title: "Pro-Pose: Unpaired Full-Body Portrait Synthesis via Canonical UV Maps" description: >- ECCV2026 paper note: Pro-Pose introduces canonical UV space modeling with donor-based reposing to decouple pose from texture, leveraging 470K unpaired in-the-wild images and 30K paired samples to synthesize controllable full-body avatars in standardized base clothing, serving as a clean canvas for virtual try-on. tags: - ECCV2026 - 3D Vision - Pose-guided Synthesis - Canonical UV Space - Flow Matching - Virtual Try-On date: 2026-09-19 content_hash: 210a39c251b86b58

Pro-Pose: Unpaired Full-Body Portrait Synthesis via Canonical UV Maps¶

Conference: ECCV 2026
Paper: ECCV Paper
Project: https://pro-pose-portrait.github.io
Area: 3D Vision
Keywords: pose-guided portrait synthesis, canonical UV space, base clothing canonicalization, donor-based UV reposing, virtual try-on

TL;DR¶

Pro-Pose develops an unpaired full-body portrait synthesis framework operating in canonical UV texture space with a donor-based reposing strategy that eliminates pose-texture boundary leakage, enabling photorealistic, pose-controllable avatar generation in standardized minimal garments from a single in-the-wild image and acting as an occlusion-free geometric canvas for virtual try-on.

Background & Motivation¶

Photographs captured by everyday users in uncontrolled environments typically suffer from cluttered backgrounds, suboptimal illumination, arbitrary poses, and complex garment occlusions. Transforming such in-the-wild snapshots into studio-grade, pose-controllable portrait avatars while preserving personal biometric identity is vital for digital human synthesis and virtual try-on (VTO). However, existing 2D diffusion-based reposing methods—such as CFLD, MCLD, and LEFFA—rely almost exclusively on small-scale paired benchmarks like DeepFashion, which contains only around 100 unique identities. This narrow distribution leads to severe overfitting, identity drift on novel individuals, and an inability to disentangle or modify source clothing. Conversely, 3D avatar approaches and Gaussian splatting methods demand rigorous multi-view capture or 3D ground-truth geometry, leaving them starved of real-world identity diversity.

The fundamental tension stems from the lack of large-scale paired datasets depicting identical subjects across diverse poses and outfits. While single-view in-the-wild photographs are abundant, naively training an image-to-image generator on unpaired data where input and target coincide collapses into a trivial identity copy-paste mapping, failing to learn 3D geometric transformations and occlusion handling. Furthermore, downstream virtual try-on pipelines routinely struggle when applied directly to unconstrained everyday photos because source garments impose structural and geometric interference that contaminates novel garment warping.

To circumvent the paired data bottleneck, Pro-Pose shifts synthesis from 2D pixel space into a canonical SMPL-X UV texture domain coupled with an automated Base Clothing (BC) standardization pipeline that converts subjects into minimal black tank tops and shorts. Core Idea: operate in a canonical full-body UV texture space using a donor-based reposing mechanism—intersecting input visibility masks with unrelated donor masks to extinguish boundary-based pose shortcuts—thereby unlocking joint Flow Matching training on massive unpaired single images alongside scarce paired data for high-fidelity, identity-preserving reposing.

Method¶

Overall Architecture¶

Given a single in-the-wild reference photo $I_{p'}$ in source pose $p'$, Pro-Pose synthesizes a standardized, pose-controllable portrait $A_p$ in target pose $p$ (driven by SMPL-X mesh renderings) wearing minimal Base Clothing (BC). The model takes as input the unwrapped partial UV texture map, the SMPL-X target pose rendering, and an optional identity-anchoring facial crop. The generative backbone is built upon a frozen Flux.1 [dev] Latent Rectified Flow Transformer (DiT), predicting vector fields that transport standard Gaussian noise toward clean latent portrait representations. The comprehensive pipeline spans Base Clothing dataset preparation, dual-branch Flow Matching joint training (paired supervision and single-view donor self-supervision), and optional test-time few-shot LoRA personalization.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["In-the-Wild Image Input<br/>Arbitrary clothing & scene"] --> B["Base Clothing Standardization<br/>Gemini 2.5 Flash edited minimal outfit"]
    B --> C["Canonical Full-Body UV Parsing<br/>Extract partial UV texture & SMPL-X"]
    C --> D["Dual-Branch Flow Matching<br/>Paired supervision + donor self-supervision"]
    D --> E["Test-Time Few-Shot Adaptation<br/>Skin-masked LoRA fine-tuning"]
    E --> F["Controllable Identity Avatar<br/>Reposed portrait / UV editing / VTO canvas"]

