免蒸馏，只改一个Loss！5行代码实现扩散模型4步生成
Ranking
Overall
86
Content
95
Popularity
64
Observed public metrics from 1 member.
Merged summary
TL;DR - Perceptual Flow Matching replaces MSE with a semantic perceptual loss to generate images and videos in 4–8 steps, approaching the quality of conventional 35–50-step flow-matching models without distillation or teacher models.
- Perceptual supervision penalizes off-manifold predictions, reducing the posterior-mean blurring that accumulates during few-step inference.
- The method was evaluated on text-to-image generation, image editing, and video generation.
- Experiments link stronger off-manifold discrimination in feature spaces such as DINO and SigLIP to better few-step generation.
- Huber, pixel-space, and randomly initialized ViT losses did not provide the same gains, indicating that semantically rich supervision spaces are essential.
Sources (1)
免蒸馏，只改一个Loss！5行代码实现扩散模型4步生成
Public signals
Hugging Face upvotes 19
TL;DR - Perceptual Flow Matching replaces MSE with a semantic perceptual loss to generate images and videos in 4–8 steps, approaching the quality of conventional 35–50-step flow-matching models without distillation or teacher models.
- Perceptual supervision penalizes off-manifold predictions, reducing the posterior-mean blurring that accumulates during few-step inference.
- The method was evaluated on text-to-image generation, image editing, and video generation.
- Experiments link stronger off-manifold discrimination in feature spaces such as DINO and SigLIP to better few-step generation.
- Huber, pixel-space, and randomly initialized ViT losses did not provide the same gains, indicating that semantically rich supervision spaces are essential.