Skip to content
Hot eventOngoing

lvladikov Releases the MOSS 4-Step Distilled LoRA

1 report1 sourceUpdated 23 hours ago
Subscribe to this event via RSS

Start here

AI overview

lvladikov released the MOSS-Image-SFT 4-step distilled LoRA and a ComfyUI workflow. The LoRA is trained with eDMD and LADD. The author says 22 1024² images took 26.6 seconds, while the 50-step baseline took 387.3 seconds, and fine texture averaged 1.22 times that of the 50-step teacher. When using it, load the _comfyui file, set strength to 1.0, 4 steps, and CFG 1, and do not carry over the 50-step teacher's layer skipping or negative prompt settings. It was trained across 12 resolutions, from 768×512 and 512×768 to 1824×1024 and 1024×1824; other aspect ratios are not in the training set.

Written by AI from the reports · updated 1 hour ago

Report timeline

Follow the reports to see the event from other sides.

Nov 19
  1. ComfyUI Wiki · 工作流与开源创作分享Picks
    ComfyUI Usage and Measured Comparison of the MOSS-Image-SFT 4-Step Distilled LoRA

    lvladikov's MOSS-Image-SFT 4-step distilled LoRA is trained with eDMD and LADD. In the author's tests, 22 1024² images took 26.6 seconds, compared with 387.3 seconds for the 50-step baseline, and fine texture averaged 1.22 times that of the 50-step teacher.

Heat trend for this event

Not enough consecutive observations to draw a trend yet.