
Furkan Horoz1, Ozgur Kara1, Kaan Oktay2, Adil Kaan Akan2, Atakan Kara2,
Anand Bhattad3,β , James Matthew Rehg1,β
1University of Illinois at Urbana-Champaign 2fal 3Johns Hopkins University
β Equal advising.
ReDiverse-FT adapter for SDXL. Few-step distilled students keep their teacher's activation variance across seeds but
concentrate it into far fewer directions at one block of the first denoising step, so different seeds give near-identical
samples. This rank-32 LoRA on the student's blocks up to up_blocks.0.attentions.0 is trained to match the teacher's activations there and
is active in the first sampling step only; the remaining steps run the unmodified student.
| Teacher | 4-step student | This adapter |
|---|---|---|
| SDXL-base 1.0 | DMD2, distilled from the teacher | ReDiverse-FT LoRA of the student, rediverse_ft.safetensors (216 MB) |

Usage
git clone https://github.com/H-furkan/rediverse && cd rediverse # set up the environments of envs/README.md
hf download ReDiverse/ReDiverse-SDXL --local-dir ReDiverse-SDXL
python -m sdxl.generate --arm rediverse --adapter ReDiverse-SDXL --prompt "a photo of a fox in the snow" --num-samples 4 # image env
--arm rediverse_ft uses the adapter alone; --arm rediverse (above) also adds the training-free ReDiverse-TF directions, which first runs the teacher on 25 noises for each new prompt (about 15 s on an 80 GB H100) and caches the result. Without --adapter, the code downloads the same file from ReDiverse/ReDiverse.
| DINO diversity β (GenEval) | Student | ReDiverse-FT (this adapter) | ReDiverse (FT + TF) | Teacher (multi-step) |
|---|---|---|---|---|
| SDXL | 59.4 | 64.6 | 65.0 | 66.3 |
License
Trained on the DMD2 SDXL student, whose license is CC BY-NC-SA 4.0, so this adapter is released under CC BY-NC-SA 4.0 for non-commercial use. The SDXL base weights it runs with remain subject to the use-based restrictions of Attachment A of the CreativeML Open RAIL++-M License. NOTICE credits DMD2 and SDXL.
Citation
@article{horoz2026rediverse,
title = {Restoring Diversity in Distilled Image and Video Generators by Probing Activations},
author = {Horoz, Furkan and Kara, Ozgur and Oktay, Kaan and Akan, Adil Kaan and Kara, Atakan
and Bhattad, Anand and Rehg, James Matthew},
journal = {arXiv preprint arXiv:2611.18550},
year = {2026}
}
- Downloads last month
- -
Model tree for ReDiverse/ReDiverse-SDXL
Base model
tianweiy/DMD2