Draft AI Poster Layouts Faster in ComfyUI: MOSS-Image's 8-Pass Mode Plus a Text Cache for Long Prompts
The longer your layout file, the more time MOSS-Image spends rereading it. A new community add-on stops the rereading. Here is how to set up a drafting workflow around it, and the one case where it makes things slower.
The most detailed prompts are the slowest ones, and with layout models the gap gets big.
MOSS-Image, the open poster model OpenMOSS released on October 16, takes a layout as its prompt. Every headline, price, logo and image area becomes a box with a description and, for text, the exact words to print. A simple title card might be a few lines. A dense event poster with a lineup, a schedule, sponsor logos and fine print can turn into a long document, and the model reads that document again on every pass it makes over the image.
That rereading is where your time goes. On October 19, a developer who publishes as p1atdev posted a fix: a ComfyUI node that reads the text once and reuses it, plus a small correction add-on that cleans up the side effects. On one long poster prompt, the author measured a render dropping from 7.87 seconds to 5.26. That is a third of the wait gone, on exactly the kind of image where waiting hurts.
My position: if you build posters with many boxes, this deserves a place in your setup, but only as a drafting tool. The output changes when you use it, so it belongs in the stage where you are deciding where things go, not the stage where you are deciding what you will deliver.
Why long prompts cost so much
You do not need the maths to make good decisions here, just the shape of it.
An image model like MOSS-Image builds a picture over a series of passes, each one refining the last. On every pass, the picture and the words "look at" each other. Normally that conversation runs in both directions every single time: the words get reread in light of the picture, and the picture gets adjusted in light of the words.
When the prompt is one sentence, that rereading is cheap. When the prompt is a full layout with dozens of fields, the words become a large share of the work. The model spends a growing slice of every pass on text that has not changed.
The p1atdev node makes a trade. It lets the text be read once at the start, stored, and reused while only the picture updates. The catch is that this changes how the model works, so the result drifts from what it would normally make. That is what the correction add-on is for. It was trained to pull the cached version back toward the normal one. The author is careful about how far: in their words, "Small-scale visual checks suggest partial recovery of layout and object consistency." Partial is the honest word.
The other half of the speed: 8 passes instead of 50
The text cache is built to work alongside a second add-on, a fast drafting version from the trainer lodestones (published October 16 as MOSS-Image-SFT-8-steps-dmd, Apache 2.0). Normally MOSS-Image takes 50 passes per image. This add-on teaches it to land on a usable picture in 8.
Stack the two and you change the rhythm of the work. Instead of committing to a layout and waiting, you can try a version, move a box, try again. That is how designers actually work on a comp, and the slow version of the model fights that habit.
Put this into practice
You need ComfyUI on a machine with an NVIDIA graphics card. The node's README says it supports ComfyUI 0.41.0 with native MOSS-Image support. If the text-reader loader in ComfyUI offers a type called moss_image, you are set.
1. Install the node. Clone p1atdev/comfy-moss-image-kv-cache into your custom_nodes folder and restart ComfyUI. The README says no extra packages are needed.
2. Download five files. Into diffusion_models, the MOSS-Image SFT model in its standard BF16 format (the full-precision file, not a shrunken one). Into text_encoders, the Qwen3.5 9B text reader. Into vae, the Flux2 image decoder. Into loras, two add-ons: lodestones' lora_latest.safetensors from MOSS-Image-SFT-8-steps-dmd, and p1atdev's correction file from MOSS-Image-SFT-KV-Cache-LoRA. The correction card pins a specific revision of the lodestones file; use that link if you want to match the author's results exactly. The README says to use the SFT model, so skip Base and community variants.
3. Wire it in this order. Load Diffusion Model, then a LoRA loader with the 8-pass add-on at strength 0.7, then a second LoRA loader with the correction at 1.0, then the MOSS-Image Text KV Cache node with text_timestep at 1000, then your sampler. If you would rather not wire it by hand, the repo includes an example workflow, examples/moss_kv_cal100.json.
4. Set the sampler. Euler, simple schedule, 8 passes, guidance 1, at 1024 by 1024. These are the only settings the author tested. Change them and you are outside what the correction was trained for.
5. Make a second, normal setup for finals. Same layout file, same seed, but with both the correction add-on and the cache node bypassed. This is your comparison and, if you like the drafting result, the route to a delivery render with the full 50-pass model.
6. Draft, then commit. Use the fast setup to settle composition: does the lineup fit, does the headline crowd the image, does the fine print land in the margin you drew. When the layout is right, switch to the normal setup for the version you will actually show.
A useful habit is to keep a note of which layout version you drafted with. Because the cached route changes the output, the same seed will not give you the same picture on the final run. The layout carries over. The exact image does not.
When it makes things slower
This is the part most people will skip and then wonder why ComfyUI feels sluggish.
The author measured two cases on the same graphics card. The long poster prompt, about 2,900 tokens (the small chunks a model reads text in), went from 7.87 seconds to 5.26. A short prompt, 53 tokens, which is roughly a sentence or two, went the other way: from 3.85 seconds to 4.33. The cache costs a little to set up, and on short prompts there is nothing to save.
The README gives no break-even point. So the rule of thumb is simple. Turn it on for dense layouts with lots of text fields. Leave it off for quick single-subject tests. If you are unsure, time both on your own card with your own layout; it takes two renders.
The timings also leave things out. They exclude startup, model loading and the initial reading of the prompt, and they were measured on an RTX 6000 Ada, a 48 GB workstation card. A consumer card will be slower overall, and the cache itself uses some extra memory, so a card that barely fits MOSS-Image could tip over.
Where it breaks
Drafts are not previews. The cached route changes composition, detail and how closely the image follows the prompt. The author says so directly. A draft you love may not survive the switch to the normal setup.
It was trained on almost nothing. The correction add-on learned from 10 prompts the author wrote, checked against 3 more, with no outside images. The card's own line: "Quality evaluation is limited; this is not a general quality guarantee."
The settings are narrow. Only the standard BF16 model files and an 8-pass add-on strength of 0.7 were tested. Other strengths, pass counts, smaller model files or other fast add-ons fall outside the recipe. If you swap in a different drafting add-on, results are a guess.
Loading the add-on alone does nothing. The README is blunt about it: without the cache node, the correction file does not turn on caching. People will download one file, see no change and give up.
It is a day old. The node, the correction and the 8-pass add-on underneath it are all community work from the last four days. Expect changes, and check the pages before you build a client job on today's version.
What this asks of you
The real gain here is not a few seconds per image. It is permission to work the way you already work on a comp: try, move, try again, without a 50-pass wait turning each idea into a commitment.
That only pays off if you keep the two stages honest. Draft fast, judge layout only, then render the final the slow way and judge that on its own. Set up both routes in ComfyUI this week, time them on your densest real layout, and keep the fast one only if it saves you more than it costs. Your card and your posters will give you a better answer than any chart.