LIVE WIRE Ubuntu vs Windows 11 for AI Image Generation in 2026: A Careful ComfyUI Comparison
[ SYSTEM DISPATCH // V.4.2 ]
AI INTELLIGENCE & GENERATIVE SYSTEMS
[ FILED UNDER // COMFYUI ]

MOSS Image Comes to ComfyUI: Fast, Open-Source Generation with Turbo, References & Editing

AIwithWalt October 26, 2026 6 min read

MOSS Image is now supported in ComfyUI, giving home users another free and open-source text-to-image model under the Apache 2.0 license. In this 12:26 walkthrough, Nerdy Rodent demonstrates a full modular workflow covering fast turbo generation, high-resolution output, references, bounding-box layouts, editing, and ControlNet.

A Proper FOSS Model for Local Use

The video opens by framing why MOSS Image matters: unlike some recent releases with restrictive licenses, MOSS Image is Apache 2.0, so it can be run freely at home on your own computer. The model is small, supports layouts for generations, handles a variety of image styles, and can render text in both English and Chinese.

Rodent presents his workflow using his “Rodent method” — neat, colorful, modular groups that are easy to understand and update. Viewers can pause for screenshots or rewatch sections to rebuild it.

Still from MOSS Image - With Control, References & Editing in ComfyUI!
Still from “MOSS Image – With Control, References & Editing in ComfyUI!”.

Model Loading: INT8, Attention Backend, and LoRAs

The workflow starts with model loading. Rodent uses the INT8 versions, each under 8GB, which fit comfortably in the VRAM of most cards including his RTX 3090. He enables the model attention backend for speed and mentions optional skip-layer guidance to help image quality.

What extends MOSS Image well beyond the base release is community LoRAs:

  • Standard text-to-image with Base or SFT versions
  • Turbo 4-step LoRA for lightning-fast generation
  • High-res LoRA for 4K+ images
  • Edit LoRA for image editing
  • Color / Depth Control LoRA

Image resolution defaults to typical 1-megapixel sizes, but with the high-res LoRA it can be pushed to 8 megapixels or more. A single image loader feeds references, image-to-image, inpainting, and editing.

Still from MOSS Image - With Control, References & Editing in ComfyUI!
Still from “MOSS Image – With Control, References & Editing in ComfyUI!”.

Prompting Options: Enhancer, References, and Bounding Boxes

Prompting is highly flexible. Options include a prompt enhancer, reference-image prompting with image weight and vision megapixels settings, bounding-box prompting, and image-to-text where the enhancer can look at an input image.

In non-turbo mode there is also a negative prompt and support for weighted prompts using word:1.2 or negative values. The KSampler is standard, but Rodent adds a switch to easily toggle turbo mode on and off.

Normal mode uses around CFG 2.5 and 25 steps. Turbo mode drops to CFG 1.0 and just 4 steps. He shows a text-rendering test borrowed from a Ming Image prompt, and even at 1024×1024 the English text is almost fully readable, which he rates as very impressive.

Still from MOSS Image - With Control, References & Editing in ComfyUI!
Still from “MOSS Image – With Control, References & Editing in ComfyUI!”.

High-Resolution Generation and Turbo Mode

Rodent tests 8 megapixels — 3856 x 2176 — with a prompt for a rodent spaceship captain on the bridge of a futuristic space cruiser. Without the high-res LoRA, the result has good fur quality and a nice monocle, but duplicates the character on the left.

Loading moss high res lora test v04 with otherwise identical settings fixes the duplication, producing a single captain. He then combines the high-res LoRA with the moss image SFT 4-step turbo LoRA. At high resolution he finds increasing turbo steps to 6, with slightly different sampler and scheduler settings, helps quality. The result is faster and slightly more comic-like than realistic, but very strong for that resolution.

He then drops back to 1 megapixel with turbo on and the prompt enhancer enabled to test styles. The enhancer group shows a preview of the expanded prompt, with the full system prompt available from the official MOSS Image prompt enhancer skill. Results include:

  • A geometric, painting-like cubism test
  • A well-drawn anime scene of an ugly curious cat in a French cake shop, with “patisserie” added in the background by the enhancer
  • A charcoal-style house with the phrase “the broken door swings wide”
Still from MOSS Image - With Control, References & Editing in ComfyUI!
Still from “MOSS Image – With Control, References & Editing in ComfyUI!”.

