--- license: apache-2.0 base_model: - OpenMOSS-Team/MOSS-Image-SFT - lvladikov/MOSS-SFT-Distill-4step-LoRA base_model_relation: merge pipeline_tag: text-to-image tags: [comfyui, moss-image, step-distillation, int8, convrot, 4-step] --- # MOSS-Image SFT 4-step (LoRA merged): BF16 and INT8 ConvRot [lvladikov/MOSS-SFT-Distill-4step-LoRA](https://huggingface.co/lvladikov/MOSS-SFT-Distill-4step-LoRA) merged into MOSS-Image SFT. No LoRA node needed. Not an official OpenMOSS or Comfy-Org product. | File | Format | Size | |---|---|---| | `moss_image_sft_4step_merged_bf16.safetensors` | BF16 | 10.16 GB | | `moss_image_sft_4step_merged_int8_convrot.safetensors` | int8_tensorwise + ConvRot, per-row (layout identical to Comfy-Org `moss_image_sft_int8_convrot.safetensors`) | 5.41 GB | Put a file in `ComfyUI/models/diffusion_models/` and load it with **Load Diffusion Model**. Workflows: `workflows/`. ## Settings (keep them) 4 steps · `euler` · **CFG 1** (BasicGuider) · ManualSigmas `1.0, 0.91, 0.717, 0.024, 0.0` · trained sizes 768x512 to 1824x1024 (and portrait), multiples of 16. INT8 needs an NVIDIA GPU with INT8 tensor cores (SM 7.5+) and a ComfyUI with native int8_tensorwise + ConvRot support. ## How it was made 1. `moss_sft_4step_rank_64_lora_comfyui.safetensors` (rev 2ec5241400bf) merged into `moss_image_sft_bf16.safetensors` (Comfy-Org/MOSS-Image rev 2929efe3cdd5) in fp32 at strength 1.0, ComfyUI's runtime rule `W + (alpha / rank) * up @ down`. BF16 rounds once, from fp32. 2. INT8 from the fp32 merge: group-wise regular Hadamard rotation (ConvRot, group [256]), per-row symmetric INT8 (`mseclip`). Layer selection and metadata mirror Comfy-Org's official SFT INT8 file. ## Checks that passed * BF16: header identical to the base; 211 tensors byte-identical; 168 merged tensors equal bf16(W + ΔW) exactly (median ‖ΔW‖/‖W‖ = 0.0014). * INT8: header and every `comfy_quant` identical to the official file; all 168 quantized layers reconstruct the merged fp32 weight, median relative error 0.0084 (max 0.0096); the official file's own error on the unmerged base: 0.0086. * Quantizer basis proven on the unmerged base against the official file (median exact-match 1.0000); bit-identical to comfy-kitchen 0.2.41. * Remote sha256 of both checkpoints equals the local files. Known limitations are those of the LoRA (see its card). ## Credits and license MOSS-Image by OpenMOSS-Team (Apache-2.0), ComfyUI repack by Comfy-Org. 4-step LoRA by lvladikov (Apache-2.0), NOTICE included. ConvRot math ported from comfy-kitchen (Apache-2.0). Merge and quantization by [NidAll](https://huggingface.co/NidAll).