Key Designs¶

1. Base Clothing Standardization: Canonical Canvas and Relighting Distillation Directly reposing unconstrained clothing leaks source garment silhouettes into target poses and disrupts subsequent downstream try-on. To establish a standardized benchmark, the authors construct a Base Clothing (BC) dataset across DeepFashion paired pairs, FFHQ face collections, and diverse commerce images using Gemini 2.5 Flash Image. For full-body captures, prompt-guided pixel-aligned editing replaces varied apparel with a standard black sleeveless tank top and shorts while strictly preserving bodily shape, identity, and posture. For cropped portraits such as FFHQ, generative outpainting synthesizes full-body extensions into the minimal outfit. By distilling these standardized targets, the generative network simultaneously learns to attenuate baked-in highlight artifacts and render uniform studio lighting consistent with the subject.

2. Donor-Based UV Reposing: Eliminating Self-Supervised Pose Leakage In canonical UV texture space, unwrapping a single-view photo yields a partial texture $T_p = T \odot M_p$, where $M_p \in \{0, 1\}^{H \times W}$ is the binary visibility mask dictated by pose $p$. Crucially, the occlusion boundaries of $M_p$ correlate perfectly with the source pose. In single-view self-supervision, the network can minimize loss by copy-pasting visible pixels along these mask contours rather than learning 3D geometric transport. To shatter this shortcut, the authors introduce Donor-based UV Reposing. For each sample, an unrelated "donor" mask $M_{\tilde{p}}$ exhibiting an IoU overlap between $[0.4, 0.8]$ with $M_p$ is sampled from a precomputed training pool. The input texture is masked by this donor: $$T_{p \rightarrow \tilde{p}} = T_p \odot M_{\tilde{p}} = T \odot (M_p \odot M_{\tilde{p}})$$ Because binary intersection is commutative, the resulting boundary profile no longer uniquely identifies source pose $p$, compelling the network to perform true geometric inpainting and novel-view texture completion.

3. Dual-Branch Flow Matching: Synergistic Paired and Unpaired Learning To harmonize scarce paired data with abundant single-view images, the generative backbone employs Rectified Flow Matching over the Flux.1 [dev] 16-channel VAE latent space. The velocity network $v_\theta$ learns the transport vector field between Gaussian noise $x_1$ and data latent $x_0 = \mathcal{E}(I_p)$ at timestep $t \sim \mathcal{U}[0, 1]$: $$\mathcal{L}_{\text{FM}} = \mathbb{E}_{t, x_0, x_1} \left\| v_\theta(x_t, t, c) - (x_1 - x_0) \right\|^2$$ The conditioning token set $c$ is dynamically assigned according to the branch. In the paired supervision branch, the model conditions on reference partial texture $T_{p'}$, SMPL-X target pose $p$, and a high-resolution MediaPipe face crop $I_{p'}^{\text{FC}}$. In the single-view self-supervision branch, the model conditions on the donor-reposed texture $T_{p \rightarrow \tilde{p}}$ and target pose $p$, but strictly drops the face crop condition ($I^{\text{FC}} = \emptyset$). This omission prevents trivial identity leakage, forcing the network to synthesize facial biometrics purely from the warped canonical UV texture.

4. Test-Time Personalization via Few-Shot Adaptation Under extreme orientation shifts—such as generating a frontal portrait from a rear-view reference—single-view UV textures contain zero frontal cues, causing feed-forward models to hallucinate facial characteristics. Pro-Pose mitigates this via test-time personalization. Given a compact set $S$ of $N$ reference images of a subject across arbitrary poses, cross-image pairs are formed. The network fine-tunes Rank-128 LoRA adapters in the DiT attention projections and MLP layers using a foreground skin-masked flow matching objective: $$\mathcal{L}_{\text{FT}} = \mathbb{E}_{t, x_{j,0}, x_{j,1}} \left\| M_j \odot \left( v_\theta(x_{j,t}, t, c_{\text{paired}}) - (x_{j,1} - x_{j,0}) \right) \right\|^2$$ where $M_j$ is the target latent skin segmentation mask. This rapid subject-specific adaptation anchors biometric likeness and sharpens delicate facial contours.

