Research map / Image generation / Fast diffusion sampling
ReDi: Efficient Learning-Free Diffusion Inference via Trajectory Retrieval
Kexun Zhang, Xianjun Yang, William Yang Wang, Lei Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add ReDi top-2 retrieval | PNDM · MS-COCO val, NFE 20 | FID: 0.274 -> 0.255 | Improved |
| add ReDi top-2 retrieval | DPM-Solver · MS-COCO val, NFE 10 | FID: 0.268 -> 0.261 | Improved |
| use top-2 instead of top-1 retrieved neighbors (ablation) | ReDi on PNDM · MS-COCO val, NFE 30 | FID: 0.264 -> 0.247 | Improved |
| add ReDi top-1 retrieval | PNDM · MS-COCO val, NFE 40 | FID: 0.262 -> 0.262 | No clear effect |
| use trajectory samples instead of CLIP embeddings as retrieval key (ablation) | ReDi · MS-COCO val, NFE 30 | FID: 0.2784 -> 0.2643 | Improved |
| use trajectory samples instead of CLIP embeddings as retrieval key (ablation) | ReDi · MS-COCO val, NFE 40 | L2 distance to true value: 9.03 -> 7.95 | Improved |
| add retrieval instead of vanilla skipping from k to v (ablation) | ReDi · PNDM k=40, v=20 | FID: 1.58 -> 0.267 | Improved |
| use conditional x_T initial sample instead of unconditional for knowledge base (ablation) | ReDi · PNDM, NFE 20-40 | FID: 0.268 -> 0.265 | No clear effect |
| change guidance scale away from the 7.5 used to build the knowledge base (ablation) | ReDi · MS-COCO val | FID: 0.264 -> 0.271 | Hurt |
| shrink knowledge base size from 50K to 20K (ablation) | ReDi · MS-COCO val, PNDM | FID | Hurt |
| increase gap k-v with k=40 (v from 35 to 25) (ablation) | ReDi · MS-COCO val, PNDM | FID | Hurt |
| lower key step k from 40 to 20 with k-v=15 (ablation) | ReDi · MS-COCO val, PNDM | FID | Hurt |
Later papers that compared with this method
3 later accepted papers report ReDi as a baseline in their results tables; 3 of them report a better number on most of the results they share with it.
- REGLUE Your Latents with Global and Local Semantics for Entangled Diffusion (ECCV 2026 · better on 2 of 2 shared results)
- Semantics Lead the Way: Harmonizing Semantic and Texture Modeling with Asynchronous Latent Diffusion (CVPR 2026 · better on 1 of 1 shared result)
- ReCON: Training-Free Acceleration for Text-to-Image Synthesis with Retrieval of Concept Prompt Trajectories (ECCV 2024 · better on 5 of 5 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Tracing the Roots: Leveraging Temporal Dynamics in Diffusion Trajectories for Origin Attribution (NeurIPS 2025 · similarity 0.90)
- Hierarchical Koopman Diffusion: Fast Generation with Interpretable Diffusion Trajectory (NeurIPS 2025 · similarity 0.89)
- On the Trajectory Regularity of ODE-based Diffusion Sampling (ICML 2024 · similarity 0.89)
- T-Stitch: Accelerating Sampling in Pre-Trained Diffusion Models with Trajectory Stitching (ICLR 2025 · similarity 0.89)
- A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models (ICML 2024 · similarity 0.89)
- Plug-and-Play Diffusion Distillation (CVPR 2024 · similarity 0.88)
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.
