How to build a consistent-face dataset for retraining character LoRAs on Krea2
Kryimsson · reddit · 2026-09-23
A Reddit user reflects on their first character LoRA, which didn't turn out as well as hoped, and plans to retrain with a new dataset. They already have a realism workflow they like using Krea2 raw with a turbo LoRA, but are stuck on the dataset step: how to keep the same face across the whole dataset? They ask whether ReActor or BFS face-swap tools are the way to go and invite community advice.
More from Multimodal
- Gradio Open-Sources Qwen-Image 2.1 Rewriter: Dataset, Traces, Evals and Code — Gradio · 2026-09-23
- Distillation Costed ~$16 Total: It's a Data-Cost Problem, Not a Training-Cost Problem — Gradio · 2026-09-23
- ML-Intern in HuggingChat Ran the Whole Distillation Pipeline End to End — Gradio · 2026-09-23
- Gradio Distills Qwen-Image 2.1's 9B Prompt Rewriter Down to 0.8B for Laptops — Gradio · 2026-09-23
- Two-Ref Workflow: Character Sheet + Real Shot with GPT Image 2.5 and Seedance 2.5 — techhalla · 2026-09-23
- Qwen-Image-2.1 official demo live: 40 steps plus PE thinking unlocks full power — _akhaliq · 2026-09-23