acceptodds
Under review as a conference paper at ICLR 2027

UniBioTransfer: A Unified Framework for Multiple Biometrics Transfer

Abstract

DeepFace generation has traditionally followed a task-driven paradigm. However, such a single-task setting severely limits model generalization and scalability. A unified model capable of solving multiple deepface generation tasks in a single pass represents a promising and practical future direction, but remains extremely challenging due to data scarcity and serious cross-task conflicts arising from heterogeneous attributes transformations. To this end, we propose UniBioTransfer, the first unified framework capable of handling both conventional deepface tasks, e.g., face transfer and face reenactment, and tasks with huge shape-varying transformations, e.g., hair transfer and head transfer. Far more beyond, UniBioTransfer naturally generalizes to unseen tasks, like lip, eye, and glasses transfer, with minimal model fine-tuning. In simple words, UniBioTransfer achieves the goal by addressing data insufficiency challenge in multi-task generation through a unified data construction strategy that can handle diverse and distinct attributes, and mitigates cross-task interference via an innovative BioMoE, a Mixture-of-Experts based model coupled with a novel two-stage training strategy that effectively disentangles task-specific knowledge. Extensive experiments demonstrate the effectiveness, and generalization, and scalability of UniBioTransfer, outperforming both existing multi-task models and task-specific methods across a wide range of tasks, while maintaining a comparable model size and training time. We will release our code upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
head swap, portrait edit, hair swap

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_DEfbOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  • A single SD-1.5-based model for face, hair, motion, and head transfer, with adaptation to low-level tasks (eye, lip, glasses, hat).

  • Data: swapping-based corruption, where spatially-dynamic attributes are corrupted with an external generative model.

  • Model: BioMoE, consisting of a global expert, 8 routed experts selected by a structure-aware router, and task-specific LoRA experts.

  • Training: Task-Isolated Initialization per task, then averaging the FFNs and fitting the residuals via SVD before Task-Unified Fine-tuning.

  • Evaluated on 987 FFHQ pairs against task-specific, multi-task, and unified baselines, with per-component ablations.

Critical strengths
  • The swapping-based corruption strategy is the most convincing part of the paper. The problem statement is clear (a mask alone leaks the target silhouette), and Table 5 (hair 0.492→0.421, head 0.555→0.460) and Table 6 (0.424 vs 0.421 when the external generator is swapped) support it.

  • Handling four tasks from one checkpoint and adapting to low-level tasks (Table 3) has practical value. The gap in Table 4 over naive parameter sharing (0.478) and one-stage training (0.507) versus the full model (0.637) is substantial.

Critical weaknesses
  • The claim that the unified model outperforms task-specific models is not well supported. In Table 4 the difference between the task-specific row and the full model is small (e.g., 0.622 vs 0.637), and the effect of structure-aware routing is at the level of 0.634→0.637, with no variance reported. “Comparable performance from a single model” would be a sufficient claim; the paper states more than that.

  • The comparison conditions are mixed. Some competing methods perform better on individual metrics (CanonSwap on face-transfer pose distance, 2.92 vs 3.63; Face-Adapter on motion-transfer pose distance, 6.89 vs 7.09), REFace for head transfer was retrained by the authors (A.5), and the hair/head training data is constructed with StableHair, which is then used as a baseline (B.1.2).

Questions for authors

Q1) A.5 acknowledges that the original REFace paper supports both face and head transfer, but Table 1 classifies it under “training separately for different tasks.” Does this distinction concern the method’s capability or the publicly available implementation, and did you try the released REFace checkpoint on head transfer?

Q2) What happens with only the global expert and task-specific LoRA experts, without routed experts? This combination is not in Table 4. Could you also report variance over repeated runs for Table 4?

Q3) In Stage I, how are the components other than the FFN (attention, refNet, task-specific projection MLPs) trained and carried into Stage II? The text only describes averaging the FFNs. Please also check the reference to “Appendix 3.3” in Sec. 3.3.

1 comment

We sincerely thank the reviewer for the careful reading and constructive suggestions. We have revised the manuscript accordingly and respond below.

