Human mesh recovery (HMR) from a single RGB image is inherently ambiguous, as multiple 3D poses can correspond to the same 2D observation. Recent diffusion-based methods tackle this by generating various hypotheses, but often sacrifice accuracy. They yield predictions that are either physically implausible or drift from the input image, especially under occlusion or in cluttered, in-the-wild scenes. To address this, we introduce a dual-memory augmented HMR critique agent with self-reflection to produce context-aware quality scores for predicted meshes. These scores distill fine-grained cues about 3D human motion structure, physical feasibility, and alignment with the input image. We use these scores to build a group-wise HMR preference dataset. Leveraging this dataset, we propose a group preference alignment framework for finetuning diffusion-based HMR models. This process injects the rich preference signals into the model, guiding it to generate more physically plausible and image-consistent human meshes. Extensive experiments demonstrate that our method achieves superior performance compared to state-of-the-art approaches.
Figures & tables
Figure 1: We introduce a VLM-guided HMR critique agent equipped with a dual-memory mechanism that delivers stable and semantically grounded assessments for groups of estimated 3D meshes. Building on these group-wise signals, our group preference alignment framework steers diffusion-based HMR models towards more coherent and reliable mesh generation.
Figure 2: Overview of our framework. Our purpose is to refine a diffusion-based HMR model that generates a group of human mesh predictions per input image. We propose a VLM-enhanced HMR critique agent that assigns a score for each human mesh prediction. This critique agent is equipped with a dual-memory mechanism to give stable assessments. Then, we use this critique agent to build a group-wise HMR preference dataset without the need for manual labeling. Finally, we employ this preference dataset to finetune the base model to preferentially generate predictions that are physically plausible and better aligned with the image cues.
Methods
M
3DPW [ 55 ]
Human3.6M [ 22 ]
PVE ↓
MPJPE ↓
PA-MPJPE ↓
PVE ↓
MPJPE ↓
PA-MPJPE ↓
Deterministic
HMR [ 24 ]
-
152.7
130.0
81.3
96.1
88.0
56.8
HybrIK [ 31 ]
-
86.5
74.1
45.0
65.7
54.4
34.5
PyMaf [ 67 ]
-
110.1
92.8
58.9
-
57.7
40.5
HMR 2.0 [ 18 ]
-
-
70.0
44.5
-
44.8
33.6
POTTER [ 68 ]
-
87.4
75.0
44.8
-
56.5
35.1
Table 1: Comparison with state-of-the-arts on the 3DPW [ 55 ] and Human3.6M [ 22 ] dataset. M is the number of predictions of probabilistic methods. Ours † is trained on an additional in-the-wild dataset InstaVariety [ 25 ] , using only preference signals without 3D labels.
Methods
3DPW [ 55 ]
PVE ↓
MPJPE ↓
PA-MPJPE ↓
Base Diffusion Model
73.4
63.0
37.6
+ supervised finetuning
70.2
61.3
36.5
Ours
59.5
49.9
31.9
DPO w/ critique agent
63.9
53.1
33.4
w/o critique agent
65.4
54.9
34.7
Table 2: Ablation study on the 3DPW [ 55 ] test set. All models are finetuned on the InstaVariety [ 25 ] dataset. M=100 for all.
Figure 3: Qualitative comparison between our method and the state-of-the-art probabilistic model ADHMR [ 50 ] . Examples (a) ∼ (e) are from the 3DPW [ 55 ] dataset, while (f) ∼ (h) are challenging internet images. Both overlay and side-view results are shown.
Figure 4: Visualization of our critique agent on an internet image. We compare the erroneous initial score from HMR-Scorer [ 50 ] with the corrected score from our critique agent, along with the corresponding critique.
Methods
GTA-Human II [ 7 ]
DNA-Rendering [ 9 ]
PVE
MPJPE
PA-MPJPE
PVE
MPJPE
PA-MPJPE
SRCC ↑
KRCC ↑
SRCC ↑
KRCC ↑
SRCC ↑
KRCC ↑
SRCC ↑
KRCC ↑
SRCC ↑
KRCC ↑
SRCC ↑
KRCC ↑
ScoreNet [ 61 ]
0.464
0.411
0.468
0.404
0.430
0.327
0.532
0.512
0.496
0.469
0.458
0.367
HMR-Scorer [ 50 ]
0.578
0.506
0.588
0.513
0.489
0.428
0.610
0.540
0.562
0.508
0.475
0.393
Ours
0.605
0.539
0.615
0.556
0.528
0.461
0.633
0.556
0.588
0.532
0.510
0.433
w/o rule memory
0.597
0.491
0.577
0.480
0.454
0.427
0.619
0.487
0.564
0.520
0.484
0.377
Table 3: Group-wise score prediction results. We report the SRCC and KRCC between the predicted scores and the ground-truth HMR metrics. Our method consistently outperforms all baselines and ablation variants.
Methods
GTA-Human II [ 7 ] - PLCC
DNA-Rendering [ 9 ] - PLCC
PVE
MPJPE
PA-MPJPE
PVE
MPJPE
PA-MPJPE
ScoreNet [ 61 ]
0.524
0.516
0.472
0.550
0.551
0.499
HMR-Scorer [ 50 ]
0.634
0.627
0.565
0.664
0.659
0.620
Ours
0.695
0.697
0.653
0.737
0.717
0.700
w/o rule memory
0.683
0.672
0.630
0.701
0.673
0.650
w/o prototype memory
0.645
0.629
0.597
0.677
0.659
0.629
Table 4: Point-wise score prediction results. We report the PLCC between the predicted scores and the ground-truth metrics.