MOSS-SFT-Distill-4step-LoRA

This is a rank 64 distilled LoRA trained by community author lvladikov for OpenMOSS's MOSS-Image SFT, using eDMD distribution matching combined with an LADD adversarial discriminator (the discriminator simultaneously compares against images from the original 50-step model and approximately 8000 real photos). It generates images in 4 steps with CFG 1, and training covers 12 resolutions ranging from 768×512 to 1824×1024. The author claims that on Comfy Cloud, the GPU time for generating the same 22 images at 1024×1024 resolution dropped from 387.3 seconds to 26.6 seconds. Out of 15 prompts requiring text generation, 14 rendered all text correctly (compared to 15 for the original model). The repository provides diffusers code, LoRA files calibrated for ComfyUI loading, and four types of workflows (text-to-image, layout bounding boxes, image editing, and PiD 4x upscaling), with one version each for local use and Comfy Cloud.

This is a rank 64 distilled LoRA trained by community author lvladikov for OpenMOSS's MOSS-Image SFT, using eDMD distribution matching combined with an LADD adversarial discriminator (the discriminator simultaneously compares against images from the original 50-step model and approximately 8000 real photos). It generates images in 4 steps with CFG 1, and training covers 12 resolutions ranging from 768×512 to 1824×1024. The author claims that on Comfy Cloud, the GPU time for generating the same 22 images at 1024×1024 resolution dropped from 387.3 seconds to 26.6 seconds. Out of 15 prompts requiring text generation, 14 rendered all text correctly (compared to 15 for the original model). The repository provides diffusers code, LoRA files calibrated for ComfyUI loading, and four types of workflows (text-to-image, layout bounding boxes, image editing, and PiD 4x upscaling), with one version each for local use and Comfy Cloud.

Why it matters The model is called only once per step without running negative prompts; the author claims GPU time for the same batch of images is reduced to approximately one-fourteenth. It includes ComfyUI-specific files and workflows with 8 core nodes.

Hugging Face Likes 13 · Hugging Face Upvotes 13 · Hugging Face Downloads 4

License: Commercial use allowed · ComfyUI: Mentioned in model card · VRAM: 196 GB

Reasons

  • The model card includes side-by-side comparisons with the same seed for the original 50-step model, akhaliq's 4-step LoRA, and this LoRA, along with all generated outputs across 12 resolutions
  • Apache-2.0, same as the base model; includes an online Hugging Face demo

Usage

  • ComfyUI: Place moss_sft_4step_rank_64_lora_comfyui.safetensors in models/loras, use moss_image_sft_bf16 from Comfy-Org/MOSS-Image as the base model, and drag in moss_sft_4step_lora_t2i_local.json; configure euler with Manual Sigmas "1.0, 0.91, 0.717, 0.024, 0.0", CFG 1, and LoRA strength 1.0, without adding negative prompts
  • diffusers: Use the apply_4step_lora function from the model card to attach the standard LoRA to the transformer, with num_inference_steps=4 and guidance_scale=1.0

Risks

  • Only applicable to MOSS-Image SFT; the calibration coefficients for the ComfyUI files were tested on the bf16 base model, and the int8 files from the official template have not been tested
  • Issues listed by the author: Signs not requested to contain text may display garbled characters, occasional extra objects or people appear, and dark scenes are brighter than the original; increasing strength above 1.0 results in double edges rather than more detail
  • All figures are self-tested by the author; the 196 GB listed on the card is peak VRAM during training, while inference VRAM depends on the MOSS-Image base model

Open

抱脸 text-to-image 上升榜

Explore the latest AI Radar updates