Meitu's CFT Framework Solves Instability in AI Portrait Relighting

机器之心 · wechat · 2026-08-22

Meitu Imaging Lab (MTLab) proposes Consistent Feature Transport (CFT) to address instability issues in existing diffusion model-based portrait relighting, such as chaotic shadows and identity alteration. The method reframes relighting as a consistent feature transport problem, explicitly learning the illumination transformation between source and target distributions based on the RectifiedFlow framework. It uses image pairs with different content but identical illumination changes to supervise the core loss, preventing the model from fitting content-specific artifacts.

Core Approach & Dataset

Three-Part Training: Includes conditional generation (noise to target), reconstruction (noise to source for content preservation), and the core consistency feature transport (source to target).
Large-Scale Dataset: A dataset covering 14 illumination types (indoor/outdoor) was built, emphasizing multi-color temperature and spatially structured lighting.

Experimental Results

CFT outperforms existing methods in fidelity and perceptual similarity, showing superior stability in complex scenarios like multi-light interaction and color temperature shifts.
Ablation studies confirm the full CFT model (L₁+L₂+L₃) achieves the highest scores across SSIM, PSNR, LPIPS, and FID metrics.
The method also demonstrates generalizability in style transfer tasks (e.g., OmniStyle) and is now integrated into Meitu products like Picchi and AirBrush.

Original post →

Meitu's CFT Framework Solves Instability in AI Portrait Relighting

More from Multimodal