MOSS-Image character LoRA trained successfully on RX 7900 XT 20GB / Windows (Fizgig)
retained text
A concrete AMD/Windows success report: I trained a MOSS-Image character LoRA in Fizgig, then tested the exported weights in ComfyUI. Example: Nirrti, a fictional character from Stargate with a defined reference face. **Hardware / training environment** * Radeon RX 7900 XT 20GB, Ryzen 7 5800X3D, 64GB RAM. * Native Windows, Fizgig v8.9.91-5-g20331b14 (ROCm build). * PyTorch 2.12.0+rocm7.15.0a20260728; HIP 7.15.26290. **Dataset and training** * 18 curated AI-generated images, varied angles, expressions and outfits; English captions; trigger: nirrti. No horizontal flips. * MOSS-Image Base, INT8 loading, 8 swapped blocks, \~1MP buckets, batch 1, gradient accumulation 1. * Standard LoRA, rank/alpha 32/32; AdamW, constant LR 1e-4; EMA 0.98. * Clean-end Slide: 60:7:30:15:2. * SAM 3.1 face/hair masks, visually checked and corrected; outside-mask weight 0.1; prior preservation OFF. * 70 epochs / 1,260 steps; checkpoint each epoch. SFT previews every 10 epochs plus baseline, 30 steps, CFG 4, Turbo OFF. **Measured result** * Completed in \~3h35 including caching and previews; step counter averaged \~6.3 s/step. * Fizgig-reported peak GPU memory \~17.5 GiB over the whole run, including caching/previews; system RAM peaked around 42 GiB. This is not the training-only torch allocation. * Final \~149MB safetensors validated: 756 tensors, all finite. **ComfyUI validation** Native MOSS Base BF16, Qwen3.5-9B BF16 encoder, flux2-vae\_small, LoRA strength 1.0. 25 steps, CFG 2.5, er\_sde/simple, SLG layer 22 from 0 to 0.5. Portrait 1024x1024; full body 832x1248. The separate ComfyUI environment used PyTorch 2.12.0+rocm7.14.0. The face holds up in portrait, full body and a laboratory scene with different clothing. Attached: reference crop, Base portrait/full-body tests, and a story scene using an illustration prompt. No face restoration. Some SFT previews exaggerate skin texture; anatomy still needed a pose correction in one scene. I haven't run a controlled masked/unmasked comparison. Credit where it's due: ChatGPT/Codex was an active technical assistant for dataset/caption preparation, SAM mask checks and corrections, Fizgig setup, run monitoring/checkpoint validation, and ComfyUI workflows. I supplied the character brief/references, art direction and final likeness review. Thanks to the MOSS-Image and Fizgig developers!
document fetched 2026-10-29 00:33 · canonical https://reddit.com/r/StableDiffusion/comments/1xhn1ay/mossimage_character_lora_trained_successfully_on