Suppose you start ComfyUI with --fp16-unet because your GPU lacks bf16, load MOSS-Image, and ask for a big poster. The progress bar finishes and the preview is solid black, or a field of noise. ComfyUI Models, an unofficial guide kept separate from Comfy Org, OpenMOSS and Qwen and free to use, has an entry for exactly this.
The problem is numeric. MOSS-Image produces values that sit near the upper limit of fp16, so very large images or added LoRAs overflow it. With the official weights, the guide notes, the first black image shows up at 2560×2560. Bigger canvases and extra LoRAs are the two triggers it names.
The solution has three branches. If your card handles bf16, drop the fp16 flag and run the default bf16. If it does not, download KBlueLeaf's FP16-fixed weights, which the guide suggests for anyone who needs fp16. If you are using a 4- or 8-step rapid-moss adapter to speed things up, update the rapid-moss node. The faster generation page explains that no official Turbo exists yet and describes community options from 4 to 14 steps as well as Spectrum.
Resolution choices matter as well. The OpenMOSS reference suggests 1 to 2 megapixels with sides in multiples of 16, and the ComfyUI templates default to 1 megapixel, far below the size where overflow began in the guide's example. The file picker on the same page suggests the bf16 diffusion model, 10.16 GB, for 16 GB cards and up, and the int8 model, 5.41 GB, for smaller ones. Both are listed with the folder they go in, ComfyUI/models/diffusion_models/, and a download command.
The guide has limits. It collects fixes from Comfy-Org and OpenMOSS discussions and issues and links each one, reporting what those sources say, and official VRAM requirements are not published. Reading is free and needs no account. The black image entry and the rest of the errors are on https://comfyuimodels.com/.
The full troubleshooting list is on ComfyUI Models.
