ReDiverse is a study by researchers from the University of Illinois Urbana-Champaign, fal, and Johns Hopkins University: the authors found that few-step distilled models produce nearly identical images when changing seeds because the activations of a specific module in the first step are compressed into fewer directions. This rank-32 LoRA aligns the activations of the 4-step MiniMax-H3 Turbo with those of the original MiniMax-H3 at that location, taking effect only during the first sampling step; the authors report that the diversity score on VBench increased from 36.5 to 44.1, and further rose to 50.9 when combined with the training-free ReDiverse-TF (the original multi-step model scores 52.5). A version for the 6-step Pika H3 Flex was also released concurrently, with each of the two files being 612 MB.
Why it matters Few-step distilled models often produce nearly identical shots when changing seeds; this method only affects the first step, leaving the remaining steps unchanged, and a version for Pika H3 Flex was also released concurrently.
Hugging Face likes 9 · Hugging Face trending score 9 · Hugging Face downloads 48
License: Conditional commercial use · ComfyUI: Not available · VRAM: Not specified
Regional License: The MiniMax H3 open-weight license excludes the EU, UK, South Korea, and the US (Excluded Territories in the LICENSE); separate authorization from MiniMax is required in these four regions.
Reasons
- Includes the paper, activation probing code, and same-seed comparison examples; other versions in the same series are available for SDXL, SD3, and Wan 2.1
Usage
- Clone github.com/H-furkan/rediverse and set up the video environment according to envs/README.md, agree to the terms and log in on Hugging Face, then run hf download ReDiverse/ReDiverse-MiniMax-H3 --local-dir ReDiverse-MiniMax-H3, and execute python -m h3.generate --arm rediverse_ft --adapter ReDiverse-MiniMax-H3 --prompt "…" --num-samples 4; for the Pika H3 Flex version, use generate_pika.py from its repository, and adding --compare will simultaneously generate videos with the same seed but without the adapter
Risks
- Released for research purposes and subject to the MiniMax H3 community license, which does not apply to the EU, UK, South Korea, and the US; downloading requires logging in and agreeing to the terms
- When using --arm rediverse to overlay training-free directions, each new prompt requires running the original model 25 times (the authors report this takes approximately 1 GPU hours on a 80 GB H100); it has only been trained and tested on text-to-video generation, and the authors only provided command-line scripts, with no ComfyUI nodes available yet