---
title: How to Run MOSS-Image 5B Locally in ComfyUI (VRAM, Setup)
url: "https://goremy.ai/blog/articles/moss-image-5b-comfyui-local-install"
author: Luis Chavez-Mattos
editor: Luis Chavez-Mattos
editor_title: Director, Product
published: "2026-11-03T15:00:09.884Z"
updated: "2026-11-03T15:00:12.457Z"
topic: Open Model Run Guides
description: "Install the 5B MOSS-Image text-to-image model in ComfyUI: where the files go, the node setting that matters, about 28 GB VRAM use and layout mode."
---

# How to Run MOSS-Image 5B Locally in ComfyUI (VRAM, Setup)

> Install the 5B MOSS-Image text-to-image model in ComfyUI: where the files go, the node setting that matters, about 28 GB VRAM use and layout mode.

Edited by Luis Chavez-Mattos, Director, Product · Published 2026-11-03

## How do you run MOSS-Image locally in ComfyUI?

Install the MOSS-Image plugin through ComfyUI Manager, download the model files into the matching ComfyUI model folders, load the example workflow, and turn on the newer node mode in the ComfyUI settings. Expect roughly 28 GB of VRAM in use with the SFT model, based on one hands-on test. That puts it on a 32 GB or 40 GB class NVIDIA card, not a typical 12 or 16 GB gaming GPU.

## TL;DR

- MOSS-Image is a **5 billion parameter text-to-image model** from OpenMOSS, released under Apache 2.0, with a FLUX.2 VAE and a Qwen-family text encoder.
- The project says ComfyUI support comes as a **native plugin in ComfyUI Manager**, and that merging into the official ComfyUI repository is still in progress.
- A local run of the SFT model in ComfyUI showed **about 28 GB of VRAM** in use.
- The new node mode in ComfyUI has to be on. In testing, leaving it off gave **blurry output and no bounding box editing**.
- Layout mode lets you place objects with bounding boxes, but it is **labeled experimental** and was patchy in testing, with one test flipping box positions left to right.
- Text rendering is advertised for English and Chinese, yet mixed-text layout tests showed **hit-and-miss results**.

## What is MOSS-Image?

MOSS-Image is a family of text-to-image models from OpenMOSS. There are two checkpoints: MOSS-Image-Base and MOSS-Image-SFT. The SFT model is a supervised fine-tune of Base with the same architecture and inference pipeline. The model card says SFT produces clearer text grouping and more organized layouts in its whiteboard comparison.

Both accept three prompt styles: plain natural language, tag sequences, and JSON layouts that control text placement and object arrangement. The weights and bundled inference code use the Apache 2.0 license, so commercial use is permitted under that license's terms. The model is known for rendering clean English and Chinese text, which is where it is meant to stand out.

## What hardware does MOSS-Image need?

The model card says to use an NVIDIA GPU with CUDA support. It does not publish a minimum VRAM figure. The only number from hands-on use is the roughly 28 GB seen while running MOSS-Image-SFT in ComfyUI.

| Item | Detail |
| --- | --- |
| GPU | NVIDIA with CUDA |
| Observed VRAM (ComfyUI, SFT) | About 28 GB |
| Python (script route) | 3.10 or later |
| PyTorch (script route) | 2.9 or later, matched to your CUDA build |
| Precision | Transformer and text encoder in bfloat16, VAE and initial noise in float32 |

The 28 GB figure includes the transformer, text encoder and VAE, so it is not a pure weight size. The project's quick start mentions CPU offload and compilation options for the Python and Gradio route. Those may lower the memory ceiling at the cost of speed, but the sources give no numbers, so test on your own card.

## How do you install MOSS-Image in ComfyUI?

You need a working ComfyUI install first. Then:

1. Open ComfyUI Manager and search for **MOSS-Image** to install the native plugin.
2. Download the model files from the Hugging Face model card. Open the Files tab, right-click the download arrow, and choose "save link as."
3. Save each file into the matching folder under your ComfyUI `models` directory.
4. Load the example workflow from the project's ComfyUI folder on GitHub.
5. Check the node setting described below, then queue a prompt.

Three files matter: the diffusion model checkpoint (Base or SFT), the FLUX.2 VAE, and the Qwen text encoder. The exact subfolder for each follows a file pattern in the project's ComfyUI README. ComfyUI normally separates diffusion models, text encoders and VAEs into their own folders, but use the README's layout rather than guessing. The README and the example workflow JSON are the authority here, since the integration is still moving.