W1. Comparison with Task-Specific Methods

CanonSwap’s lower pose distance should be considered together with its identity-transfer fidelity. As illustrated in Figures 4(b) and 14, CanonSwap exhibits a more conservative identity transfer, with its outputs retaining more characteristics of the target face. Retaining the original face can yield a low pose distance even when the intended identity transfer is incomplete. This is consistent with its lower ID similarity in Table 2 (0.551 versus our 0.637). The visual comparisons and user study further support our overall face-transfer quality: Figure 10(a) reports a preference rate of 64.6% for Ours versus 19.6% for CanonSwap.

Face-Adapter achieves a slightly lower motion-transfer pose distance (6.89 versus 7.09), while Ours achieves higher identity similarity (0.602 versus 0.503) and a lower expression distance (2.09 versus 2.27). These results support a better balance of identity preservation and expression transfer while maintaining competitive pose accuracy.

We have refined our claims to emphasize the overall benefits of task consolidation while acknowledging the trade-offs on individual metrics.

W2. Comparison Conditions

(a). REFace for head transfer

The original REFace method supports face and head transfer through separate retraining, resulting in two task-specific models. As no head-transfer checkpoint is publicly released, we retrain it by expanding the masking region to include hair and accessories, using its official hyperparameters.

(b). StableHair for data construction

StableHair is used to corrupt the training inputs, while the supervision target remains the original real image. Our model therefore learns to reconstruct real images rather than imitate StableHair’s outputs. Furthermore, our models trained using corruptions generated by StableHair and HairCLIPv2 achieve comparable performance, supporting low dependence on the choice of external corruption generator (Table 6).

Q2. BioMoE Ablation and Multi-seed Results

We repeat the model-design ablations in Table 4 and add two variants: w/o Task-specific Experts and w/o Routed Experts. Results are aggregated over 12 training–inference seed combinations (see Appendix B.3 for details).

Each cell reports mean ± STD, followed by the gain of Ours over that variant with an unpaired Welch 95% CI. Positive gains favor Ours.

SettingFace ID ↑Hair ↓Motion ↓Head ↓
w/o BioMoE (Naive Parameter Sharing)0.480 ±\pm 0.008; +0.1535 [0.1471, 0.1599]0.477 ±\pm 0.005; +0.0570 [0.0535, 0.0604]9.80 ±\pm 0.18; +2.81 [2.69, 2.93]0.576 ±\pm 0.005; +0.1212 [0.1169, 0.1255]
w/o BioMoE (Task-Specific Model)0.623 ±\pm 0.016; +0.0103 [−0.0006, 0.0213]0.428 ±\pm 0.003; +0.0080 [0.0062, 0.0098]7.10 ±\pm 0.09; +0.11 [0.04, 0.18]0.461 ±\pm 0.004; +0.0065 [0.0029, 0.0102]
w/o Structure-Aware Routing0.630 ±\pm 0.003; +0.0036 [−0.0009, 0.0081]0.425 ±\pm 0.002; +0.0049 [0.0032, 0.0066]7.04 ±\pm 0.10; +0.05 [−0.02, 0.13]0.470 ±\pm 0.008; +0.0150 [0.0096, 0.0204]
w/o Task-specific Experts0.580 ±\pm 0.004; +0.0531 [0.0485, 0.0578]0.432 ±\pm 0.002; +0.0118 [0.0101, 0.0135]7.82 ±\pm 0.06; +0.83 [0.77, 0.88]0.491 ±\pm 0.006; +0.0357 [0.0312, 0.0402]
w/o Routed Experts0.629 ±\pm 0.003; +0.0047 [0.0001, 0.0093]0.426 ±\pm 0.003; +0.0060 [0.0037, 0.0083]7.15 ±\pm 0.17; +0.16 [0.04, 0.28]0.466 ±\pm 0.007; +0.0114 [0.0062, 0.0166]
Ours0.633 ±\pm 0.0070.420 ±\pm 0.0016.99 ±\pm 0.080.455 ±\pm 0.005

Overall, the full model achieves the best mean scores across all four tasks, with predominantly positive gain CIs supporting the effectiveness of each design choice.

Q3. Training and Initialization Across Stages

The refNet is a reference network that follows the same parameter-handling strategy as the main denoising network. In both networks, FFNs undergo the averaging and residual-based initialization described in Section 3.3. Attention projections and the projection MLPs following the CLIP encoder remain task-specific and retain their Stage-I weights in Stage II, introducing only a small parameter overhead because these projections are much smaller than FFNs (see Appendix A.1).

