The fused model introduces a 2.1M-parameter module trained on approximately 160k paired frames, while the unfused model uses frozen independent priors (Appendix A.1, Table 9). Thus, improvements may reflect additional data and capacity rather than learned hand-face dependencies. Controls with matched independent adapters or shuffled hand-face pairings would help isolate this contribution. Moreover, the precision values in Table 8 are computed in different spaces and are not directly comparable.
  2. The generation baselines are insufficiently specified. Appendix A.6 describes adapting whole-body models to expressive regions, but the official VPoser implementation models 21 body joints, not hand articulation or facial expression. The authors should clarify the checkpoints, retraining, and parameter mappings used for Tables 1, 3, and 5. A diffusion baseline trained on the same data and representation would also help establish whether the gains come from the proposed architecture.
  3. The face evaluation raises questions about annotation provenance and numerical consistency. Appendix A.1 uses ARCTIC expression coefficients, but the official documentation and processing code do not explicitly provide them. Their source and the reference face geometry need clarification. Additionally, Table 6 reports MICA errors of 9.0/11.1/9.2 mm, compared with 0.90/1.11/0.92 mm on the official NoW leaderboard. This factor-of-ten discrepancy requires explanation.
  4. The supplementary video reveals hand misplacement and apparent mesh penetration In real-test2.mp4 (around 2.67 s), the reconstructed hand jumps above the subject's head while the input hand remains near the torso. The wrist/sleeve region around 2.17 s also shows apparent mesh penetration or clipping artifacts. The authors should clarify whether these issues arise from body/camera fitting, expressive optimization, or rendering, and report their frequency. Comparisons with the initialization and matched baseline would help assess the robustness of the complete recovery pipeline, without attributing these failures solely to the proposed prior.
Questions for authors
  1. Does the fusion benefit persist against controls matched for data, capacity, and training budget, and what happens when hand-face pairings are shuffled?
  2. What are the exact baseline implementations and ARCTIC expression-label sources, and how should the NoW numerical discrepancy be reconciled?
  3. Which stage causes the hand misplacement at 2.67 s, and how do the initialization and matched baseline behave on that frame?
5 comments

Weakness 1

We agree with all three points. Table R1 adds both controls and replaces Table 8. It enters the revision as Table 9, and Sec. 4.2 is rewritten to match.

Table R1. Controls (left) and corrected generation metrics (right). F-indep, F-shuf and ExDiff-F share the same 2.1M head and 160k paired frames. The left part reports PA-MPVPE (mm, All) with SMPLer-X initialisation. The right part compares 5k generated with 5k ARCTIC test samples using 3 nearest neighbours. Gap is the mean marginal precision minus the joint precision.

VariantCross-part attentionTrue pairingARCTIC All ↓BEAT2 All ↓Joint Prec. ↑Hand-marg. Prec. ↑Face-marg. Prec. ↑Gap ↓
ExDiff-B✗–6.159.240.750.820.840.080
F-indep✗ (block-diag.)✓5.527.500.800.850.860.055
F-shuf✓✗ (shuffled)5.587.650.790.850.860.065
ExDiff-F✓✓5.126.520.830.870.880.045

A. Data and capacity. We agree that part of the gain comes from data and capacity. Going from ExDiff-B to ExDiff-F adds 2.1M parameters, 160k in-domain ARCTIC and BEAT2 frames, and the hand–face link at once. The controls separate them, and about 60% of the gain is data and capacity.

B. Matched controls. Both controls keep the head, data, schedule and geometry-loss weight of ExDiff-F and change one thing. F-indep uses block-diagonal attention, so it has no link. F-shuf pairs the face of one frame with the hands of a random other frame, so it has no true pairing. They recover 1.6 to 1.7 mm of the 2.7 mm gain on BEAT2 and 0.6 of the 1.0 mm gain on ARCTIC. The remaining 1.0 to 1.1 mm and 0.4 to 0.5 mm need the true pairs. Joint precision agrees, rising from 0.75 for ExDiff-B to 0.79 and 0.80 for the controls and 0.83 for ExDiff-F.

C. The precision values in Table 8. The point is correct. Its joint rows used ARCTIC in the full hand and face space, while its part rows came from Tables 3 and 5, which use FreiHAND and WCPA in smaller spaces. The comparison of 0.83 and 0.846 against 0.75 was therefore not fair, and it is removed. The right half of Table R1 replaces Table 8. It uses one reference set and the same samples, reports joint and marginal precision, and reads co-occurrence within one row as the Gap. The controls shrink the Gap from 0.080 to 0.055 and 0.065, and ExDiff-F reaches 0.045.


Weakness 2

We agree. Table R2 gives the source of every baseline row and replaces the incorrect sentence in A.6.

A. VPoser models body joints only. The reviewer's point is correct, and our A.6 sentence was incorrect. The VPoser and NRDF rows of Tables 3 and 5 are the hand and face versions retrained by the DPoser-X authors, and we report their published numbers (DPoser-X Tables S-8 and S-11). Neither has a verified whole-body hand–face checkpoint, so both rows are removed from Table 1.

Table R2. Source of every baseline row in Tables 1, 3 and 5.

