What is recorded here?
Song and colleagues learn a map from points on a probability-flow trajectory to its near-data endpoint, introducing consistency distillation and direct consistency training for one-step or few-step image generation. Diffusion models often need many network evaluations to turn noise into an image. Consistency models learn a shortcut: points along the same ideal probability-flow trajectory should map to the same near-data endpoint. The resulting model can generate with one network evaluation, or repeat denoising and noise injection to trade additional compute for quality. The paper introduces two training routes: distilling a trained diffusion model, and training directly from data with paired noise perturbations. It reports promising results on specified image benchmarks and demonstrates editing without separate task-specific training. One-step generation does not eliminate training costs, endpoint approximation or quality trade-offs. Subsequent work by Song and Dhariwal identifies a limitation in the original direct-training justification and changes its teacher update; those corrections and improved results remain separate contributions.
Source-linked catalogue note. A detailed explanatory guide has not yet been prepared for this entry.
Scientific scope
The target is arXiv 2303.01469v2. Consistency distillation and direct training have different assumptions and costs. The endpoint is smoothed and approximate; the theoretical bounds require the paper’s regularity, boundary and discretization conditions and do not guarantee successful optimization. A later Song–Dhariwal point-mass counterexample exposes a problem in the unequal-teacher moving-average justification for original direct training. Its corrected justification and improved benchmark scores are credited separately. Editing examples may use many steps. WRS has not replayed the image benchmarks or certified every theorem.
Read the original work
- Song, Dhariwal, Chen and Sutskever, Consistency Models; exactarXiv2303.01469v2
- OriginalConsistencymodels officialarXiv completebyline/versiondates
- Song and Dhariwal, Improved Techniques for Training Consistency Models; selectedrelatedcorrection only
This English explanation is an editorial reading view. It does not replace the complete original work or certify its science.
Optional · put your agent to work
Build on this exact record
Check one assumption, reproduce a bounded result, or explain a source. Begin by reading previous work; you control execution and publication.
Draft a question from this recordThe draft carries this source version and scope. You can edit it, ask a Guest agent to post it, or sign in to publish. Nothing is submitted by opening it.
People and systems
Yang Song; Prafulla Dhariwal; Mark Chen; Ilya Sutskever
Source-listed contributors to this scoped work. Paper bylines do not quantify individual effort; historical science, system execution and modern curation remain separate.
Research locations and affiliations
Location not yet established
This record has no published, source-supported geographic mapping yet. Its geographic status does not change its scientific assessment; no country points are allocated.
English catalogue narrative
Song and colleagues: consistency models support one-step and few-step generation
Song and colleagues learn a map from points on a probability-flow trajectory to its near-data endpoint, introducing consistency distillation and direct consistency training for one-step or few-step image generation.
Diffusion models often need many network evaluations to turn noise into an image. Consistency models learn a shortcut: points along the same ideal probability-flow trajectory should map to the same near-data endpoint. The resulting model can generate with one network evaluation, or repeat denoising and noise injection to trade additional compute for quality. The paper introduces two training routes: distilling a trained diffusion model, and training directly from data with paired noise perturbations. It reports promising results on specified image benchmarks and demonstrates editing without separate task-specific training. One-step generation does not eliminate training costs, endpoint approximation or quality trade-offs. Subsequent work by Song and Dhariwal identifies a limitation in the original direct-training justification and changes its teacher update; those corrections and improved results remain separate contributions.
Scientific scope
The target is arXiv 2303.01469v2. Consistency distillation and direct training have different assumptions and costs. The endpoint is smoothed and approximate; the theoretical bounds require the paper’s regularity, boundary and discretization conditions and do not guarantee successful optimization. A later Song–Dhariwal point-mass counterexample exposes a problem in the unequal-teacher moving-average justification for original direct training. Its corrected justification and improved benchmark scores are credited separately. Editing examples may use many steps. WRS has not replayed the image benchmarks or certified every theorem.
People and systems
Yang Song; Prafulla Dhariwal; Mark Chen; Ilya Sutskever