Research map / Image generation / Fast diffusion sampling
Distribution Backtracking Builds A Faster Convergence Trajectory for Diffusion Distillation
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 DisBack, which records a degradation path from the teacher to the initial student generator and then backtracks those intermediate distributions as a convergence trajectory to speed up score distillation for one-step diffusion generation.
Problem
Score distillation for one-step diffusion generation suffers from a score mismatch early in training because the teacher's score is evaluated on samples far from its training distribution, slowing convergence and hurting final quality.
Approach
Fine-tune a copy of the teacher on the student's initial samples and save intermediate checkpoints to form a degradation path; reversing it gives a convergence trajectory, and the student is distilled sequentially toward each checkpoint before the final teacher target.
Key results
- DisBack reaches FID 1.38 on ImageNet 64x64 one-step generation, versus 5.57 for Diff-Instruct and 1.51 for DMD2.
- On FFHQ 64x64 DisBack converges 1.97x faster than the variant without the trajectory constraint, and 5.67x faster on CIFAR-10.
- Distilled SDXL reaches FID 18.96 at one step on COCO 2014 and is preferred over original SDXL in 61.3% of user-study comparisons.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the convergence trajectory constraint (ablation) | DisBack · FFHQ 64×64 | FID: 12.26 -> 10.88 | Helped |
| add the convergence trajectory constraint (ablation) | DisBack · AFHQv2 64×64 | FID: 10.29 -> 9.97 | Helped |
| add the convergence trajectory constraint (ablation) | DisBack · ImageNet 64×64 | FID: 5.96 -> 1.38 | Helped |
| add the convergence trajectory constraint (ablation) | DisBack · LSUN-bedroom 256×256 | FID: 7.43 -> 6.99 | Helped |
| add the convergence trajectory constraint (ablation) | DisBack · LSUN-cat 256×256 | FID: 10.63 -> 10.30 | Helped |
| constrain the convergence trajectory instead of using the endpoint-only Diff-Instruct variant (ablation) | Score distillation · FFHQ 64×64 | convergence speed: 1.97 times faster | Helped |
| constrain the convergence trajectory instead of using the endpoint-only Diff-Instruct variant (ablation) | Score distillation · CIFAR10 | convergence speed: 5.67 times faster | Helped |
| constrain the convergence trajectory instead of using the endpoint-only Diff-Instruct variant (ablation) | Score distillation · ImageNet 64×64 | convergence speed: 1.79 times faster | Helped |
| distill LCM-LoRA using DisBack rather than the original LCM-LoRA | LCM-LoRA, SD v1.5 teacher · COCO 2014, 512×512, one-step | FID: 78.26 -> 36.37 | Helped |
| apply DisBack to SDXL for one-step generation | SDXL · COCO 2014 | Patch-FID: 26.89 vs. 20.86 | Hurt |
Surprising findings
- Reported FID for DisBack on ImageNet 64x64 at 1.38 is better than the 15-NFE EDM teacher in the same table? Not stated; the table instead lists EDM at 511 NFE with 1.36, slightly better than DisBack.
- Patch-FID of distilled SDXL is worse than the one-step DMD2 and multi-step baselines, showing a trade-off on text-to-image metrics.
Limitations
- Performance is bounded by the teacher model's quality.
- Requires careful design of the degradation path and hyperparameters such as epochs per intermediate node.
- Theoretical grounding is weak: the score-field path may not be a gradient field, so it may not correspond to a valid density path.
Experimental setup
Datasets: FFHQ 64x64 AFHQv2 64x64 ImageNet 64x64 LSUN-bedroom LSUN-cat CIFAR-10 CelebA LAION-Aesthetics COCO 2014
Benchmarks: FFHQ 64x64 AFHQv2 64x64 LSUN-bedroom LSUN-cat ImageNet 64x64 COCO 2014
Metrics: FID Patch-FID CLIP score IS user preference mismatch degree
Baselines: Diff-Instruct Progressive Distillation Consistency Model Consistency Trajectory Model DMD2 SDXL-Turbo LCM-SDXL SDXL Lightning Score GAN FastGAN
Builds on: Diff-Instruct Variational Score Distillation EDM Stable Target Field
Closest papers by meaning
- Score identity Distillation: Exponentially Fast Distillation of Pretrained Diffusion Models for One-Step Generation (ICML 2024 · similarity 0.92)
- One-Step Diffusion Distillation through Score Implicit Matching (NeurIPS 2024 · similarity 0.91)
- Diffusion Model is Effectively Its Own Teacher (CVPR 2025 · similarity 0.90)
- Accelerating Convergence of Score-Based Diffusion Models, Provably (ICML 2024 · similarity 0.90)
- One-step Diffusion with Distribution Matching Distillation (CVPR 2024 · similarity 0.90)
- Simple and Fast Distillation of Diffusion Models (NeurIPS 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.