MOSS-Image Text to Image | 4-Step Turbo + SeedVR2 2x

▶ Video walkthrough: every switch and the tests behind them. https://youtu.be/CF8pIIlV9-0

Make pictures with MOSS-Image, the new open model that writes the words you give it. Signs, menus, posters and labels come out spelled right when you put the words in "double quotes". Three switches help: Auto prompt writes a full prompt from a few words, Layout boxes puts things and words where you draw them, and Enlarge 2x with SeedVR2 makes the finished picture twice as big and sharper. Your picture redraws one of yours with Img2img, as close to it as you like, or changes the words in it and keeps the rest. A Steps pill picks 50 steps, or 8 or 4 with a speed LoRA for a picture in about two seconds.

How to use it

1. Type your prompt. Put every word you want in the picture in "double quotes".

2. Set the size in Latent Size.

3. Run.

The switches

• Auto prompt. The text encoder MOSS already uses turns a few words of yours into a full prompt, and writes out every word that belongs on a sign, a menu or a poster. What it wrote shows in Show Text. A second or two more, no extra download.

• Layout boxes. Draw a box for each thing or line of words and say what goes in it. Your prompt still describes the whole picture. The canvas starts empty.

• Enlarge 2x with SeedVR2. SeedVR2 3B makes the finished picture twice as big on each side and sharper: 1024 x 1536 becomes 2048 x 3072, about 3 to 4 seconds more on an RTX 5090. It also takes away the softness of 8 and 4 steps. It sharpens a blurred background too, and small blurry writing can get made-up letters. It needs two more files (below); with the switch off they are not loaded.

All three are off as the file opens.

Your picture

• Off: a new picture from your prompt.

• Img2img: load a picture of yours. MOSS starts from your picture and draws over it, so the layout, the pose and the colors stay. The encoder describes your picture, so your prompt is optional, Auto prompt on or off: leave it empty, or type only what should change ("make it a rainy night", "the sign says LAST TRAIN"). Clear the example prompt first: whatever is in the box is taken as a change. Denoise, under the picture, sets how much changes: 0.8 as the file opens. Lower stays closer to your picture; 1 draws a new picture from the description. A slider shows yours beside the new one.

• Change words: load a picture and it shows on the boxes canvas. Draw a box over the words you want to change, pick text and type the new words. MOSS redraws only inside your boxes, and the rest of your picture stays as it was, at its own size. The encoder describes your picture for MOSS, so your prompt is not used. With no box drawn, your picture comes back unchanged. About 4 to 5 seconds at 4 steps on an RTX 5090.

Off is picked as the file opens.

Steps

• 50 steps: the settings OpenMOSS gives, the most detail. About 18 seconds for 1.5 megapixels on an RTX 5090.

• 8 steps or 4 steps: a speed LoRA makes the picture in about 2 seconds or 1.5. It comes out a little softer than 50 steps; Enlarge 2x with SeedVR2 makes it sharp again.

• Only one can be picked. The pill sets the steps and CFG and turns the matching row on in LoRA Loader. If its LoRA is missing, the run stops and names the file.

50 steps is picked as the file opens.

What you need

• ComfyUI 0.42.0 or newer. It is the first release with MOSS-Image (2026-10-18): update ComfyUI first. A git install: git pull, then pip install -r requirements.txt.

• AusBoss nodes 2.16.1 or newer. In ComfyUI-Manager, search "AusBoss". Already have them? Update first.

  Can't update yet? Download v1.2 from this page: the same workflow without Change words. Its second file, MOSS-Image Text to Image (AusBoss, core boxes).json, is for AusBoss nodes 2.13 or 2.14.

• The MOSS-Image files from Comfy-Org/MOSS-Image:

  • moss_image_sft_int8_convrot.safetensors → models/diffusion_models · 5.41 GB

  • qwen3.5_9b_int8_convrot.safetensors → models/text_encoders · 9.75 GB

  • flux2-vae.safetensors → models/vae · 336 MB

• For 8 and 4 steps, the speed LoRAs → models/loras · 1.19 GB each:

  • lora_latest.safetensors by lodestones, for 8 steps

  • MOSS-Image-SFT-4-steps by akhaliq, for 4 steps. It downloads as lora.safetensors: rename it MOSS-Image-SFT-4-steps.safetensors.

