Heretic or Abliterated?
retained text
Looking for recommendations for local LLM to convert an image into a usable prompt for Krea2 with max accuracy. Theres Heretic and Abliterated models for Qwen, Gemma and Deepseek. I am running a 5090 so hopefully a model that fits in the Vram along side the Krea model and text encoder. The krea2 workflow uses a normal non-abliterated text encoder.
document fetched 2026-10-12 00:30 · canonical https://reddit.com/r/StableDiffusion/comments/1x30bum/heretic_or_abliterated