DiSReg: New LoRA Training Method Uses Self-Regulation to Eliminate Tag Pollution
bdsqlsz · x · 2026-09-06
hakomikan introduced DiSReg (Differential Self-Regulation), a new LoRA training technique shipped in TrainTrain that targets tag pollution: when training a character LoRA, tags like 1girl or solo absorb learning too, so the character activates without the trigger word.
How it works
- Classic regularization images barely help, since the model still trains on them rather than genuinely unlearning the unwanted associations.
- DiSReg inserts a self-regulation step after each normal training step: it compares noise predictions with and without the LoRA under a no-trigger condition, minimizing their difference with a stop-gradient, directly canceling the drift on non-trigger tags.
- The author argues this makes the regularization error strictly zero — an explicit "roll back" signal rather than an indirect self-correction loop.
The post includes the loss formulas for both the character-training and self-regulation steps, and the feature is already implemented in TrainTrain — useful for anyone training character/style LoRAs needing precise trigger control.
More from Multimodal
- Prompt Share: Monochrome High-Key Elegance for Fine-Art Black and White Portraits — azed_ai · 2026-09-06
- Danny Boyle limits AI to 30 seconds in new film 'Ink', sets usage boundaries — taherdhanera · 2026-09-06
- Google's WeatherNext 3 learns weather from satellite data, skipping physics sims — The Decoder · 2026-09-06
- MiniMax's H3 Max Director hits fal: realtime steerable video at $0.02/s promo — isidentical · 2026-09-06
- astra claims breakthrough: generating videos by rendering Blender 3D scenes instead of pixel-by-pixel diffusion — IndraVahan · 2026-09-06
- Yoroll launches YoLive, a never-ending AI livestream film with 4s video generation — vista8 · 2026-09-06