We have clarified these details and corrected “Appendix 3.3” to “Section 3.3” in the revised paper. Thanks for your valuable reviews!

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

UniBioTransfer unifies face, hair, motion, and head transfer in a diffusion-based framework, with further fine-tuning for facial parts, accessories, and attribute combinations. The paper addresses two problems: training masks can leak the desired geometry, and shared training can cause interference between tasks. Its data construction strategy corrupts selected attributes while retaining the original real image as the supervision target; for shape-varying attributes, it uses generated swaps instead of simple masks. Its BioMoE architecture combines a shared expert, lightweight routed experts guided by facial landmarks, and task-specific LoRA experts. Training first learns tasks separately, then initializes the shared model by weight averaging and allocates LoRA ranks using singular values of task residuals before joint training.

Evaluation includes 987 FFHQ image pairs, comparisons with specialized and unified methods, facial-part adaptation, component ablations, and additional tests involving extreme poses, exaggerated expressions, and occlusions. A 16-participant preference study supplements the automatic metrics. Results generally favor UniBioTransfer across the four main tasks, although it does not win every metric. The main contribution is a practical combination of training-data construction and parameter sharing that broadens task coverage while retaining competitive transfer quality.

Critical strengths
  1. A well-motivated data construction contribution with direct empirical support. Swapping-based corruption addresses a concrete shortcut in self-reconstruction training: the input mask can reveal the desired silhouette. Importantly, the supervision target remains the original real image rather than the external generator’s output. Table 5 shows improved hair/head CLIP distances over masking alone (0.421/0.460 versus 0.492/0.555), and Table 6 shows similar downstream results with two different corruption generators. This makes the data contribution more convincing than a claim based solely on using a stronger external model.

  2. Useful task consolidation with broad supporting experiments. One framework handles four tasks, including substantial hair/head shape changes, while remaining competitive with specialized models. Table 4 shows a considerable improvement over naive sharing, including face identity similarity of 0.637 versus 0.478 and motion pose distance of 7.09 versus 9.74. The difficult-condition evaluations, facial-part adaptation results, and preference study provide complementary support. Maintaining this breadth without a large quality penalty is a meaningful practical contribution even where the margins over task-specific models are modest.

Critical weaknesses
  1. The evidence is stronger for the overall system than for the necessity of every BioMoE component. Table 4 establishes that naive sharing performs poorly, but does not separately remove routed experts and task-specific experts to isolate their respective contributions. Structure-aware routing yields small improvements on several metrics, such as face identity similarity increasing from 0.634 to 0.637. Table 10 appropriately matches total parameters for rank-allocation comparisons and supports a substantial motion-transfer benefit over uniform ranks, but shows nearly unchanged performance on other tasks, including slightly better hair distance for uniform ranks. Without repeated runs or uncertainty estimates, the evidence does not fully establish the robustness of the smaller gains or the necessity of the complete architecture. This limits the strength of the methodological claims beyond the demonstrated system-level improvement.

  2. The generalization and scalability claims extend beyond the quantitative evidence. Table 3 supports adaptation to facial-part tasks, but accessory transfer and cross-/intra-level compositions are supported mainly by examples. The paper explains expert inheritance, adding a router, and constructing composite references, but it does not provide a comparably complete quantitative protocol for these combinations, including adaptation data, training budget, and preservation of previously learned capabilities. The user study helps with the four main tasks but includes only five questions per task, so it cannot support the broader compositional claims. This matters because efficient expansion to new tasks is a prominent motivation for the unified framework.

Questions for authors
  1. In Table 4, are the naive-sharing, single-task, and one-stage variants matched for training data and total optimization budget? Do you have ablations separating the routed-expert and task-specific-expert contributions, or repeated-run results that establish the stability of the smaller routing gains?

  2. For accessory and compositional adaptation, which parameters are updated, how many training examples and steps are used, and is the stated 3 A40 GPU-days adaptation cost per new task? Are the original tasks retained in the same checkpoint after adaptation, and have you measured their scores? Please distinguish demonstrated results from anticipated scalability.

