Krea2 训练 LoKR 替代 LoRA 翻车，作者求调参思路
AkaToraX · reddit · 2026-11-02
A Krea2 user reports that character LoRAs give 90% consistency, but swapping to LoKR — reportedly better in the community — produced results "not close to my character at all." They can't find a LoKR loading node in ComfyUI (all nodes are LoRA-named) and ask whether KSampler settings need adjusting.
They shared their full AiToolkit recipe: 30 images, 1200-2400 steps, Krea2 raw, low VRAM, Qfloat8 quantization, linear rank 32, bf16, AdamW8Bit, lr 0.0001, weight decay 0.0001, linear balanced timesteps, MSE loss, cached text embeddings/latents, resolutions 512/768/1024. Inference: ComfyUI krea2 turbo int8, LoRA strength 0.75-1.25, 8 steps, CFG 1, ersde sampler, simple scheduler, denoise 1. The thread is effectively a request for how to retune rank/learning rate/strength when moving from LoRA to LoKR — useful reference for character-consistency training.
「多模态」频道最新
- Veo 3.1 系列扩展：AI 生成预告片「Planet Hell」首集发布 — Uncanny_Harry · 2026-11-03
- 用 Minimax H3 混搭〈道格〉+〈嘿阿诺〉画风做复古动画预告片 — sloppoz · 2026-11-03
- 作者分享用 Grok 写器乐配乐提示词的实战指南 — creatoroff · 2026-11-03
- Fable 5.5 进军动态设计：创意到动效全程可编辑代码，直出 AE/Premiere — garrytan · 2026-11-03
- RTX 5060 上练 MOSS 人脸 LoKr 百轮仅六成像，求复现配方 — Confident_Ad2351 · 2026-11-03
- EachLabs Lockstep 视频换角实测：质量领先且便宜 3.3 倍 — alifcoder · 2026-11-03