⚠️ Experimental – first version. MOSS-Image is new and this kit was built and tested in a few days. Expect bugs and rough edges. Feedback and bug reports in the comments are very welcome.

🖥️ Now available: Fumika, the standalone app. The same MOSS trainer – plus trainers for Anima (1.0 / 2.9B / 3.8B), SDXL and sliders – in a free desktop app for Windows and Linux: one installer, models downloaded and managed for you, a dataset check with memory estimate before every run. No nodes, no workflows. GitHub · Civitai article. This ComfyUI kit stays available if you prefer ComfyUI.

Train MOSS-Image LoRAs (characters, outfits, styles) from inside ComfyUI, with no command line. Then pick the best epoch from one grid.

How it works (easy mode)

1. Train a LoRA (Templates → ComfyUI-MOSS-LoRA-Trainer):

1. Make a folder, and inside it a folder named 4_mychar with your images (20–100). Next to every image goes a .txt file with the same name and one sentence describing the image, always starting the same way: Safe image. ohwx, an adult woman with long black hair, red eyes, white tank top. She stands in a park and smiles.

2. Enter the folder and a name, then press Run. A report tells you whether the dataset is OK and how long training takes.

3. Switch start on and press Run again. A console window opens and trains, about 1 h on an RTX 5090. First time only: the trainer installs itself (5–15 min) into its own folder; it does not touch your ComfyUI Python.

4. Press Run any time to see the progress.

2. Test my LoRA: enter the same name and press Run. You get a grid with one row per saved file, plus a row without the LoRA. Take the earliest row where the character looks right. The test prompts are built automatically from your captions.

Settings

The tested recipe from our released MOSS LoRAs:

• rank 32 / alpha 16

• AdamW8bit, learning rate 1e-4, cosine schedule

• logit-normal timesteps

• text encoder frozen and cached

• bf16 at 1024 px

The number of epochs is calculated from your dataset size. Optional fields let you change:

• base model, text encoder, VAE

• epochs

• rank

• learning rate

• resolution

Captions: the one thing that matters

MOSS reads sentences, not tag lists. Use the same trigger word and the same character description in every caption, and the same sentence later when you prompt. Already have a Danbooru-tag dataset (Anima / Illustrious)? The advanced workflow contains a node that converts tags into sentence captions.

Requirements

• ComfyUI with MOSS-Image support

• models in your normal folders (any sub-folder / name; the nodes find them by name or you pick them):

  • MOSS-Image Base: train on Base, not on a turbo or fine-tune

  • Qwen3.5-9B text encoder

  • FLUX.2 VAE

  • 8-step turbo LoRA, used for the test grid

• NVIDIA GPU: tested on 32 GB, 24 GB should work. Close other GPU work while training.

• git; uv is used when installed (faster), otherwise Python's venv.

Downloads (right-click → save into the folder in brackets):

• MOSS-Image Base: bf16, 24 GB+ VRAM or int8, smaller GPUs (models/diffusion_models)

• Qwen3.5-9B text encoder (models/text_encoders)

• FLUX.2 VAE (models/vae)

• 8-step turbo LoRA (models/loras)

Install

Extract the zip into ComfyUI/custom_nodes/, restart ComfyUI, press F5 and open Templates.

To make whole doujins with your LoRA, use the MOSS Doujin Kit. Full guide: Training MOSS-Image LoRAs for doujins. License: MIT. Training runs kohya-ss sd-scripts (moss_image_train_network.py).