Guo-Jun Qi

AI Scientist/Professor, Foundation Models/MLLM, Diffusion Models, Multimodal/Image/Video Generative Models, IEEE/IAPR Fellow

[Guo-Jun Qi @guojunq]

Check our work on scaling diffusion models at test time by self-guidance. It only uses a diffusion model on its own with no network/memory overhead.

This presents how test-time scaling law in @OpenAI o1 model succeeds in diffusion models.

The report: https://t.co/Iu1MZgWBeK.

[Guo-Jun Qi @guojunq]

OpenAI @OpenAI used Reinforcement FineTuning (ReFT) for COT in LLM.

Last week, we submitted a report https://t.co/yZ4RFTzt5b where we presented the ReFT for diffusion and flow based models.

Diffusion time process is like COT reasoning step by step on how to generate images.

[Guo-Jun Qi @guojunq]

A brave idea of training with "reinforcement learning from reward" to dynamically schedule the diffusion time.

TPDM samples images more rapidly than the benchmark flux. The longer and more complex prompt, the more slowly the diffusion time decreases to generate more details.

[Guo-Jun Qi @guojunq]

Check out our one-step diffusion generator called Score Implicit Matching (SIM) at NeurIPS 2024 at https://t.co/ozErwKpobX #neurips2024