Reference Images and Layout Control

Using reference-image prompting, he feeds in the ComfyUI example character and asks for a charcoal drawing of a colorful character next to a terribly drawn house. The output clearly takes influence from the reference — large ears and lots of color — while keeping a charcoal feel.

With an empty prompt and only a reference image, MOSS produces a very similar variation, making it a useful way to generate variants.

For positional control, he uses the MOSS Image Prompt Builder from KJNODES rather than the stock Nodes 2.0 layout tool. It provides a prompt section plus an editor where you can draw boxes for object, character, or text. In 4-step turbo mode, his test places a rodent with text at the top and spooky Halloween pumpkins at the bottom, and the layout is respected.

Editing Is Experimental, ControlNet Works Well

Image editing is possible with the edit LoRA, using by default the AI Toolkit version of MOSS Image Edit. Rodent loads arm v5 moss edit plus an SFT LoRA blocks 1 version to reduce striping seen with the SFT model. With a stylized portrait and the prompt “change her hair color to bleach blonde and her expression to happy,” the edit partly works — hair becomes blonder and the face becomes happier — but it replaces the origami geometric mouth with a realistic mouth and shifts all colors significantly. His verdict: very experimental at the moment.

ControlNet is more convincing. He loads the MOSS color and depth control LoRA, again with SFT LoRA blocks 21 to 49 to remove striping. Using an alien creature in sand as input, he generates a depth map with Depth Anything V3 Mono Large, feeds it into the native MOSS Image Color Stroke Depth Control node, and prompts for “a pencil drawing of a weird alien at home.” Shown in an image-compare node, depth is preserved well while converting to pencil sketch.

Upscaling: Latent Upscale and SeedVR2

Even though large images can be generated directly, Rodent shows two upscaling paths. A typical image upscaler at 2x with ~0.25 denoise produces a 2048 x 1664 result with little change but higher resolution in a few steps.

SeedVR2 with a 4x scale multiplier produces a much larger 4096 x 3328 image. In comparison, SeedVR keeps the image very similar but sharpens it considerably, clearing up blur from the original.

He closes by noting the workflow plumbing: image, latent, and masked latent groups plus switches to auto-select latent, prompt priority — AI prompt first, then bounding box, then manual prompt — and conditioning switches with Conditioning Switch 2 applying ControlNet if enabled.

Key takeaways

  • MOSS Image is Apache 2.0 open-source, small, and ComfyUI-ready, with INT8 checkpoints under 8GB suitable for cards like a 3090.
  • Normal generation is ~25 steps at CFG 2.5; turbo LoRA drops to 4 steps at CFG 1.0 for seconds-fast output, with 6 steps helpful at high resolution.
  • The high-res LoRA enables 8+ megapixel images and fixes duplication issues seen at large sizes.
  • Prompt enhancer, reference images, and KJNodes bounding-box layouts add strong stylistic and compositional control, plus good English/Chinese text rendering.
  • Image editing works but is experimental and shifts colors; Depth/Color ControlNet and SeedVR2 upscaling give more reliable results.

Who it is for

This video is for ComfyUI users who want a fast, locally-run open model with modern extras. If you have a mid-range GPU, like building modular workflows, and want turbo-speed drafts, high-res outputs, image references, layout control, and early editing and ControlNet support without restrictive licensing, this walkthrough is directly relevant.

Watch it

Full demonstration and workflow breakdown in the original video: MOSS Image – With Control, References & Editing in ComfyUI! by Nerdy Rodent.

Leave a Reply

Your email address will not be published. Required fields are marked *

>_ JOIN THE INTELLIGENCE NETWORK

One email when a new breakdown drops: the workflow, the settings, and what actually worked. No spam, unsubscribe anytime.

Join the Intelligence Network

New AI video & image breakdowns, straight to your inbox. Free, no spam.