MOSS-Image in ComfyUI: workflow and model comparison

This tutorial tests MOSS-Image for text inside generated pictures and explains the creator's ComfyUI workflow. The speaker describes a 5 billion parameter model with an Apache 2.0 license. Base targets training; the demonstrated workflow uses SFT. The supplied metadata specifies ComfyUI v0.42.0 or newer.

The examples include business cards, menus and Chinese lettering. The speaker reports near-perfect spelling up to about 40 words, with longer book-page text starting to scramble. Layout boxes control object placement and text size. A separate editing workflow redraws a selected box while, according to the demonstration, preserving everything outside it pixel for pixel.

Four switches cover automatic prompt expansion, starting from a picture, layout boxes and SeedVR2 2x upscaling. Prompt expansion uses the text encoder already loaded. Generation options include 50 steps, eight steps and four steps with a speed LoRA. The creator reports roughly two seconds per four-step image on an RTX 5090 with the model loaded. For local AI use, the stated minimum is 8 GB VRAM with 16 GB RAM; the speaker recommends INT8 files for cards with 24 GB VRAM or less.

Comparisons with Qwen Image 2.1 and FLUX 3 Klein cover lettering, characters and poses. MOSS earned both the most favorites and the most rejects in the creator's blind grading. Reported flaws include anatomy errors and faint SFT background stripes; the speaker says two corrective LoRAs are needed for the stripe fix.