Loss & Training¶

The framework is trained on roughly 500K total images: 470K unpaired single-view photographs (from FFHQ and commercial repositories) and 30K paired images from DeepFashion (encompassing all ~100 identities). Optimization is performed with AdamW at a learning rate of $10^{-4}$ across 128 TPUv5 accelerators with a global batch size of 128 for 100K iterations. Condition dropout is enforced during training: dropping all conditions with probability $0.05$, or dropping individual inputs with probabilities $0.3$ for texture $T_{\text{in}}$, $0.3$ for face crop $I_{\text{in}}^{\text{FC}}$, and $0.1$ for target pose $p_{\text{target}}$. Single-image inference takes approximately 51 seconds.

Key Experimental Results¶

Main Results¶

Evaluations are conducted on the in-domain DeepFashion In-Shop test split (8,570 pairs) and the challenging out-of-domain WPose dataset (2,305 pairs, measured with foreground-masked M-PSNR, M-SSIM, and M-LPIPS), benchmarking against specialized diffusion reposing models and state-of-the-art vision-language foundational systems.

Dataset	Method	PSNR / M-PSNR↑	FID↓	SSIM / M-SSIM↑	LPIPS / M-LPIPS↓	OKS↑	FaceSim↑	DINO↑	HPSv3↑
DeepFashion (In-Domain)	CFLD	17.65	7.15	0.748	0.182	0.48	0.3180	0.9731	4.15
	MCLD	18.21	7.08	0.756	0.176	0.49	0.3440	0.9654	4.29
	LEFFA	14.02	4.23	0.755	0.119	0.44	0.5794	0.9409	4.41
	OnePoseTrans	13.57	8.74	0.605	0.307	0.46	0.5750	0.9476	4.32
	UniHuman	14.05	6.25	0.796	0.156	0.46	0.5810	0.9434	4.03
	Gemini 2.5 Flash Image	16.98	4.59	0.738	0.179	0.43	0.5815	0.9691	7.19
	Gemini 3 Pro Image	17.51	4.30	0.775	0.109	0.45	0.5856	0.9705	7.22
	Ours (Unpaired + Paired)	19.36	4.24	0.818	0.075	0.48	0.6047	0.9759	7.24
WPose (Out-of-Domain)	CFLD	15.43	96.07	0.744	0.208	0.31	0.0885	0.6412	1.94
	MCLD	15.64	94.23	0.759	0.201	0.35	0.0995	0.6478	1.96
	LEFFA	16.71	67.85	0.776	0.193	0.32	0.0914	0.5725	2.01
	OnePoseTrans	17.23	27.43	0.818	0.151	0.33	0.1735	0.7205	4.44
	UniHuman	17.64	27.75	0.807	0.161	0.34	0.1121	0.7207	2.89
	Gemini 2.5 Flash Image	16.67	9.55	0.779	0.149	0.32	0.4713	0.7005	7.35
	Gemini 3 Pro Image	17.19	7.15	0.795	0.145	0.33	0.5241	0.7119	7.40
	Ours (Unpaired + Paired)	19.95	5.99	0.860	0.121	0.38	0.5571	0.7394	7.55

Ablation Study¶

The ablation investigates the indispensability of unpaired self-supervision and the impact of test-time adaptation.

Config	PSNR↑	FID↓	SSIM↑	LPIPS↓	FaceSim↑	Note
Unpaired Only (WPose)	16.13	6.95	0.761	0.215	0.3805	Lacks direct cross-pose alignment supervision; degraded pose warping
Paired Only (WPose)	18.30	6.65	0.820	0.155	0.4959	Overfits DeepFashion identities; poor out-of-domain generalization
Full Model (Unpaired + Paired, WPose)	19.95	5.99	0.860	0.121	0.5571	Combines massive identity variety with accurate pose conditioning
Ours Zero-Shot	18.59	6.88	0.823	-	0.4837	Base feed-forward model on challenging blurred or degraded inputs
Ours Fine-Tuned	19.34	6.76	0.835	-	0.5722	FaceSim gains +18.3% relative, restoring sharp individual facial traits

