OneTrainer LoRA recipe: Prodigy optimizer plus a 3-version merge trick
jiggydancer · reddit · 2026-10-02
A Reddit user shares their working OneTrainer setup for training LoRAs for Krea on an RTX 5070 Ti: Prodigy optimizer, cosine scheduler, LR 1.0, batch size 1 with 2 accumulation steps, 768 resolution, noising strength 0.2–0.8, masked training on. The key workflow is step-count probing: train 2000 steps, generate test images for overfitting (blur/noise), then iterate ±20–25% to find the overfitting threshold, and finally merge three versions (slightly overfit plus two runs at 20% fewer steps) in ComfyUI for the best results.
More from Multimodal
- Garry Tan: This Realtime AI Video Agent Is Indistinguishable From a Human — garrytan · 2026-10-02
- Black Forest Labs adds up-to-4K image generation, opens commercial weights inquiries — bfl_ai · 2026-10-02
- Black Forest Labs launches FLUX 3 Image: pixel-perfect edits, 4K, open weights soon — bfl_ai · 2026-10-02
- Qwen 2.1 image editing struggles with pose changes while 2511 shines — Axymeus · 2026-10-02
- NoSpoon Studios offers free AI music videos and microdramas on its last public day — Kyrannio · 2026-10-02
- 36 interactive and playable "machines of generative art" coming to a living Generative Art Timeline — Merzmensch · 2026-10-02