Reddit proposal: train a corrective LoRA for FLUX Klein from Krea-corrected outputs
Friendly-Fig-6015 · reddit · 2026-11-13
A Reddit poster proposes an automated pipeline that turns a model's own mistakes into training data: generate images with FLUX Klein, auto-detect defect regions (e.g. malformed hands), use Krea or another editing model to correct those regions while preserving composition, filter failed corrections, caption accurately, then train a small corrective LoRA for FLUX Klein and evaluate on unseen prompts/seeds. Krea acts only as a teacher producing better training examples — no weight transfer, no Krea LoRA on FLUX. Known risks: the editor may introduce its own errors, style leakage into the LoRA, limited generalization, and license questions around training on generated outputs. The author runs an RTX 5060 Ti 16GB + 32GB RAM and wants to test how much runs on consumer hardware, starting narrow with hands; seeking collaborators on ComfyUI automation, inpainting, dataset filtering and LoRA training.
More from Multimodal
- Adobe brings Generative Remove to Premiere beta for non-destructive object erasure — jnack · 2026-11-13
- Blogger recreates Adobe keynote's video action control in ComfyUI, training-free with MiniMax H3 and LTX — TomLikesRobots · 2026-11-13
- HeyGen integrates Gemini Omni Flash 1.2 with 15s audio-synced video — HeyGen · 2026-11-13
- 192-prompt image model showdown: H3's realistic faces impress vs Krea 2 — dh7net · 2026-11-13
- AI artist tetsuoai drops new "Unit 01" robot-themed generated images — tetsuoai · 2026-11-13
- Reddit user puts MiniMax H3-Image through a busy backstage scene test with one text-to-image pass — Lazy_Importance_3135 · 2026-11-13