{"paper":{"title":"PermuQuant: Lowering Per-Group Quantization Error by Reordering Channels for Diffusion Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Reordering channels to group similar statistics reduces per-group quantization error in diffusion models.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Junxian Li, Kai Liu, Kaiwen Tao, Renjing Pei, Yongsen Cheng, Yulun Zhang, Zhikai Chen, Zhixin Wang","submitted_at":"2026-05-10T12:26:50Z","abstract_excerpt":"Large-scale visual generative models have achieved remarkable performance. However, their high computational and memory costs make deployment challenging in resource-constrained scenarios, such as interactive applications and personal single-GPU usage. Post-training quantization (PTQ) offers a practical solution by compressing pretrained models without expensive retraining. However, existing PTQ methods still suffer from severe quality degradation under extremely low-bit settings. In this paper, we identify channel ordering as an important but underexplored factor in per-group quantization. In"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"PermuQuant consistently reduces quantization error and outperforms existing PTQ baselines. On FLUX.1-dev with an RTX 5090, PermuQuant achieves up to a 1.8× single step speedup and reduces the DiT memory footprint by 3.5× under W4A4 NVFP4 quantization.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That a permutation chosen on calibration data via the joint second-moment criterion will generalize to the full input distribution at inference time and will not introduce new artifacts in the generated outputs.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"PermuQuant reduces per-group quantization error in diffusion models by sorting channels with similar activation and weight statistics into the same groups using a calibration-checked permutation.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Reordering channels to group similar statistics reduces per-group quantization error in diffusion models.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"f4943accd8ed586b19be0e6617004ba1b2428ada61acfacd21b03dbdbf95d9b8"},"source":{"id":"2605.09503","kind":"arxiv","version":2},"verdict":{"id":"8c63ad3d-c07e-4742-b55a-a0c4e3a52aa5","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-12T04:17:00.312464Z","strongest_claim":"PermuQuant consistently reduces quantization error and outperforms existing PTQ baselines. On FLUX.1-dev with an RTX 5090, PermuQuant achieves up to a 1.8× single step speedup and reduces the DiT memory footprint by 3.5× under W4A4 NVFP4 quantization.","one_line_summary":"PermuQuant reduces per-group quantization error in diffusion models by sorting channels with similar activation and weight statistics into the same groups using a calibration-checked permutation.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That a permutation chosen on calibration data via the joint second-moment criterion will generalize to the full input distribution at inference time and will not introduce new artifacts in the generated outputs.","pith_extraction_headline":"Reordering channels to group similar statistics reduces per-group quantization error in diffusion models."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.09503/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"claim_evidence","ran_at":"2026-05-20T07:42:01.380920Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T16:41:37.255516Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T13:01:17.787173Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T10:10:27.411584Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"7f6d5d5810d9779f60c7b85a0866222517c7ca8056bd740c810b98e5e21dce8d"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}