Can Consistent Feature Transport Preserve Identity While Relighting Portraits? ECCV 2026 CFT Method

Problem

Portrait relighting often requires changing lighting conditions while keeping the subject’s identity, geometry, and background unchanged. Current diffusion‑based methods fall into three categories: control‑based, decomposition‑based, and direct image‑to‑image translation. Each suffers from issues such as dependence on accurate control signals, error propagation causing inconsistent shadows or color shifts, and insufficient modeling of lighting‑specific feature displacement.

Limitations of Existing Datasets

Public relighting datasets are small and captured in controlled environments, lacking diverse lighting effects such as multi‑color‑temperature, spatially varying shadows, and non‑uniform illumination, which hampers learning of robust, controllable relighting models.

CFT Proposal

MT Lab introduces Consistent Feature Transport (CFT), reformulating portrait relighting as a lighting‑consistent feature‑transfer problem. CFT learns the transformation between source and target image distributions explicitly within the Rectified Flow framework, rather than treating the task as generic image translation.

Core Idea and Losses

CFT jointly models three distributions: noise, source image, and target image. Training uses three loss components:

L₁ (Noise → Target) : Given source image x_{src}, target image x_{tgt} and target text condition c_{tgt}, the model learns a velocity field that maps prior noise to the target latent representation.

L₂ (Noise → Source) : Under the same noise, a neutral condition c_{src} reconstructs the source distribution, reinforcing preservation of identity, geometry, and scene structure.

L₃ (Source → Target – Consistent Feature Transport) : Inspired by non‑invertible editing, CFT exploits the linearity of Rectified Flow. Using the parallelogram rule, the direct transport is defined as z_{tdirect}=z_{src}+z_{ttgt}-z_{tsrc} with velocity field

v_{tdirect}=v_θ(z_{ttgt},t,x_{src},c_{tgt})-v_θ(z_{tsrc},t,x_{src},c_{src})

. Supervision for L₃ uses image pairs that share the same lighting transformation but differ in identity or scene content, preventing the model from conflating lighting changes with other variations.

Dataset Construction

To support complex lighting learning, the team built a large‑scale portrait relighting dataset covering indoor and outdoor scenes, with 14 distinct lighting effect categories. The dataset emphasizes spatially structured illumination, multiple color temperatures, and intricate shadows, addressing the scarcity of such examples in existing benchmarks.

Experimental Results

On a self‑constructed test set, CFT outperforms prior methods across pixel fidelity (SSIM, PSNR), perceptual similarity (LPIPS), distribution realism (FID), and dedicated lighting metrics. Qualitative examples show clear advantages in multi‑light interaction, structured lighting, and large color‑temperature shifts.

Configuration   SSIM   PSNR   LPIPS   FID
Only L₁        0.9153 22.92  0.0985  20.38
L₁+L₂          0.9141 22.83  0.1018  21.61
L₁+L₃          0.9161 23.38  0.0972  19.32
L₃ (original)  0.9141 22.80  0.1003  20.30
L₃ (random)    0.9135 22.80  0.1035  19.55
Full CFT (L₁+L₂+L₃) 0.9202 23.51 0.0946 18.73

User study with 30 participants rating 40 random samples shows that the combination Flux‑Kontext + CFT achieves the highest scores on lighting quality (IQ), content consistency (CC), physical plausibility (PP), and image aesthetics (IA).

Ablation Study

Adding L₂ alone does not improve relighting quality, but when combined with L₃ it provides extra structural preservation. Using original (same‑identity) pairs for L₃ supervision degrades performance, confirming the benefit of “same lighting, different content” pairing. Variance analysis of the learned velocity fields indicates that the full CFT model yields the lowest variance on both training and test sets, reflecting more stable lighting feature transport.

Generalization to Style Transfer

The “consistent feature transport” principle is not limited to lighting. By constructing a style‑transfer dataset (OmniStyle‑150k) and applying CFT on OmniStyle and Qwen‑Image‑Edit backbones, the method consistently improves content structure and perceptual fidelity while achieving better style balance, demonstrating its applicability to other attribute‑editing tasks.

Deployment

The relighting capability powered by CFT has been integrated into Meitu’s products such as Picchi, AirBrush, and BeautyPlus, enabling diverse lighting effects without lengthy textual prompts.