Krea 2 Pixel Art LoRAs Lock Generations to Clean 32/64/128px Grids
Incognit0ErgoSum · reddit · 2026-09-16
A Reddit creator found that Krea 2 was trained on several pixel-art resolutions and released 32x32, 64x64 and 128x128 LoRAs that keep generations on clean pixel grids, avoiding ugly mixed-pixel output.
- Models are up on CivitAI and HuggingFace, with a companion ComfyUI node doing the minimal post-processing; heavier cleanup nodes were deliberately avoided since they over-guess row/column structure
- Handles long prompts well: demos include a single 128x128 generation of a 64-item icon sheet and a four-direction character sprite sheet
- The author notes Krea's weak direction-following but frames results as a usable starting point; 24/48/256px and full sprite-sheet LoRAs are planned
More from Multimodal
- Generative AI short film submitted to Lumara Film Festival — Kyrannio · 2026-09-17
- Google releases Gemma 3n: 2GB RAM multimodal model, first sub-10B to top 1300 on LMArena — joemeno · 2026-09-17
- NetEase Youdao open-sources Confucius R2T2, a 2B speech model with 200ms real-time transcription — dr_cintas · 2026-09-17
- Fountain 0 releases ODYSSEY: The Fall, an AI feature film shot entirely with Kling 3.0 — CurieuxExplorer · 2026-09-17
- YuE2 generates a 2-minute clip in just 41 seconds in real-world test — cocktailpeanut · 2026-09-17
- Hypit, a 7k-star open-source tool, clones viral videos into agentic workflows for Claude Code — nikola_mr64990 · 2026-09-17