{"paper":{"title":"Improved Denoising Diffusion Probabilistic Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Simple modifications let denoising diffusion models achieve competitive log-likelihoods while supporting much faster sampling.","cross_cats":["cs.AI","stat.ML"],"primary_cat":"cs.LG","authors_text":"Alex Nichol, Prafulla Dhariwal","submitted_at":"2021-02-18T23:44:17Z","abstract_excerpt":"Denoising diffusion probabilistic models (DDPM) are a class of generative models which have recently been shown to produce excellent samples. We show that with a few simple modifications, DDPMs can also achieve competitive log-likelihoods while maintaining high sample quality. Additionally, we find that learning variances of the reverse diffusion process allows sampling with an order of magnitude fewer forward passes with a negligible difference in sample quality, which is important for the practical deployment of these models. We additionally use precision and recall to compare how well DDPMs"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We show that with a few simple modifications, DDPMs can also achieve competitive log-likelihoods while maintaining high sample quality. Additionally, we find that learning variances of the reverse diffusion process allows sampling with an order of magnitude fewer forward passes with a negligible difference in sample quality.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the chosen modifications to the noise schedule and variance parameterization do not introduce unmeasured biases in the learned distribution or sampling dynamics beyond what the reported metrics capture.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Targeted tweaks to DDPMs produce competitive likelihoods and high-quality samples, with learned reverse variances enabling 10x faster sampling and predictable scaling with compute.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Simple modifications let denoising diffusion models achieve competitive log-likelihoods while supporting much faster sampling.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"2118f6fc99ebc0835cf236c045cbb67c81f02ce3cffb4d1721a9025c2de7ef3f"},"source":{"id":"2102.09672","kind":"arxiv","version":1},"verdict":{"id":"e6030429-ad94-4c1a-b774-cc0fa9ea9572","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T19:15:07.553096Z","strongest_claim":"We show that with a few simple modifications, DDPMs can also achieve competitive log-likelihoods while maintaining high sample quality. Additionally, we find that learning variances of the reverse diffusion process allows sampling with an order of magnitude fewer forward passes with a negligible difference in sample quality.","one_line_summary":"Targeted tweaks to DDPMs produce competitive likelihoods and high-quality samples, with learned reverse variances enabling 10x faster sampling and predictable scaling with compute.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the chosen modifications to the noise schedule and variance parameterization do not introduce unmeasured biases in the learned distribution or sampling dynamics beyond what the reported metrics capture.","pith_extraction_headline":"Simple modifications let denoising diffusion models achieve competitive log-likelihoods while supporting much faster sampling."},"references":{"count":17,"sample":[{"doi":"","year":null,"title":"Large Scale GAN Training for High Fidelity Natural Image Synthesis","work_id":"244e6f06-bad2-4f34-8186-ff370286427f","ref_index":1,"cited_arxiv_id":"1809.11096","is_internal_anchor":true},{"doi":"","year":2011,"title":"Very deep vaes generalize autoregressive models and can outperform them on images","work_id":"214ff54a-9ea1-46bc-94a9-8daf75f1d2b9","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2017,"title":"Gans trained by a two time-scale update rule converge to a local nash equilibrium","work_id":"ee703933-f9ee-4a48-8621-5b66c0322d6d","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1902,"title":"Flow++: Improving flow-based generative models with variational dequantization and architecture design","work_id":"58dad468-996a-4b27-af63-673e9a09dfe9","ref_index":4,"cited_arxiv_id":"1902.00275","is_internal_anchor":true},{"doi":"","year":2009,"title":"Kynk¨a¨anniemi, T., Karras, T., Laine, S., Lehtinen, J., and Aila, T","work_id":"57a3305f-67e0-48e5-b939-33ac062b4b60","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":17,"snapshot_sha256":"8c6f8996e9a08ce74790d6a0faa84ff44c437f9badfc692f93d43c6244e6722e","internal_anchors":3},"formal_canon":{"evidence_count":2,"snapshot_sha256":"d76da9291b9c1eaf4f5b75f28ef3a0f2bdd14086981db0298a76c4bc840cacf5"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}