<blockquote>⚠️ 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.</blockquote><blockquote>🖥️ 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. <a target="_blank" rel="ugc" href="https://github.com/7err4/fumika">GitHub</a> · <a target="_blank" rel="ugc" href="https://civarchive.com/articles/37215/fumika-a-free-local-studio-for-comics-images-and-lora-training">Civitai article</a>. This ComfyUI kit stays available if you prefer ComfyUI.</blockquote>Train MOSS-Image LoRAs (characters, outfits, styles) from inside ComfyUI, with no command line. Then pick the best epoch from one grid.<h2 id="how-it-works-(easy-mode)">How it works (easy mode)</h2>1. Train a LoRA (Templates → ComfyUI-MOSS-LoRA-Trainer):<ol><li>Make a folder, and inside it a folder named <code>4_mychar</code> with your images (20–100). Next to every image goes a <code>.txt</code> file with the same name and one sentence describing the image, always starting the same way: <code>Safe image. ohwx, an adult woman with long black hair, red eyes, white tank top. She stands in a park and smiles.</code></li><li>Enter the folder and a name, then press Run. A report tells you whether the dataset is OK and how long training takes.</li><li>Switch <code>start</code> 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.</li><li>Press Run any time to see the progress.</li></ol>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.<h2 id="settings">Settings</h2>The tested recipe from our released MOSS LoRAs:<ul><li>rank 32 / alpha 16</li><li>AdamW8bit, learning rate 1e-4, cosine schedule</li><li>logit-normal timesteps</li><li>text encoder frozen and cached</li><li>bf16 at 1024 px</li></ul>The number of epochs is calculated from your dataset size. Optional fields let you change:<ul><li>base model, text encoder, VAE</li><li>epochs</li><li>rank</li><li>learning rate</li><li>resolution</li></ul><h2 id="captions:-the-one-thing-that-matters">Captions: the one thing that matters</h2>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.<h2 id="requirements">Requirements</h2><ul><li>ComfyUI with MOSS-Image support</li><li>models in your normal folders (any sub-folder / name; the nodes find them by name or you pick them):<ul><li>MOSS-Image Base: train on Base, not on a turbo or fine-tune</li><li>Qwen3.5-9B text encoder</li><li>FLUX.2 VAE</li><li>8-step turbo LoRA, used for the test grid</li></ul></li><li>NVIDIA GPU: tested on 32 GB, 24 GB should work. Close other GPU work while training.</li><li>git; <code>uv</code> is used when installed (faster), otherwise Python's venv.</li></ul>Downloads (right-click → save into the folder in brackets):<ul><li>MOSS-Image Base: <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MOSS-Image/resolve/main/diffusion_models/moss_image_base_bf16.safetensors">bf16, 24 GB+ VRAM</a> or <a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MOSS-Image/resolve/main/diffusion_models/moss_image_base_int8_convrot.safetensors">int8, smaller GPUs</a> (<code>models/diffusion_models</code>)</li><li><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MOSS-Image/resolve/main/text_encoders/qwen3.5_9b_int8_convrot.safetensors">Qwen3.5-9B text encoder</a> (<code>models/text_encoders</code>)</li><li><a target="_blank" rel="ugc" href="https://huggingface.co/Comfy-Org/MOSS-Image/resolve/main/vae/flux2-vae.safetensors">FLUX.2 VAE</a> (<code>models/vae</code>)</li><li><a target="_blank" rel="ugc" href="https://huggingface.co/lodestones/Kromoss/resolve/main/8-steps-dmd-lora.safetensors">8-step turbo LoRA</a> (<code>models/loras</code>)</li></ul><h2 id="install">Install</h2>Extract the zip into <code>ComfyUI/custom_nodes/</code>, restart ComfyUI, press F5 and open Templates.To make whole doujins with your LoRA, use the <a target="_blank" rel="ugc" href="https://civarchive.com/models/3010228">MOSS Doujin Kit</a>. Full guide: <a target="_blank" rel="ugc" href="https://civarchive.com/articles/37028">Training MOSS-Image LoRAs for doujins</a>. License: MIT. Training runs kohya-ss sd-scripts (<code>moss_image_train_network.py</code>).

MOSS LoRA Trainer for ComfyUI - train MOSS-Image LoRAs in one node

v1.0 – first experimental release. Easy workflows "Train a LoRA" (3 fields, installs the trainer automatically, epochs from dataset size) and "Test my LoRA" (epoch grid with prompts built from your captions), plus 2 advanced workflows (full config, tag-to-sentence caption converter, epoch grid from stock nodes).

Files with SHA256: 9f66b9a85e1912e464c31e985a1dc3f023208da0060a3d904ee0c9ef75b1f878