Integration Matters: Rollout-Based Training for Constrained Diffusion Models

arXiv cs.AI · 6 hours ago

Researchers introduced a fine-tuning method for constrained diffusion models that integrates constraint guidance during training by differentiating through the fixed noise schedule of the denoising process. The approach improved constraint satisfaction across multiple tasks while maintaining competitive sampling quality compared to existing methods. By aligning training with the actual sampling trajectory, the method addresses the mismatch between training-time and sampling-time distributions that affects existing constrained generation approaches.