Key Findings¶

Unpaired Single Images Drive In-the-Wild Robustness: Relying exclusively on paired training data achieves strong in-domain DeepFashion results (PSNR 19.38) but degrades sharply on out-of-domain WPose (M-PSNR falls to 18.30, FID rises to 6.65). Integrating 470K unpaired images elevates WPose M-PSNR to 19.95 and drives FID down to 5.99, demonstrating that unpaired data is essential for wild generalization.
Resilience to Extreme Pose Degradation: In stratified evaluations partitioning test cases by Euler-angle head rotation differences, competing models (LEFFA, MCLD, UniHuman, OnePoseTrans) suffer catastrophic identity collapse ($\text{FaceSim} < 0.4$) in over 97% of cases. Pro-Pose maintains identity fidelity with a failure rate of only 13.5%, closely matching the natural identity variance of the Ground Truth.
Empowering Downstream Virtual Try-On: In a 30,000-vote human preference study on WPose (10 raters evaluating 500 subjects across 6 target garments), applying Google VTO directly onto original photos yielded only a 28.02% preference rate with a 31.04% garment artifact rate. Preprocessing subjects with Pro-Pose into canonical avatars boosted VTO preference to 71.98% and slashed garment errors to 17.34%.

Highlights & Insights¶

Elegantly Bypassing 3D Online Rendering: Conventional self-supervision requires expensive online rendering of textured SMPL-X meshes to construct synthetic pairs, which lacks high-frequency hair geometry. Applying donor mask intersections ($T \odot (M_p \odot M_{\tilde{p}})$) directly in 2D UV space eliminates pose leakage with minimal computational overhead.
De-risking Virtual Try-On via Minimal Garment Canvases: Standardizing subjects into neutral black tank tops and shorts decouples body geometry from source fashion, eliminating collar and sleeve leakage that historically plagued VTO pipelines on in-the-wild inputs.
Direct UV Texture Editing Emergence: Because synthesis is anchored strictly to canonical UV parameters, localized modifications applied to the 2D texture (such as drawing a tattoo or altering makeup) geometrically propagate across novel target poses with subpixel consistency without retraining.

Limitations & Future Work¶

Facial Hallucination Under Total View Occlusion: Synthesizing a frontal portrait from a rear-view reference forces the feed-forward model to hallucinate facial identity due to zero UV overlap, necessitating multi-frame test-time fine-tuning for faithful reconstruction.
Sensitivity to Single-View SMPL-X Shape Fitting: Single-view 3D body estimation errors—such as underestimating limb volume or misinterpreting silhouette boundaries—propagate directly into the reposed avatar, causing distorted anatomical proportions.
Reconstruction Limits in Deep Self-Occlusions: Intricate skin and anatomical details concealed within tight self-occluded folds cannot be recovered from single views without learning multi-reference consistency priors.

vs MCLD / LEFFA (2D Diffusion Reposing): While MCLD and LEFFA rely on paired DeepFashion data and suffer high identity breakage (>97%) under real-world pose shifts, Pro-Pose unlocks 470K unpaired images through donor UV reposing, outperforming them on WPose (M-PSNR 19.95 vs 15.64/16.71) while stripping away clothing occlusions.
vs UniHuman / OnePoseTrans (Unified Human Editing): UniHuman attempts generalized editing in pixel space but lacks explicit UV geometric decoupling, producing weak identity preservation on large pose shifts (FaceSim 0.11-0.17 vs 0.557 for Pro-Pose).
vs Gemini 2.5 Flash / Gemini 3 Pro Image (Multimodal LLM Inpainting): Although large multimodal vision models exhibit impressive zero-shot inpainting, Pro-Pose delivers tighter keypoint alignment (OKS 0.48 vs 0.45) and superior biometric likeness in extreme pose variations due to explicit SMPL-X flow matching guidance.

Rating¶

Novelty: ⭐⭐⭐⭐⭐ (The donor-based UV reposing mechanism elegantly prevents pose leakage, pioneering scalable unpaired avatar synthesis.)
Experimental Thoroughness: ⭐⭐⭐⭐⭐ (Thorough validation across DeepFashion and WPose, pose-stratified identity curves, extensive ablations, and a 30K-vote VTO blind study.)
Writing Quality: ⭐⭐⭐⭐⭐ (Crisp motivation, rigorous mathematical formulation in UV space, and clear diagrams.)
Value: ⭐⭐⭐⭐⭐ (Provides a dual breakthrough: high-fidelity digital avatar reposing and an indispensable canonical intermediate representation for commercial virtual try-on.)