moss/README.md

MOSS-Image LoRA Training Hub (moss/)

Training notes and templates for character LoRAs on MOSS-Image Base (OpenMOSS, 5B), trained with maltrainer on ComfyUI's single-file checkpoints.

---

🧭 Directory Map

Path	Description
📘 Combined Optimizer Training Guide	The method: train the dataset with AdamW 8-bit, Automagic3 and Prodigy at shorter-than-usual lengths, merge the three LoRAs at 0.7 each, and resize the merge to rank 8. Why it works, how to run it, and what's still open.
📗 MOSS-Image Training Guide	maltrainer's guide to MOSS-Image LoRA training: model files, how it matches ComfyUI, the timestep distribution, the templates and the first real run.
🔬 Stacking Findings	The finding that started the combined method: two under-strength LoRAs, stacked at 1.0 each, gave the best results so far.
🛠️ templates/	The three member templates: adamw8bit (60 epochs), automagic3 (20 epochs), prodigy (30 epochs).

---

⚡ The Method in Four Lines

1. Train three times, once per optimizer, each run kept short.
2. Merge the three LoRAs exactly, at strength 0.7 each (merge_loras.py).
3. Resize the rank-48 merge to rank 8 (resize_lora.py).
4. Load the result at strength 1.0.

It grew out of a finding on MOSS-Image: two separately trained, under-strength LoRAs stacked at 1.0 each looked better than either one turned up. The subject adds up across runs, while each run's own errors partly cancel.

📦 Model Files (Comfy-Org/MOSS-Image)

Role	File
DiT (train on the Base model)	diffusion_models/moss_image_base_bf16.safetensors
VAE	vae/flux2_vae.safetensors
Text encoder (caching only)	text_encoders/qwen3.5_9b_int8_convrot.safetensors