SVDtrunc

Compressed FLUX.1 models from Importance-Aware Low-Rank Distillation of Diffusion Transformers (ECCV 2026).

Denis Zavadski^1,2,*, Sebastian Heid^1,2,*, Damjan Kalšan¹, Stefan Roth^2,3,4, Carsten Rother^1,2

_{¹Heidelberg University · ²Zuse School ELIZA · ³TU Darmstadt · ⁴hessian.AI · ^*Equal contribution}

_{Each row uses the same prompt: FLUX.1-dev on the left, then SVDtrunc at 68%, 57%, 49% and 39% of its transformer parameters. SVDtrunc-m and SVDtrunc-s are the 68% and 57% models released here.}

Model	Base model	Transformer parameters	File
SVDtrunc-m	FLUX.1-dev	8.1B (68%)	FLUX.1-dev/svdtrunc-m.safetensors
SVDtrunc-s	FLUX.1-dev	6.8B (57%)	FLUX.1-dev/svdtrunc-s.safetensors
SVDtrunc-schnell-70	FLUX.1-schnell	8.3B (70%)	FLUX.1-schnell/svdtrunc-schnell-70.safetensors

Only the transformer is compressed. The text encoders and the autoencoder are those of the base model.

Usage

The checkpoints are loaded with the SVDtrunc code:

git clone https://github.com/vislearn/SVDtrunc && cd SVDtrunc && uv sync
uv run svdtrunc generate --checkpoint CVL-Heidelberg/SVDtrunc/FLUX.1-dev/svdtrunc-m.safetensors \
    --prompt "a red fox sitting in fresh snow, golden hour" --output-dir outputs

The text encoders and the autoencoder come from the base model, FLUX.1-dev or FLUX.1-schnell. Both are gated: accept the license on the base model's Hub page and log in with hf auth login first. SVDtrunc-schnell-70 generates images in 4 steps, like FLUX.1-schnell. The code repository also shows how to use the models from Python and how to compress FLUX.1 to other sizes.

License

The weights follow their base model:

• FLUX.1-dev/: SVDtrunc-m and SVDtrunc-s are derived from FLUX.1-dev and released under the FLUX.1 [dev] Non-Commercial License. See NOTICE.
• FLUX.1-schnell/: SVDtrunc-schnell-70 is derived from FLUX.1-schnell and released under the Apache 2.0 license.

Citation

@inproceedings{zavadski2026svdtrunc,
  title     = {Importance-Aware Low-Rank Distillation of Diffusion Transformers},
  author    = {Zavadski, Denis and Heid, Sebastian and Kal{\v{s}}an, Damjan and Roth, Stefan and Rother, Carsten},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}