🤖 ReDiverse Restores Diversity to Few-Step Generators

A group from UIUC, fal, and Johns Hopkins explained the 'diversity collapse' in distillates: different seeds produce nearly identical images and clips. The student retains the teacher's final variance but compresses it into fewer directions—locally, in the first denoising block.

🌍 Distillation via DMD2, TDM, lightx2v, and Turbo is the standard for accelerating generators; the side effect is explained at the activation level and fixed with a cheap LoRA or without training: SDXL 59.4→65.0, Wan 2.1 31.6→52.2 with the same 4 steps, while CLIPScore is preserved.

👤 Running SDXL, SD3, Wan 2.1, MiniMax-H3, or Pika H3 Flex: ReDiverse-FT adapters are on Hugging Face, ReDiverse-TF is training-free; diversity is nearly as good as the teacher. MiniMax-H3 and Pika H3 Flex adapters are research-only: the license is not valid in the EU, UK, South Korea, and the US.

Source 1: https://arxiv.org/abs/2611.18550 Source 2: https://github.com/H-furkan/rediverse