LoRA on MOSS-Image generates pixel art that lands on a clean grid
berthol · reddit · 2026-10-29
The author found MOSS-Image-SFT produces fairly clean pixel grids out of the box, then trained a small LoRA to make the pixel scale predictable:
- Trained on 30 pixel-snapped 64×64 sprites upscaled to 512×512, targeting an 8× scale (each logical pixel = 8×8 block)
- Working sizes: 512×512→64×64 grid, 768×768→96×96, 1024×1024→128×128
- Grid can still drift and edges bleed, but results land close to the target
- Cleanup uses Sprite Fusion Pixel Snapper (also available as a Comfy node)
- Model, workflow, and training details are public on CivitAI and Hugging Face
More from Multimodal
- One prompt to MV: Magnific Move turns comments into v1→v7 edits across 8 motion designs — techhalla · 2026-10-29
- Magnific Move ships with a library of motion templates, tutorial thread wraps up — techhalla · 2026-10-29
- Character consistency trick: one base portrait, then swap only the mood word per prompt in Magnific One — techhalla · 2026-10-29
- Magnific Move drops: tutorial shows AI music video editing 'stopped being the hard part' — techhalla · 2026-10-29
- Reve publishes technical report on training infra that scaled to 10K GPUs with 10x fewer chips — cantrell · 2026-10-29
- Kling 4.0 video model lands on Cloudflare AI Gateway at $0.157/sec for 720p — ritakozlov · 2026-10-29