Community Pins MOSS-Image SFT Vertical Banding to Layers 37-41, Ships De-Banding LoRA
MatlowAI · reddit · 2026-10-18
Reddit user MatlowAI traced the vertical banding artifact in MOSS-Image SFT to specific layers. Inspired by Kijai's LoRA, they compared SFT vs Base, swapped transformer blocks, and isolated most banding to layers 37-41 — fixed by a LoRA that reverts just those layers to Base weights while keeping the rest of SFT intact, with an optional layers 32-36 fix for stubborn cases. Benchmarks show the L37-41 fix LoRA cutting the banding metric from 1.29 to 1.07 (1.0 = clean) while retaining 77.4% SFT similarity. Higher resolution also helps but can morph into wider patterns. Full comparisons, ComfyUI workflows and the breakdown are on Hugging Face (MATLOWAI/MOSS-Image-SFT-DeBand).
More from Multimodal
- Flux3 vs Flux2 9B style transfer: better composition fidelity at 4-5x the speed — Winter_unmuted · 2026-10-18
- Agent spawns its own 'procedural audio composer' subagent and synthesizes a vaporwave track in Python — teortaxesTex · 2026-10-18
- A 3D journey tracing one message from a café table to an overseas data centre — techartist_ · 2026-10-18
- Running Fish Audio Drama 3 on a 16GB AMD GPU: Red Curtain ports the 35B MoE that officially wants 48GB — fechyyy · 2026-10-18
- Kling 4.0 opened 48 hours ago: 7 wild examples with full prompts — minchoi · 2026-10-18
- AI-generated 'Motion Blur' fashion images stun with long-exposure look — OVolosin82152 · 2026-10-18