vvvvvjdy/D-OPSD

What it solves

D-OPSD addresses the challenge of continuously tuning step-distilled diffusion models (models designed for fast, few-step inference) without losing their original knowledge or their ability to generate high-quality images in just a few steps.

How it works

It uses an on-policy self-distillation framework. The system assigns the same model two different roles—a teacher and a student—by providing them with different contexts. This allows the model to perform supervised tuning on its own generated outputs (roll-outs) without needing an external reward function or additional modules.

Who it’s for

Researchers and developers working with text-to-image diffusion models, specifically those using LLM/VLM encoders who want to fine-tune models for new concepts, styles, or domain preferences while maintaining fast inference speeds.

Highlights

• On-Policy Self-Distillation: Enables supervised tuning using the model's own outputs.
• Preservation of Capability: Maintains few-step inference speed and previous knowledge during tuning.
• Versatile Application: Validated across different settings, including full fine-tuning for domain adaptation (e.g., anime) and small LoRA training for learning new concepts from limited data.
• No External Rewards: Operates without the need for external reward functions or extra modules.