Improved image generation by predicting rather than denoising pretrained features
Dino Forcing Flow Models: Do not denoise what you can predict
Computer Vision and Pattern Recognition
Summary
Generating images with AI often involves cleaning up noisy pictures step by step, which can be slow and complicated. The paper shows how predicting certain image features directly, instead of denoising them, makes the image creation process faster and produces better results. The authors tested this approach on large image sets and it outperformed previous methods, requiring fewer training steps. This new way of guiding the AI simplifies the process while improving quality.
What this means in practice
- •For computer vision engineers: Train image generation models faster and achieve higher-quality images by predicting pretrained features instead of denoising them iteratively.
- •For graphics software developers: Integrate an improved AI image synthesis method that requires fewer training resources while enhancing generated image quality for creative applications.$Commercial implications: Enables companies to offer better AI-driven image generation tools with lower training costs to artists and content creators.
Authors
Arijit Ghosh, Lucas Degeorge, Paul Couairon, Alexei A Efros, Vicky Kalogeiton, David Picard
Abstract
Co-denoising pretrained representations such as DINO can substantially improve the training speed and quality of flow matching models, but it introduces a second denoising trajectory and requires carefully designed schedules. We propose a simpler alternative: predict the pretrained representation directly, then condition the model on its own prediction. This removes the need for a second ODE and any representation-specific denoising schedules, while retaining the benefits of representation guidance. Our approach converges substantially faster and achieves better generation quality as measured by FID score. On ImageNet, it outperforms the state of the art in latent space at 2x fewer epochs than prior methods; in pixel space, it improves FID over comparable prior methods by more than 20%. These results support a simple principle: do not denoise what you can predict. Our code is openly available at https://github.com/arijit-hub/dino_forcing.