{"paper":{"title":"eDiff-I: Text-to-Image Diffusion Models with an Ensemble of Expert Denoisers","license":"http://creativecommons.org/licenses/by/4.0/","headline":"An ensemble of stage-specialized diffusion models improves text alignment in image synthesis at the same inference cost.","cross_cats":["cs.LG"],"primary_cat":"cs.CV","authors_text":"Arash Vahdat, Bryan Catanzaro, Jiaming Song, Karsten Kreis, Miika Aittala, Ming-Yu Liu, Qinsheng Zhang, Samuli Laine, Seungjun Nah, Tero Karras, Timo Aila, Xun Huang, Yogesh Balaji","submitted_at":"2022-11-02T17:43:04Z","abstract_excerpt":"Large-scale diffusion-based generative models have led to breakthroughs in text-conditioned high-resolution image synthesis. Starting from random noise, such text-to-image diffusion models gradually synthesize images in an iterative fashion while conditioning on text prompts. We find that their synthesis behavior qualitatively changes throughout this process: Early in sampling, generation strongly relies on the text prompt to generate text-aligned content, while later, the text conditioning is almost entirely ignored. This suggests that sharing model parameters throughout the entire generation"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Our ensemble of diffusion models, called eDiff-I, results in improved text alignment while maintaining the same inference computation cost and preserving high visual quality, outperforming previous large-scale text-to-image diffusion models on the standard benchmark.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The synthesis behavior qualitatively changes throughout the generation process such that early stages rely on text conditioning while later stages largely ignore it, making a single shared-parameter model suboptimal.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"An ensemble of stage-specialized text-to-image diffusion models improves prompt alignment over single shared-parameter models while preserving visual quality and inference speed.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"An ensemble of stage-specialized diffusion models improves text alignment in image synthesis at the same inference cost.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"d19838852ca4de31ef0d83cb2cbcd3555be749baa7f1b69390136d0a2a8993bb"},"source":{"id":"2211.01324","kind":"arxiv","version":5},"verdict":{"id":"ea0e3657-7f80-4a78-848a-196756727a3f","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T01:40:27.202143Z","strongest_claim":"Our ensemble of diffusion models, called eDiff-I, results in improved text alignment while maintaining the same inference computation cost and preserving high visual quality, outperforming previous large-scale text-to-image diffusion models on the standard benchmark.","one_line_summary":"An ensemble of stage-specialized text-to-image diffusion models improves prompt alignment over single shared-parameter models while preserving visual quality and inference speed.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The synthesis behavior qualitatively changes throughout the generation process such that early stages rely on text conditioning while later stages largely ignore it, making a single shared-parameter model suboptimal.","pith_extraction_headline":"An ensemble of stage-specialized diffusion models improves text alignment in image synthesis at the same inference cost."},"references":{"count":95,"sample":[{"doi":"","year":2021,"title":"V ., Du, J., Iyer, S., Pasunuru, R., et al","work_id":"711d98e0-b9e1-4350-b81d-a621de85a6bb","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Blended latent diffusion","work_id":"1aef642e-c5e8-42d8-a391-f41998245355","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Blended diffusion for text-driven editing of natural images","work_id":"b7d69bd8-2483-45c0-93f6-0bf59101b935","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Estimating the optimal covariance with imperfect mean in diffusion probabilistic models","work_id":"6a02cf87-87e3-4e01-8807-1815408583fe","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Analytic- DPM: An analytic estimate of the optimal reverse variance in diffusion probabilistic models","work_id":"83033437-6996-4dfd-9f0f-41f417388b99","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":95,"snapshot_sha256":"534b6679319b65d105d3f306b7c44eee8a40b75cbb5a95d217ab7b4a06494e91","internal_anchors":11},"formal_canon":{"evidence_count":2,"snapshot_sha256":"25a2d4e30e413f2993b5b0dc4bbf4653a831193a26daf8e758a4a4c926d03d5a"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}