Back to projects
Diffusion / Post-training
Self-Evolving Diffusion
Qwen-VL and on-policy self-distillation for self-evolving diffusion models.
?
An ongoing project exploring how diffusion models can improve through their own generated experience. The current direction combines Qwen-VL with on-policy self-distillation to study self-evolving training loops for multimodal generation and reasoning.
The project is still in progress; more details and release materials will be added as the experiments mature.