• For Enlarge 2x, SeedVR2 from Comfy-Org/SeedVR2:

  • seedvr2_3b_int8_convrot.safetensors → models/diffusion_models · 3.46 GB

  • seedvr2_ema_vae_fp16.safetensors → models/vae · 501 MB

The INT8 file gives the same picture as the full-size one (moss_image_sft_bf16, 10.2 GB), about 2.4 times faster. For a smaller card there is a smaller text encoder, qwen3.5_9b_w4a8_gptq (7.34 GB): pick it in Text encoder.

Tips

• Spell out every word. MOSS writes exactly the words you give it. Words it has to make up come out as gibberish, so put them in quotes, or let Auto prompt write them. Up to about 40 words came out right in my tests; a full page of text is too much.

• Write "a young woman in her twenties" rather than "girl". MOSS draws a child for "girl". Auto prompt does this for you.

• Img2img: lower Denoise to keep more of your picture. 0.6 or 0.7 keeps more of it, faces included; 0.9 changes more. At 8 and 4 steps the same number changes a little more than at 50. If a person still comes out looking different, say how they look in your words: "she has pale skin and freckles".

• Change words: draw the box a little wider than the new words, and type them exactly as they should read. The box's words decide the letters.

• Faint vertical stripes on a smooth sky or wall are a known quirk of this model. MATLOWAI's two DeBand LoRAs together, both at strength 1 in LoRA Loader, removed them in my tests; one alone barely helps.

• Settings: 50 steps at CFG 3 with layer 22 skipped on the negative, the settings OpenMOSS gives. On an RTX 5090 a 1.5 megapixel picture takes about 18 seconds, about 20 with Auto prompt or a picture. 8 and 4 steps run at CFG 1, as the speed LoRAs ask.

• Size: 1 to 2 megapixels works best. A new picture has the size you set in Latent Size. Img2img keeps your picture's shape at about 1 megapixel: set megapixels in the picture's loader for a bigger one. Change words keeps the size of your picture. Enlarge 2x doubles it on each side.

Uncensored

• Auto prompt and Img2img are written by MOSS's own text encoder, from an instruction in the Write the prompt box (double-click it). You can change both. Their only content rule is that every person is an adult.

• If a picture looks censored or tamer than you asked, read what the writer wrote in Show Text. Then change its instruction, or turn Auto prompt off and write the prompt yourself.

• For an uncensored writer there is an abliterated text encoder, Qwen3.5-9B-abliterated_int8_convrot → models/text_encoders · 9.47 GB: pick it in Text encoder. MOSS reads your prompt with it too: in my tests it drew the same pictures, words spelled right.

• The negative prompt is OpenMOSS's, with "nsfw" taken out. It is in the Draw with MOSS-Image box and is used only at 50 steps.

What it does not do

• No free editing. MOSS-Image draws new pictures from words. Img2img draws over yours, so the layout stays and details change. Change words changes only what is inside your boxes: words on a sign, a shirt or a screen.

Credits

MOSS-Image is by the OpenMOSS team, under the Apache 2.0 license. The ComfyUI files are Comfy-Org's. The speed LoRAs are by lodestones and akhaliq, under Apache 2.0. SeedVR2 is by ByteDance Seed, under Apache 2.0; its ComfyUI files are Comfy-Org's. The abliterated text encoder is wangzhang's Qwen3.5-9B-abliterated, converted for ComfyUI by beycanai, under Apache 2.0.

v1.4 (2026-10-21): Img2img, your picture as the start, with a Denoise for how close it stays; an Uncensored note; no "nsfw" in the negative. v1.3 (2026-10-19): Change words, a third choice for your picture. v1.2 (2026-10-19): Enlarge 2x with SeedVR2, and Layout Boxes 🆎 for the boxes, in both views. v1.1 (2026-10-18): the Steps pill, 50, 8 or 4 steps. v1.0 (2026-10-18): first release.

Questions or bugs: GitHub issues or the comments here. Videos on YouTube @ausbossdev. More from me on X @Zanzibased and GitHub.