## Which ComfyUI setting do you have to change?

Open the ComfyUI menu at the top left and make sure the newer node mode is enabled. In the video this was called "node two," which most likely means ComfyUI's Nodes 2.0 interface. Check the exact label in your version.

This is the step people will miss. With it off, the tester reported blurry results and found that the bounding box drawing tool would not work at all. If your first images look soft, check this before you blame the model.

## What are the default sampling settings?

The model card lists these defaults, and the example workflow should be close to them.

| Setting | Default |
| --- | --- |
| Sampler | Flow-matching Euler |
| Steps | 50 |
| CFG | 3 |
| Shift | 5 |
| Layer skipping | Negative branch, layer index 22 (zero-based) |
| Negative prompt | Empty |

Default guidance uses two transformer forwards per step, so 50 steps means 100 forwards. That is why generation is heavier than the parameter count suggests. If you use the Python API instead of ComfyUI, image dimensions must be positive multiples of 16. Sticking to multiples of 16 in ComfyUI is a sensible habit too. For 8-step LoRAs and other runtimes, see this [guide to running Moss Image locally](https://goremy.ai/blog/articles/run-moss-image-locally-comfyui-sglang-loras).

## How does layout mode work?

Layout mode is the feature that sets MOSS-Image apart. Instead of describing positions in prose, you give the model a JSON layout with a scene caption and a list of items. Bounding boxes use the `yxyx` format in an absolute centralized coordinate space. In the Python route you serialize the JSON and pass it as the prompt. In ComfyUI you draw boxes with the mouse and attach a description (or text) to each one.

In the video test, three boxes were drawn: a five-story pagoda at top left, a woman in an indigo silk kimono at center, and a stone lantern at bottom right. The result was a decent image, but the horizontal placement came out flipped relative to what was drawn. The pagoda also did not clearly have five stories. A second test with four boxes and mixed Chinese and English signage placed the puppet and the satay cart correctly. The Chinese signboard landed top right instead of top center, and the requested English text did not appear.

The model card labels layout prompts as experimental, and these results fit that. Treat layout as a way to get rough composition, not exact placement.

## How good is the output quality?

Single-prompt scenes were the strong point. A Swiss lake scene with a meadow, manor, rowboat and Alps came out sharp from foreground to background. Close-up portraits, such as a Huli wigman in a cloud forest, and a Bohemian glassblower scene were also strong. The tester found the model good at faces and outdoor scenery, with occasional blur. A macro flower image looked more like a watercolor than a photo.

Text was weaker than the billing suggests. An Urdu Nasta'liq prompt on a truck art panel got the script style right but the words wrong. That is expected, since the model targets English and Chinese. The mixed-text layout test missed even the English. Verify any text-heavy output rather than assuming it works.

## Is running MOSS-Image locally worth it?

It depends on your hardware. If you already have a 32 GB or larger NVIDIA card, a 5B model with Apache 2.0 licensing, solid realism and an experimental layout tool is worth an afternoon. If you have a 16 GB card, try the offload options first, and expect slower runs. Rented GPUs are the other route, though the sources give no pricing.

The ComfyUI integration is also not yet in the official repository, so expect changes. Pin your plugin version once you have a working setup.

## Frequently Asked Questions

### How much VRAM does MOSS-Image use in ComfyUI?

One local test of MOSS-Image-SFT in ComfyUI showed around 28 GB of VRAM in use. The model card gives no official minimum. Offload and compilation options exist in the project's quick start, but their savings are not documented in the sources.

### What is the difference between MOSS-Image-Base and MOSS-Image-SFT?

SFT is a supervised fine-tune of Base with the same architecture and inference pipeline. According to the model card, SFT gives clearer text grouping and more organized layouts in its comparison example. The video tested the SFT version.

### Why are my MOSS-Image results blurry in ComfyUI?

The most likely cause is the newer node mode being switched off in the ComfyUI menu. In testing, leaving it off gave blurry images and disabled bounding box creation. Also confirm that the VAE and text encoder are in the correct folders and that you are using the example workflow.

### Can MOSS-Image render text?

It is built for clean English and Chinese text and can follow a text field in a layout item. Results were inconsistent in testing, especially when text was combined with layout boxes. Other scripts, such as Urdu, are not supported and came out wrong.

### What license does MOSS-Image use?

The model weights and bundled inference code are released under Apache License 2.0, according to the model card.