Training LoRAs on Krea2? Skip These txtfusion Layers or Training Explodes
Capitan01R- · reddit · 2026-09-15
Reddit user Capitan01R- shares a key pitfall for training LoRAs on Krea2: the internal MLP blocks inside txtfusion cause training to explode and must be excluded for character/style or any general-purpose LoRA. The fix is adding an ignoreifcontains list covering txtfusion.layerwiseblocks.0/1, txtfusion.projector, txtfusion.refinerblocks.0/1, and txtmlp to the training config. The author also notes their previously teased project is still in progress.
More from Multimodal
- ElevenLabs MCP adds voice, music, image, video generation — and a user got a buyer — heyshrutimishra · 2026-09-15
- Phota Identity V2: train a persistent identity from 5 photos in 1 minute — stuffyokodraws · 2026-09-15
- How to remove the "AI sheen": making a generated character look real — DeviceCapital7041 · 2026-09-15
- GPT Image 2.5 consistency demo: 16-frame grid turned into GIF — gefei55 · 2026-09-15
- "AI music is unreasonably good": Redditor argues it already tops most human artists — Neither_Factor_5296 · 2026-09-15
- Reviving a forgotten 1934 French comic weekly with fal, Three.js and GPT 6 Astra — OdinLovis · 2026-09-15