MOSS-Image character LoRA training on 12GB RTX 3060: 7.1GB VRAM, 54s inference
Mycelia_Bygonepath · reddit · 2026-10-31
A team integrated MOSS-Image (5B DiT int8 weights + Qwen3.5-9B text encoder + Flux2 VAE) into a local ComfyUI studio on a single 12GB RTX 3060 and trained a 21MB Rank-16 character LoRA, cutting generation from 345s to 54–65s. Key lessons:
- Flux2 128-channel latent trap: Flux2 VAE packs 2×2 spatial sub-patches into channels (32×2×2=128); bilinear F.interpolate on these latents mixes sub-patch phases and produced a blurred silhouette in the first run. Fix: use clean uninterpolated 36×36 latent crops (576×576 patches)
- Non-differentiable int8 attention: attentioncomfykitchenint8 gives zero gradients to proj.qkv LoRA adapters. Setting blk.comfyattention.function = None for layers 28..41 restored autograd — loss dropped to 0.3865 in 240 steps (10 min) at 7.10GB peak VRAM
- Bypass beats Merged at inference: LoraLoaderModelOnly re-quantizes weights (1.0156 max quantization error); LoraLoaderBypassModelOnly keeps int8 base untouched with only 1.8% step-time overhead
All gallery PNGs embed the full ComfyUI workflow for reproduction.
More from Multimodal
- Crayon-style cartoon set generated with Midjourney v8.2 and sref 1786649973 — azed_ai · 2026-10-31
- AI-Generated Sitcom Clip Shows Virtual Character Wendy Skateboarding Down a Street — Old_Put1306 · 2026-10-31
- Heavy Wan 2.2 Character LoRA Meets Video: Animate vs VACE vs Scail Workflow Showdown — Business-Chocolate-4 · 2026-10-31
- MOSS AMA Roadmap: Full-Duplex Live Model, Image Turbo, Consumer-GPU Text2Vid — tomByrer · 2026-10-31
- Intel Arc B580/B70 ComfyUI Benchmarks: omnixpu Beats Native pytorchxpu — Willing-Grade5566 · 2026-10-31
- Recreating the Life of King Tut With AI: "The House of Amun" — retarded_raptor · 2026-10-31