Training a LoRA for MossImage: Full Walkthrough
MOSS-Image Base is the checkpoint to fine-tune. This page collects what OpenMOSS has actually published about LoRA work, how to caption text-heavy images, and which parts of the recipe you still have to tune yourself.
Read this before you start
As of 2026-10-20, OpenMOSS has not published a LoRA training script or default hyperparameters. There is no official rank, alpha, learning rate, batch size or dataset manifest. This page does not invent them. It separates verified guidance from the choices you must make and measure on your own hardware.
01 / Start from Base
Why not SFT?
The official guide is explicit: Base is the model intended for LoRA work.
MOSS-Image-Base has not undergone post-training or inference parameter distillation such as CFG. OpenMOSS says the goal is to preserve the training information in full and make it easier for the community to train LoRAs. SFT is a curated fine-tune for finished images, so it is the wrong starting point if you want to add a new style.
The second rule from the official guide: keep the text encoder frozen. Do not fine-tune the Qwen3.5-9B encoder.
02 / Published vs not
What the official documentation covers
| Item | Published by OpenMOSS? | What to do |
|---|---|---|
| Base recommendation | Yes | Fine-tune Base, not SFT. |
| Text encoder | Yes: freeze it | Do not train the Qwen encoder. |
| Captioning rules | Yes | Preserve visible text exactly; describe layout. |
| Caption schema | Yes | JSON with long/medium/short captions, watermark boxes and border flags. |
| Training script | No | Choose your own trainer and log the configuration. |
| Rank / alpha / learning rate | No | Tune on your dataset; there are no official defaults. |
| Dataset format | No | A folder of images plus matching captions is the common baseline. |
| Training recipe | Announced, not released | Watch the official repository for the release. |
03 / Data
Caption text-heavy images correctly
Text rendering is the reason to use MOSS-Image. A caption that paraphrases the words teaches the wrong thing.
- Preserve every visible word exactly — no translation, paraphrase, romanization or normalisation.
- Describe the layout: line breaks, reading order, relative position, alignment, size hierarchy and spacing.
- Describe the type: serif/sans, weight, colour, texture and whether the text is printed, handwritten or painted.
- Record watermarks with normalised 0–1000 boxes in
[y_min, x_min, y_max, x_max]order, and record borders as left/right/up/down or none. - The official guide suggests a strong vision-language model for captioning; it names Gemma 4 and Qwen 3.5 27B/35B-A3B as examples.
{
"caption": {
"long": "A detailed description including every visible word.",
"medium": "A shorter description that still preserves the text.",
"short": "A one-line summary."
},
"watermark": [
{ "box_2d": [0, 0, 0, 0], "label": "watermark text" }
],
"borders": ["none"]
}04 / What you tune
The missing hyperparameters
There is no official LoRA recipe yet. These are the choices you must make and record.
Rank, alpha, learning rate, batch size, steps, resolution, caption dropout and the number of training images all change the result. Start from the defaults of your trainer, keep a fixed seed and evaluation prompt set, and compare runs side-by-side. Do not copy a recipe from a different model family and assume it transfers.
For evaluation, use prompts with long text, mixed Chinese/English strings and a layout. A style LoRA that improves portraits but destroys glyphs is not a good MOSS-Image LoRA.
05 / Inference
Load the LoRA in ComfyUI
The official ComfyUI README says the built-in LoRA loader works for LoRAs whose model keys match. Use the official MOSS-Image node and workflow, then insert a standard LoRA loader between the model and the sampler. For SLG or block patches, use eager execution because TorchCompileModel filters transformer-option guards.
Follow the ComfyUI workflow guide for the loader path and sampler settings.
06 / Community
Built a LoRA?
We are collecting MOSS-Image LoRA experiments for a future showcase. Send the model link, the base checkpoint, and one before/after prompt.
Email support@getmossimage.com with the subject “MOSS-Image LoRA submission”. Include the dataset size, the text-encoder state, and whether the LoRA keeps glyphs legible. We will only publish results we can reproduce.
07 / FAQ
LoRA FAQ
Does OpenMOSS provide a LoRA training script?
No. As of 2026-10-20 the repository has no training or LoRA directory, and the user guide does not publish rank, alpha, learning rate, batch size or dataset format. The project says a training recipe is forthcoming.
Which checkpoint should I start from?
Start from MOSS-Image-Base. The official guide says Base has not undergone post-training or CFG distillation, which preserves the training information and makes it the intended starting point for LoRA work.
Should I fine-tune the text encoder?
No. The official user guide says to keep the text encoder frozen when you post-train MOSS-Image-Base.
How should I caption images that contain text?
Preserve the visible text exactly. Do not translate, paraphrase, romanize or normalize it. Describe line breaks, ordering, alignment, size hierarchy, font style, color and the arrangement of text blocks. The official guide includes a JSON caption schema with long, medium and short captions plus watermark boxes and border flags.
Can ComfyUI load a LoRA for MOSS-Image?
Yes, for LoRAs whose model keys match. The official ComfyUI README says the built-in LoRA loader works for matching keys, and recommends eager execution when using SLG or changing block patches.
Next