  3. Appendices A.5 and C.3 exclude preprocessing failures separately for affected methods. How many failures occur for each method, and do the conclusions hold on a common successful subset? For the user study, how were the five examples per task selected, how was the output order randomized, and how were ties aggregated into the reported preference rates?

2 comments
AuthorsOpenReview ↗

This is Part 1/2 of our reply

We sincerely thank the reviewer for the detailed review.

We have revised the manuscript accordingly and respond below. Due to word count limitations, we will answer all your questions in two parts. This is the first part.

W1. BioMoE Component Contributions

(a). Component ablations and multi-seed results

We repeat the model-design ablations in Table 4 and add two variants: w/o Task-specific Experts and w/o Routed Experts. Results are aggregated over 12 training–inference seed combinations (see Appendix B.3 for details).

Each cell reports mean ± STD, followed by the gain of Ours over that variant with an unpaired Welch 95% CI. Positive gains favor Ours.

SettingFace ID ↑Hair ↓Motion ↓Head ↓
w/o BioMoE (Naive Parameter Sharing)0.480 ±\pm 0.008; +0.1535 [0.1471, 0.1599]0.477 ±\pm 0.005; +0.0570 [0.0535, 0.0604]9.80 ±\pm 0.18; +2.81 [2.69, 2.93]0.576 ±\pm 0.005; +0.1212 [0.1169, 0.1255]
w/o BioMoE (Task-Specific Model)0.623 ±\pm 0.016; +0.0103 [−0.0006, 0.0213]0.428 ±\pm 0.003; +0.0080 [0.0062, 0.0098]7.10 ±\pm 0.09; +0.11 [0.04, 0.18]0.461 ±\pm 0.004; +0.0065 [0.0029, 0.0102]
w/o Structure-Aware Routing0.630 ±\pm 0.003; +0.0036 [−0.0009, 0.0081]0.425 ±\pm 0.002; +0.0049 [0.0032, 0.0066]7.04 ±\pm 0.10; +0.05 [−0.02, 0.13]0.470 ±\pm 0.008; +0.0150 [0.0096, 0.0204]
w/o Task-specific Experts0.580 ±\pm 0.004; +0.0531 [0.0485, 0.0578]0.432 ±\pm 0.002; +0.0118 [0.0101, 0.0135]7.82 ±\pm 0.06; +0.83 [0.77, 0.88]0.491 ±\pm 0.006; +0.0357 [0.0312, 0.0402]
w/o Routed Experts0.629 ±\pm 0.003; +0.0047 [0.0001, 0.0093]0.426 ±\pm 0.003; +0.0060 [0.0037, 0.0083]7.15 ±\pm 0.17; +0.16 [0.04, 0.28]0.466 ±\pm 0.007; +0.0114 [0.0062, 0.0166]
Ours0.633 ±\pm 0.0070.420 ±\pm 0.0016.99 ±\pm 0.080.455 ±\pm 0.005

Overall, the full model achieves the best mean scores across all four tasks, with predominantly positive gain CIs supporting the effectiveness of each design choice.

(b). Training data and budget

The w/o BioMoE (Naive Parameter Sharing), w/o BioMoE (Task-Specific Model), and w/o Two-Stage Training variants use the same training data and matched training costs, as do the other ablation variants unless otherwise specified.

(c). Adaptive LoRA rank allocation

Our adaptive rank allocation aims to improve overall multi-task performance under a fixed parameter budget, rather than outperform uniform allocation on every individual task. Uniform allocation under-allocates capacity to motion transfer, whose task-specific weights differ more substantially from those of the other tasks. Our strategy addresses this imbalance by assigning more capacity to motion transfer while maintaining similar performance on the remaining tasks, yielding the best overall performance among the compared strategies (+7.42% over uniform allocation; Overall: the equally weighted average of direction-adjusted relative gains across all four tasks).

Rank allocationFace ID ↑Hair ↓Motion ↓Head ↓Overall relative gain (%) ↑
Uniform0.6360.41910.130.4600.00
Gradient-based0.6340.4337.350.464+5.73
Result-based (Ours)0.6370.4217.090.460+7.42

W2. Generalization and Scalability

Our scalability claim concerns adding the evaluated tasks and combinations with limited data and training cost through lightweight task-specific modules. We evaluate facial-part adaptation and hat+lip composition quantitatively, and accessory transfer and hair+glasses composition qualitatively. Expansion to substantially larger task sets remains future work.