RowTable 1 (joint)Table 3 (hand)Table 5 (face)
VPoserRemoved (no verified whole-body checkpoint)DPoser-X Table S-8 (VPoser-hand)DPoser-X Table S-11 (VPoser-face)
NRDFRemoved (no whole-body model, needs rotations)DPoser-X Table S-8 (NRDF-hand)Not available (needs rotations)
DPoser-XReleased whole-body checkpoint, sampled by us, with hands, jaw and expression keptDPoser-X Table S-8 (DPoser-hand)DPoser-X Table S-11 (DPoser-face)
ExDiffOur protocol in A.6, 5k generated against 5k ARCTIC test samples, 3 nearest neighboursDPoser-X's released script, 50k samples against its hand reference batchSame script, 50k samples against its face reference batch

B. Checkpoints, retraining and mappings. Table R2 lists them. The baseline rows of Tables 3 and 5 are as published in DPoser-X, and our rows use the same released evaluation code in the same MANO and FLAME spaces.

Table R3. Same-data diffusion baseline for Table 1 (5k generated against 5k ARCTIC test samples, 3 nearest neighbours).

ModelTraining dataAPD ↑FID ↓Prec. ↑Rec. ↑dNN ↓
DPoser-X (released)authors' whole-body mix5.360.100.700.781.31
DPoser-X† (retrained)ours, part-only and paired5.730.090.730.801.25
ExDiff-Bours, part-only5.910.080.750.821.18
ExDiff-Fours, part-only and paired6.520.040.830.880.96

C. Same-data diffusion baseline. For hands and face it already exists. DPoser-hand and DPoser-face are diffusion priors trained on the same datasets and parameters as ExDiff-B, so Tables 3 and 5 already compare architectures. ExDiff-B is better on most metrics but not on hand FID or expression dNN, and the revision says so. For the joint setting we add DPoser-X†, the DPoser-X recipe trained on exactly ExDiff-F's data and representation (Table R3). With data matched, ExDiff-F still raises APD from 5.73 to 6.52, lowers FID from 0.09 to 0.04 and raises precision from 0.73 to 0.83, so the gain comes from the architecture.

Weakness 3

We agree on both points.

A. ARCTIC expression and reference face. ARCTIC provides jaw and eye poses but no expression coefficients, and our A.1 sentence was incorrect. For ARCTIC frames the expression is set to zero, so the expression–hand link is learned from BEAT2 only. The reference face on ARCTIC, used in the Face column of Table 2, is ARCTIC's own SMPL-X face with subject shape, jaw pose and zero expression. That column therefore measures jaw and shape, not expression, for every method. A.1 and the Table 2 caption now say this.

B. The factor of ten on NoW. The unit in Table 6 is incorrect, but the ranking is not. Every row comes from DPoser-X's pipeline, the baselines from its Table 5 and ours from its released code. Its unit is ten times the official NoW millimetre, so the MICA row is the official 0.90, 1.11 and 0.92 mm multiplied by ten. Since all rows share the unit, their order holds. We are re-running our rows and MICA with the official NoW script on the validation set and will post the millimetre values here if they finish. The revision reports Table 6 in official millimetres and calls the metric scan-to-mesh distance throughout.


Weakness 4

Both artifacts are real. One comes from the first-stage body fit and one from rendering, and neither comes from the prior. The demo fits every frame on its own, starting from the mean pose, with 500 iterations on ViTPose keypoints and no temporal term (Appendix A.2).

A. Hand jump at 2.67 s, caused by body and camera fitting. It is one frame, frame 16 of 36. The detector is right. It places the right wrist at chest height with a score of 0.99, and the mean hand confidence of 0.85 is far above the threshold of 0.3. The mean-pose initialisation has both arms down, so the error arises in the first-stage fit of camera and body, which stopped in a wrong local minimum for the right arm. The prior models only finger articulation in the wrist frame and cannot move the hand. When we fit the same frame with DPoser-X as the prior, it gives the same misplacement. The revision restarts the arm fit when the wrist reprojects more than 10% of the box height away from a confident keypoint, and this check catches frame 16.

B. Wrist clipping at 2.17 s, caused by rendering. We drop the forearm when rendering (Appendix A.4), so the hand mesh is open at the wrist, and its dark inside shows when the opening faces the camera. The other views show an intact hand and no interpenetration. The revision closes the wrist with a cap for rendering.

C. Frequency and comparisons. In real-test2.mp4 the hand jump occurs in 1 of 36 frames (2.8%) and the wrist opening in 13 of 36 frames (36%). The revised supplementary reports both counts for every clip. It also adds a comparison from the same keypoints and initialisation, without a prior, with DPoser-X and with ExDiff-F. The Limitations section now states that first-stage arm errors are outside what the prior can fix.

We would like to thank you again for your careful review and constructive suggestions. We hope our responses have addressed your concerns regarding the paper. Please let us know if you have any further questions or if any additional information is required.

Comments on the paper

1 comment

General Comment

We would like to thank all the reviewers for their careful review and constructive suggestions. We hope our responses have addressed your concerns regarding the paper. Please let us know if you have any further questions or if any additional information is required.

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

Related papers

Discussion (0)

Sign in to comment.