Fizgig v8.5 adds MOSS-Image training: full LoRA suite on a 12GB card, tuned ComfyUI settings included
shootthesound · reddit · 2026-10-18
Fizgig v8.5.0 brings MOSS-Image (OpenMOSS's 5B Apache-2.0 image model) training support: LoRAs, LoKRs, sliders, full fine-tuning and multi-subject LoRAs, all loadable straight into ComfyUI. Training runs on the officially recommended Base checkpoint while previews render on the SFT checkpoint people actually generate with.
The author's tested-best settings: Fast preset (Rank 16, fixed LR 1e-4, 50 epochs at 0.25MP for facecrops), and sampling with resmultistep, 30 steps, CFG 4, shift 5, skipping block 22 on the negative—which beat the reference 50-step Euler in side-by-sides. A full recommended ComfyUI config table is included, plus using MOSS's bundled small FLUX.2 VAE to avoid spots and keeping the text encoder at high precision (officials say it's more fragile than the model).
Small cards: 12GB trains with SFT previews under 10.5GB VRAM; on 8GB the text encoder streams layer-by-layer from system RAM for exact caption caching. Total download 35GB.
More from Multimodal
- kimo: An Open-Source GUI-less Video Editor That AI Agents Drive via MCP — LieTop7965 · 2026-10-18
- MOSS-Image-SFT 8-step DMD distilled model trends on Hugging Face — lodestones · 2026-10-18
- Paying $0.50 per 10 seconds for video gen APIs — when is self-hosting worth it? — Key-Lychee-913 · 2026-10-18
- VFX pro builds entire Houdini scene by voice via Claude, 50-hour render on RTX 5090 — bilawalsidhu · 2026-10-18
- Asked 4 AI Assistants to Design Their Own Humanoid Lobsters — Styles Give Them Away — peteypeso · 2026-10-18
- Fix NVIDIA RTX VFX extension install failures in ComfyUI with one pip command — Mad4reds · 2026-10-18