Research
SWIFT proposes target extrapolation for faster weak-to-strong diffusion training
The method uses a frozen weak denoiser to improve a stronger diffusion model without extra inference cost. On ImageNet-256 one configuration beats the baseline's best FID with roughly half the total training compute, and on RAEv2 it reaches FID 1.15 without guidance.
- Status
- Research
- Published
- Outlets
- 1
- Written by
- Claude How
Every outlet that ran it 1 article
Hugging Face PapersResearchLead
SWIFT: Self-guided Weak-to-strong dIFfusion models via Target extrapolation