Research map / Image generation / Fast diffusion sampling

D-OPSD: On-Policy Self-Distillation for Continuously Tuning Step-Distilled Diffusion Models

NeurIPS 2026 · Image generation › Fast diffusion sampling · approach: Distillation for few-step generation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes D-OPSD, an on-policy self-distillation method that continually fine-tunes step-distilled diffusion models by using the same model as a text-only student and a text+target-image teacher, matching velocities on the student's own few-step rollouts.

Problem

Continually fine-tuning step-distilled few-step diffusion models with standard SFT on offline noised target images degrades their few-step inference quality, while online RL needs reward functions that secondary developers typically lack.

Approach

Modern diffusion models with LLM/VLM encoders inherit in-context ability, so feeding the encoder text plus the target image yields a stronger teacher condition; the model acts as student (text-only) and teacher (multimodal, EMA or frozen) and the student is trained to match teacher velocities on its own few-step rollouts, with no reward model.

Key results

• In LoRA customization on Z-Image-Turbo, D-OPSD reaches Quality-S 3.7965 and CLIP-S 0.3664, versus 2.4236 Quality-S for Vanilla SFT, while keeping 8-step inference.
• In full fine-tuning on Z-Image-Turbo, D-OPSD obtains FID 40.49 (base 48.69) and DINO-D 0.1088, with GenEval 0.7170 vs base 0.7543.
• Vanilla SFT on Z-Image-Turbo drops GenEval to 0.1588 in full fine-tuning, showing severe loss of prior knowledge; D-OPSD costs roughly 4x FLOPs and 2x training time per iteration versus SFT.

Surprising findings

• Modern diffusion models with LLM/VLM encoders inherit in-context ability: conditioning on text plus target image produces concept-preserving variations without any extra training.
• Naively replacing Qwen3-4B with Qwen3-VL-4B for the teacher introduces high-frequency artifacts; swapping in the Qwen3-4B LLM weights while keeping ViT and connector fixes it.

Limitations

• Training fails when the teacher cannot generate concept-consistent images under multimodal conditioning, since it then gives no useful supervision.
• Relies on the base model's in-context ability of the LLM/VLM encoder; the VLM-based teacher needed an encoder weight swap to avoid artifacts.
• Roughly 4x FLOPs and 2x training time per iteration compared with vanilla SFT.

Experimental setup

Datasets: DreamBooth dataset in-house 25K anime image-text dataset

Benchmarks: GenEval DPG

Metrics: DINO distance LPIPS distance FID VLM judgment (VLM-J) CLIP score Quality score Aesthetic score GenEval score DPG score

Baselines: Vanilla SFT SFT + LoRA on distilled Dreambooth PSO DMD (two-stage) Base-Model

Builds on: on-policy distillation on-policy self-distillation DreamBooth Diffusion-DPO DMD

Closest papers by meaning

• Plug-and-Play Diffusion Distillation (CVPR 2024 · similarity 0.90)
• Diffusion Model is Effectively Its Own Teacher (CVPR 2025 · similarity 0.89)
• Tuning Timestep-Distilled Diffusion Model Using Pairwise Sample Optimization (ICLR 2025 · similarity 0.89)
• Diffusion Self-Distillation for Zero-Shot Customized Image Generation (CVPR 2025 · similarity 0.89)
• Random Conditioning for Diffusion Model Compression with Distillation (CVPR 2025 · similarity 0.89)
• On improved Conditioning Mechanisms and Pre-training Strategies for Diffusion Models (NeurIPS 2024 · similarity 0.89)