| { |
| "schema_version": 1, |
| "base_model": { |
| "repo_id": "black-forest-labs/FLUX.1-dev", |
| "revision": "3de623fc3c33e44ffbe2bad470d0f45bccf2eb21", |
| "license_name": "flux-1-dev-non-commercial-license", |
| "license_url": "https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/3de623fc3c33e44ffbe2bad470d0f45bccf2eb21/LICENSE.md" |
| }, |
| "artifacts": { |
| "FLUX.1-dev-w8a8-convrot.safetensors": { |
| "policy": "whole_dit_w8a8_convrot", |
| "convrot_group_size": 256, |
| "counts": { |
| "w8a8": 304, |
| "preserve_bf16": 476 |
| }, |
| "limitations": [ |
| "Paper scope is the FLUX.1 diffusion transformer; CLIP, T5, VAE, biases, norms, and non-block linears remain BF16.", |
| "ComfyUI fuses single-block q/k/v/MLP input projections into linear1, expanding component-level decisions to the fused tensor." |
| ], |
| "source_manifest_sha256": "859aa6f5321c0a528426a0518d1c3b20d0b1701d84896c1cf9c34611a8a8b322" |
| }, |
| "FLUX.1-dev-int8_convrot.safetensors": { |
| "policy": "partial_profile_int8_convrot", |
| "convrot_group_size": 256, |
| "limitations": [ |
| "Profile-based selective quantization rather than the whole-DiT paper ablation.", |
| "Larger checkpoint because more sensitive tensors remain at higher precision." |
| ], |
| "source_manifest_sha256": null |
| }, |
| "FLUX.1-dev-paper-w4a4-convrot.safetensors": { |
| "policy": "paper_whole_dit_w4a4_convrot", |
| "convrot_group_size": 256, |
| "quantization_group_size": 64, |
| "counts": { |
| "w4a4": 301, |
| "preserve_bf16": 479 |
| }, |
| "limitations": [ |
| "Memory-first experimental variant with substantially lower paired PSNR than INT8 variants.", |
| "ComfyUI fuses single-block q/k/v/MLP input projections into linear1, expanding component-level decisions to the fused tensor." |
| ], |
| "source_manifest_sha256": "aaa49f1d13d719eb813addf57e380f4d77c01032db1b33e03b8d9c1156fb7390" |
| }, |
| "t5xxl_flux1_int8_convrot.safetensors": { |
| "policy": "t5xxl_encoder_w8a8_convrot", |
| "convrot_group_size": 256, |
| "counts": { |
| "w8a8": 168, |
| "preserve_bf16": 51 |
| }, |
| "limitations": [ |
| "Experimental T5-XXL policy not evaluated by the ConvRot paper.", |
| "Image similarity does not establish text-comprehension equivalence." |
| ], |
| "source_manifest_sha256": "5ef2d35fef8945a9e9e5c68c028aa74a0ea47f2e834e80614fc4ab9542199bfb" |
| } |
| }, |
| "paper": { |
| "title": "ConvRot: Rotation-Based Plug-and-Play 4-bit Quantization for Diffusion Transformers", |
| "arxiv": "2512.03673" |
| } |
| } |
|
|