(a). Adaptation details

Adaptation uses 5K training samples for hat and hat+lip transfer, and 7K for glasses-related tasks. Each new task is adapted using four A40 GPUs for about 20K steps, requiring approximately 3 GPU-days, less than 10% of unified pretraining.

During adaptation, we train only task-specific modules, mainly LoRA experts, gates, and projections. All other modules, including the global shared expert and routed experts, are frozen to preserve previously learned capabilities.

For compositional transfer, we inherit two task-specific experts from the parent tasks and introduce a dedicated gate to weight their contributions (see Appendix A.6 for adaptation details).

AuthorsOpenReview ↗

This is Part 2/2 of our reply

(b). Quantitative evaluation

Compositional attribute transfer is a unique capability of our unified framework, so no directly comparable method is available. We therefore construct a two-stage baseline for quantitative comparison, using hat+lip transfer as a representative task. The baseline applies Qwen-Image-Edit for hat transfer followed by FuseAnyPart for lip transfer. Hat and lip CLIP distances measure transfer fidelity, while SSIM over the remaining regions measures preservation.

MethodHat CLIP dist. ↓Lip CLIP dist. ↓Other-region SSIM ↑
Qwen + FuseAnyPart0.4720.3930.823
Ours0.4460.3850.890

Ours achieves lower CLIP distances for both attributes and higher preservation SSIM, supporting better compositional transfer and preservation than the two-stage baseline (see Appendix B.6).

(c). Preservation of previously learned capabilities

We support two adaptation modes:

  • Reuse: We update the inherited parent LoRA experts without duplicating them.
  • Copy: We train independent copies of the parent task-specific modules, including LoRA experts, gates, and projections, while leaving the original modules, such as the global expert and routed experts, unchanged.

Both modes use the dedicated gate introduced in (a) / Appendix A.6. We evaluate their effect on the parent lip task before and after hat+lip adaptation, as shown in the table below.

Lip-transfer evaluationCLIP dist. ↓Expr. dist. ↓SSIM ↑
Before hat+lip adaptation0.3780.5790.937
After hat+lip adaptation (reuse)0.3800.5840.896
After hat+lip adaptation (copy)0.3780.5790.937

Reuse leads to a moderate drop in parent-task performance, which may be due to catastrophic forgetting, a common phenomenon in continual learning. Continual learning strategies such as replay could help mitigate forgetting; however, we do not explore them further here because this work focuses on showing the generalization and rapid task adaptation; Copy mode preserves the original task capabilities with only a small parameter overhead (see Appendix B.6).

Q3. Evaluation Protocol

(a). Preprocessing failures and common samples

MethodsPreprocessing failure rate
RigFace73.3%
All other methods≈1%

Due to its reliance on multiple pretrained detection models, RigFace exhibits a high preprocessing failure rate (73.3%, as shown in the table above). In the main comparison tables, all other methods are evaluated on their common successful subset, yielding 987 of the original 1,000 FFHQ pairs (see Appendix A.4). RigFace is evaluated separately because including its much smaller successful set in this intersection would substantially reduce sample count and diversity in face size, head pose, and expression (see Appendix A.5). As an additional check, we lower RigFace’s detection threshold to increase coverage and compare all methods on a common successful set, where RigFace’s results deteriorate and the ranking remains unchanged.

Complex-scene metrics are calculated on a common successful set for all compared methods within each scenario (see Appendix C.3).

(b). Human evaluation

Sixteen participants evaluate five examples randomly sampled from the first 100 test pairs for each of the four tasks, choosing among Ours and two baselines presented in shuffled order. A tie contributes equally to all three methods, and the accumulated credits are normalized by the total number of responses per task (see Appendix C.4).

Ethics Concerns

We use the datasets under their respective usage conditions and have documented their sources and permitted uses in the revised ethics statement. We do not treat public availability as consent to identity manipulation.

Our release will include explicit restrictions against non-consensual identity manipulation, impersonation, and harassment, together with documentation of intended research uses and known limitations. We will provide data-preparation instructions rather than redistribute the source datasets.

