Research map / Architectures and efficiency / Quantization
FourTune: Towards Fully 4-Bit Efficient Post-Training for Diffusion Models
Bowen Xue, Zihan Min, Xingyang Li, Zhekai Zhang, Haocheng Xi, Lvmin Zhang, Maneesh Agrawala, Jun-Yan Zhu, Song Han, Yujun Lin, Muyang Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| quantize weights, activations and gradients to 4 bit (W4A4G4) instead of BF16 LoRA | FourTune · FLUX.1-dev customization, end-to-end | memory / speed: 29.35 -> 14.51 GB; 2.27x speedup on RTX 5090 | Improved |
| W4A4G4 vs NF4 QLoRA training speed | FourTune · FLUX.1-dev customization, RTX 5090 | speedup: 2.79x | Improved |
| add frozen numerical stabilizer branch to naive W4A4G4 (ablation) | FourTune · knowledge distillation gradient norm | gradient stability | Improved |
| replace group-wise 1x16 with block-wise 16x16 quantization (ablation) | FourTune · FLUX.1-dev reconstruction, MJHQ-30K | LPIPS / PSNR: LPIPS 0.203 -> 0.227; PSNR 21.5 -> 20.4 | Mixed |
| lower activation and gradient precision from 16 to 4 bit (W4A16G16 -> W4A4G4) (ablation) | FourTune · RL on FLUX.1-dev, Aesthetic Score | Aes: 6.0816 -> 6.1779 | No clear effect |
| fuse LoRA-related kernels vs unfused 4-bit baseline (ablation) | FourTune · DiT training step latency RTX 5090 | speedup: 1.82x | Improved |
| additionally fuse MLP components (ablation) | FourTune · DiT training step latency RTX 5090 | speedup: 1.10x additional | Improved |
| use SparseLoRA instead of W4A4G4 for 4-step distillation | SparseLoRA · pi-Flow distillation, HPSv2 prompts | FID: 15.50 -> 202.51 | Hurt |
| W4A4G4 vs BF16 LoRA on customization | FourTune · Customization similarity, FLUX.1-dev | similarity: 0.771 -> 0.783 (Identity) | No clear effect |
| W4A4G4 vs BF16 full finetune on RL | FourTune · SRPO RL, FLUX.1-dev | Aes: 6.3447 -> 6.3119 | No clear effect |
| W4A4G4 vs BF16 LoRA on Qwen-Image Subject customization | FourTune · Customization, Qwen-Image | prompt following: 0.999 -> 0.983 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SparseLoRA — SparseLoRA: Accelerating LLM Fine-Tuning with Contextual Sparsity (ICML 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- SVDQuant: Absorbing Outliers by Low-Rank Component for 4-Bit Diffusion Models (ICLR 2025 · similarity 0.89)
- EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models (ICLR 2024 · similarity 0.88)
- Q-DiT: Accurate Post-Training Quantization for Diffusion Transformers (CVPR 2025 · similarity 0.88)
- PTQ4DiT: Post-training Quantization for Diffusion Transformers (NeurIPS 2024 · similarity 0.88)
- IntLoRA: Integral Low-rank Adaptation of Quantized Diffusion Models (ICML 2025 · similarity 0.88)
- Pioneering 4-Bit FP Quantization for Diffusion Models: Mixup-Sign Quantization and Timestep-Aware Fine-Tuning (CVPR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
