How to Run MOSS-Image 5B Locally with ComfyUI

What is MOSS-Image and why does it matter?

MOSS-Image is an open-weight text-to-image model with 5 billion parameters, built around a Flux-style variational autoencoder and a Qwen 3 text encoder (3.59 billion parameters) for prompt understanding. It runs locally through ComfyUI, renders both English and Chinese text in generated images, and supports a JSON-based layout mode that lets you place specific objects inside bounding boxes before generation. It’s released under an open license, which means anyone with a capable GPU can download and run it without an API key or subscription.

That combination matters because most image models either generate decent scenes with no control over composition, or they let you sketch rough layouts but produce weak detail. MOSS-Image tries to do both: photorealistic output plus explicit spatial control over where objects land in the frame.

TL;DR

MOSS-Image is a 5B-parameter open-weight text-to-image model that pairs a Flux-based VAE with a Qwen 3.59B text encoder, released under an open license for local use.
It needs around 28GB of VRAM when run locally through ComfyUI, so it targets high-end consumer or prosumer GPUs, not laptops.
Installation follows the standard ComfyUI pattern: download the model files from the Hugging Face model card and drop them into the matching subfolders inside your ComfyUI models directory.
A JSON layout mode lets you draw bounding boxes and assign text descriptions to each region, giving rough control over where a pagoda, a figure, or a prop appears in the final image.
Text rendering only officially supports English and Chinese, and even then it’s inconsistent. Other scripts, like Urdu Nasta’liq tested in one demo, come out visually close but not actually legible.
Layout control is still patchy: bounding box positions sometimes flip or shift, and signage text placed inside boxes often lands in the wrong spot or renders incorrectly.
General scene generation, faces, and landscapes are a strength, with testing across dozens of prompts (from Swiss lakes to Papua New Guinea portraits) producing consistently sharp, detailed results without retouching.

Every team builds with AI.
Nothing gets lost.

How do you install MOSS-Image in ComfyUI?

The installation follows ComfyUI’s normal custom-model workflow, no special installer required.

Make sure ComfyUI is already installed. If you haven’t set it up yet, that’s a prerequisite step on its own, covered in general ComfyUI installation tutorials.
Open the MOSS-Image model card on Hugging Face and go to its Files tab.
Download each required file (the diffusion/UNet weights, the VAE, and the text encoder) by right-clicking the download arrow next to each file and choosing “save link as.”
Place each file in its matching ComfyUI subfolder. ComfyUI expects models sorted by type, so the text encoder goes into the text encoder folder, the VAE into the VAE folder, and so on, all nested under your ComfyUI installation’s models directory.
Load a MOSS-Image workflow (a ComfyUI graph with the correct nodes already wired up) rather than building one from scratch, since the model depends on specific node configurations to behave correctly.
Enable the relevant node toggle in the workflow before generating. In testing, leaving this node switched off produced blurry results and also disabled the bounding-box layout feature entirely.

Once the files are in place and the workflow loads cleanly, generation works like any other ComfyUI text-to-image pipeline: type a prompt, hit queue, wait for the sampler to finish.

How much VRAM does MOSS-Image need?

Running MOSS-Image locally consumed approximately 28GB of VRAM in testing. That puts it out of reach of most consumer cards with 8GB, 12GB, or even 16GB of memory. It’s realistically a model for 24GB+ cards pushed to their limit, or for 32GB+ professional and workstation GPUs. Anyone without that kind of hardware on hand can rent a cloud GPU by the hour instead of buying one outright, which is a common workaround for testing models at this size before committing to local hardware.

This VRAM footprint comes from running the full pipeline together: the diffusion model, the Flux-based VAE, and the 3.59B-parameter Qwen text encoder all loaded at once. There’s no official lighter quantized version mentioned, so lower-VRAM setups would need to wait for community-made quantized builds, a common pattern with other open image models after their initial release.

How good is the layout control feature?

MOSS-Image’s standout feature is JSON layout mode, where you draw bounding boxes directly on the canvas and attach a text description to each one, for example “a five-story pagoda” in the top-left box, “a woman in a deep indigo silk kimono” in the center, and “an ancient stone lantern” in the bottom-right.

Build apps with AI that pass the audit.

In practice, results were a mixed bag. Object placement roughly followed the intended regions, but boxes sometimes flipped left to right, and in one test with four boxes (a Chinatown street scene with English and Chinese signage), a puppet and a food cart landed correctly while a Chinese signboard ended up in the wrong position with incorrect characters. The feature clearly works as a general steering mechanism rather than a precise placement tool. It nudges the model toward a rough composition, but it doesn’t guarantee pixel-accurate object placement or correct embedded text.

Is MOSS-Image worth running locally?

For raw scene generation, the results are strong. Tests across a wide range of prompts, from a Swiss alpine lake scene to a Siberian reindeer-herder night shot to a close-up portrait of a Huli wigman in Papua New Guinea, consistently produced sharp, well-lit, detailed images without any retouching or multiple attempts. Faces, landscapes, lighting, and texture all held up well across very different cultural and geographic settings.

Where it falls short is text rendering and precise layout control, the two features that make MOSS-Image distinct from a plain text-to-image model. Text only works reliably in English and Chinese, and even then it’s inconsistent (a Chinese signboard test came out with the wrong characters and in the wrong position). Non-Latin, non-Chinese scripts like Urdu came out visually plausible but not actually correct, which is expected since the model wasn’t trained for that. Layout boxes guide composition in the right general direction but shouldn’t be treated as exact placement commands.

If the goal is high-quality general image generation on hardware with 24 to 32GB of VRAM, MOSS-Image is a reasonable open-weight option to test locally. If the goal is reliable embedded text or exact object placement for production design work, it’s not there yet.

Frequently Asked Questions

What license is MOSS-Image released under?

MOSS-Image is distributed as an open-weight model, which allows local download and use without an API or subscription. Check the model’s Hugging Face card for the exact license terms before any commercial use.

Does MOSS-Image support languages other than English and Chinese?

Officially, text rendering is designed for English and Chinese only. Testing with other scripts, such as Urdu Nasta’liq, produced text that looked stylistically plausible but wasn’t accurate, which lines up with the model’s stated language scope.

Can I run MOSS-Image without a high-end GPU?

Not easily in its current form. Local testing showed around 28GB of VRAM usage, which requires a high-end consumer or professional GPU. Renting a cloud GPU is a practical alternative for anyone without that hardware locally.

How accurate is the bounding-box layout feature?

It’s approximate rather than precise. Objects generally appear near their intended region, but positions can flip or shift, and text placed inside layout boxes (like signage) often renders incorrectly or lands in the wrong spot.

What’s the best use case for MOSS-Image right now?

General-purpose photorealistic scene and portrait generation. It handles landscapes, lighting, and human faces well across varied cultural and geographic prompts, making it a solid local option for creative image generation where exact text or object placement isn’t required.