We observe weaker performance on Asian faces, which may be largely attributable to the training-data distribution, a limitation shared by most works that also rely solely on public datasets. Reliable ethnicity annotations are unavailable for our evaluation samples, so this observation remains qualitative.

Reviewer_w1s1OpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

UniBioTransfer is a single framework for face, hair, motion, and head transfer. It combines a unified data construction strategy with a MoE model and two-stage training, and can be fine-tuned for additional tasks such as facial-part and accessory transfer.

Critical strengths

Handling all four tasks in one model is useful, and the results are competitive with both unified and task-specific methods. The data construction strategy addresses a concrete problem. The evaluation is fairly broad.

Critical weaknesses
  1. Training cost is reported using different GPU and the paper claims comparable model size, but mainly reports inference memory rather than total parameters or checkpoint size.
  2. BioMoE offers improvements compared to directly sharing parameters, but the paper does not report differences in parameter count and computational complexity between the two.
  3. For some methods in RigFace and complex scenarios, the author excluded samples that failed to process each. Therefore, is it possible that the average index of different methods corresponds to different image pairs?
Questions for authors
  1. Does BioMoE still perform better with matched model size and training cost? Are the routing gains stable across runs?
  2. Does the training cost include both stages and data preparation? How many parameters does the model use?
  3. How often does each method fail? Do the rankings hold on the same test samples?
  4. Can you add quantitative results for compositional transfer and attribute preservation?
4 comments
AuthorsOpenReview ↗

This is Part 1/2 of our reply

We sincerely thank the reviewer for the careful reading and detailed suggestions.

We have revised the manuscript accordingly and respond below. Due to word count limitations, we will answer all your questions in two parts. This is the first part.

W1. Model Size and Training Cost

We add total parameter counts and estimated A100-equivalent training costs to Table 7.

MethodTaskVRAM (GB)Time (s/img)Train (GPU-days)Estimated A100-equivalent GPU-daysTotal parameters (B)
REFaceFace11.81218 (A100)18.01.3
StableHairHair19.316UnknownUnknown2.7
HunyuanPortraitMotion18.4103384 (A100)384.02.2
GHOST2.0Head8.0<1232 (A100)232.00.5
OursAll14.51534 (A40)19.42.3
Ours (Adapt.)Unseen15.4163 (A40)1.72.4~2.6

Parameter counts for all compared methods are listed below.

MethodREFaceCanonSwapHunyuanPortraitStable-HairHairFusionGHOST2.0Face-AdapterRigFaceOurs
Total parameters (B)1.30.42.22.71.90.52.22.12.3

Our model supports all four tasks with 2.3 B parameters, comparable in size to most diffusion models.

Based on NVIDIA’s official GPU specifications, we conservatively estimate our training cost at 19.4 A100-equivalent GPU-days. Our model supports all four tasks at a training cost comparable to or lower than those reported for task-specific baselines.

W2. BioMoE versus Direct Parameter Sharing

Unless otherwise specified, all ablation variants use the same training data and matched training costs. As shown in the table below, we compare BioMoE with naive parameter sharing at its original model size and after increasing the baseline’s parameter count to match that of BioMoE (see Appendix B.3 for implementation details).

SettingFace ↑Hair ↓Motion ↓Head ↓Time (s/img)VRAM (GB)Parameters (B)
w/o BioMoE (Naive Parameter Sharing)0.4780.4749.740.57413.511.81.5
w/o BioMoE (Naive Parameter Sharing) (matched size)0.4820.4779.900.55914.814.12.3
Ours0.6370.4217.090.46014.514.52.3

BioMoE improves performance across all four tasks with comparable efficiency. Even with matched model size, naive sharing consistently underperforms BioMoE, supporting the effectiveness of our design in mitigating task interference.

Q1. Multi-seed Ablations on BioMoE

We repeat the model-design ablations in Table 4 and add two variants: w/o Task-specific Experts and w/o Routed Experts. Results are aggregated over 12 training–inference seed combinations (see Appendix B.3 for details).

Each cell reports mean ± STD, followed by the gain of Ours over that variant with an unpaired Welch 95% CI. Positive gains favor Ours.

