← sources

OpenMOSS released MOSS-Image, and it’s now supported in ComfyUI.

source: reddit (reddit:comfyui)published 2026-10-19 19:44ingested 2026-10-20 02:39

retained text

**What it is:**

* 5B DiT, using a Qwen3.5-9B text encoder and the FLUX.2 VAE
* Two versions: **Base**, and **SFT**, which is fine-tuned on curated data
* Apache 2.0

**Why it’s worth trying:**

* **Text rendering.** Legible text in English and Chinese, which most open models still struggle with.
* **Layout prompts.** You can pass structured JSON with bounding boxes to control where text and objects go. It’s useful for posters, ads, and anything with a real layout.
* **Aesthetics.** The outputs look good out of the box without much prompt wrangling.

**Settings to start with:** 1024×1024, 50 steps, CFG 3, shift 5. Neither model is step-distilled, so it’s not fast.

**VRAM:** The full BF16 files are big (\~10 GB DiT plus \~17 GB text encoder). The community already has INT8 and INT4 quants out, and people are reporting it running on 8 GB cards.

Run the workflow: [https://huggingface.co/Comfy-Org/MOSS-Image](https://huggingface.co/Comfy-Org/MOSS-Image)

document fetched 2026-10-20 00:41 · canonical https://reddit.com/r/comfyui/comments/1xa6wmo/openmoss_released_mossimage_and_its_now_supported