Redder bypasses Krea2 suppression by matching LoRA RMS changes instead of copying weights
Capitan01R- · reddit · 2026-09-12
Reddit user Capitan01R shares a clean experimental fix for Krea2's refusal/suppression behavior in text-to-image generation.
- Motivation: LoRAs stacked together can carry textfusion changes and muddy outputs, and you can't directly set final RMS per parameter with a LoRA.
- Method: after training hundreds of LoRAs, the author hooked a node to measure the RMS changes a LoRA applies to the model's text-path parameters, then rescales existing text-path tensors to match that RMS — reproducing the effect without copying the rank-64 learned update, with less destructiveness.
- Next step: once values are finalized, the node will be published with baked-in values to drop after the last model patch in the workflow, aiming for no bleed and no artifacts.
More from Multimodal
- Dev shows off Freaky Friday, a trippy short made with his new app Basic Slop — bennash · 2026-09-12
- Reddit user benchmarks 4 new MiniMax video models: Fused fastest at 1:11 with best audio — agapes1270 · 2026-09-12
- AI Painting of the Day: "The Music Copyist" Made with ChatGPT Image 2.0 — DeryaTR_ · 2026-09-12
- Higgsfield went 0 to $5.4B in 17 months: growth lead shares playbook starting with localizing launches — xiaohu · 2026-09-12
- MiniMax H3 text-to-video recreates the "disrespectful to dirt" meme scene — TheOrangeSplat · 2026-09-12
- NoSpoon auto-generates continuous AI drama episodes, 5-30 minutes each — TinfoilTricorn · 2026-09-12