SettingFace ID ↑Hair ↓Motion ↓Head ↓
w/o BioMoE (Naive Parameter Sharing)0.480 ±\pm 0.008; +0.1535 [0.1471, 0.1599]0.477 ±\pm 0.005; +0.0570 [0.0535, 0.0604]9.80 ±\pm 0.18; +2.81 [2.69, 2.93]0.576 ±\pm 0.005; +0.1212 [0.1169, 0.1255]
w/o BioMoE (Task-Specific Model)0.623 ±\pm 0.016; +0.0103 [−0.0006, 0.0213]0.428 ±\pm 0.003; +0.0080 [0.0062, 0.0098]7.10 ±\pm 0.09; +0.11 [0.04, 0.18]0.461 ±\pm 0.004; +0.0065 [0.0029, 0.0102]
w/o Structure-Aware Routing0.630 ±\pm 0.003; +0.0036 [−0.0009, 0.0081]0.425 ±\pm 0.002; +0.0049 [0.0032, 0.0066]7.04 ±\pm 0.10; +0.05 [−0.02, 0.13]0.470 ±\pm 0.008; +0.0150 [0.0096, 0.0204]
w/o Task-specific Experts0.580 ±\pm 0.004; +0.0531 [0.0485, 0.0578]0.432 ±\pm 0.002; +0.0118 [0.0101, 0.0135]7.82 ±\pm 0.06; +0.83 [0.77, 0.88]0.491 ±\pm 0.006; +0.0357 [0.0312, 0.0402]
w/o Routed Experts0.629 ±\pm 0.003; +0.0047 [0.0001, 0.0093]0.426 ±\pm 0.003; +0.0060 [0.0037, 0.0083]7.15 ±\pm 0.17; +0.16 [0.04, 0.28]0.466 ±\pm 0.007; +0.0114 [0.0062, 0.0166]
Ours0.633 ±\pm 0.0070.420 ±\pm 0.0016.99 ±\pm 0.080.455 ±\pm 0.005

Overall, the full model achieves the best mean scores across all four tasks, with predominantly positive gain CIs supporting the effectiveness of each design choice.

AuthorsOpenReview ↗

This is Part 2/2 of our reply

W3. Preprocessing Failures and Evaluation Fairness

MethodsPreprocessing failure rate
RigFace73.3%
All other methods≈1%

Due to its reliance on multiple pretrained detection models, RigFace exhibits a high preprocessing failure rate (73.3%, as shown in the table above). In the main comparison tables, all other methods are evaluated on their common successful subset, yielding 987 of the original 1,000 FFHQ pairs (see Appendix A.4). RigFace is evaluated separately because including its much smaller successful set in this intersection would substantially reduce sample count and diversity in face size, head pose, and expression (see Appendix A.5). As an additional check, we lower RigFace’s detection threshold to increase coverage and compare all methods on a common successful set, where RigFace’s results deteriorate and the ranking remains unchanged.

Complex-scene metrics are calculated on a common successful set for all compared methods within each scenario (see Appendix C.3).

Q2. Training Cost Accounting

Following the convention in prior work, our reported training cost covers both training stages and online preprocessing, but excludes offline training-data construction (see Appendix A.1).

Q4. Compositional Transfer and Attribute Preservation

Compositional attribute transfer is a unique capability of our unified framework, so no directly comparable method is available. We therefore construct a two-stage baseline for quantitative comparison, using hat+lip transfer as a representative task. The baseline applies Qwen-Image-Edit for hat transfer followed by FuseAnyPart for lip transfer. Hat and lip CLIP distances measure transfer fidelity, while SSIM over the remaining regions measures preservation.

MethodHat CLIP dist. ↓Lip CLIP dist. ↓Other-region SSIM ↑
Qwen + FuseAnyPart0.4720.3930.823
Ours0.4460.3850.890

Ours achieves lower CLIP distances for both attributes and higher preservation SSIM, supporting better compositional transfer and preservation than the two-stage baseline (see Appendix B.6).

Results on attribute preservation for the four primary tasks, including identity similarity, pose/expression distances, and non-hair SSIM, can be found in Tables 1 and 2.

Reviewer_w1s1OpenReview ↗

Maintaining Acceptance

Thank you for the detailed response. The matched-size comparison and evaluation clarifications largely address my concerns. I will maintain my acceptance rating.

Thank you for your follow-up. We are glad that our response has addressed your concern, and we sincerely appreciate your constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.