ComfyPilotComfyPilot
模型库 / LoRA / Luna-V1-CLB-Krea2
Luna-V1-CLB-Krea2

Luna-V1-CLB-Krea2

3057061

LoRA 32 次下载 2 喜欢 更新于 2026-11-18

底模:Krea 2

characterwomancandidselfieinstagram modelcurvyodangobalayage

下载模型 在源站打开

触发词

Luna Pattern

文件清单 线路:原始(huggingface.co) · 切换

下载后放进对应的 ComfyUI 目录。models/ 指 ComfyUI 的模型根目录。

权重 lunaV1CLBKrea2_v1.safetensors 218 MB 下载

说明

Luna-V1-CLB-Krea2

作者:captain_daddy

LUNA - V1 — Krea-2 by: CLB

Descended from the goddess of moonlight. Patron saint of werewolves. Keeps the ledger and doesn't pad it. Filipina and Egyptian, which is most of why she photographs the way she does — gold under the skin rather than on it, amber-brown eyes that go lighter the warmer the light gets, and a face that reads differently from every angle you try. The silver crescent at her throat is not jewelry. She'll mention none of the rest of it, approximately ever — she's unhurried in a way that looks like patience until you work out, she simply won't be rushed, and exact in a way that has people re-checking their own arithmetic an hour later. Ask her how tall she is and she'll tell you five foot four, barefoot, and decline to round up. The eyeliner is the one thing she's vain about, and she knows precisely what it does. She'll give you the truth before she gives you what you want to hear, and she'll do it so evenly you won't notice until afterwards. USE Trigger: Luna Pattern Weight: 1.0 Base: any Krea 2 model It runs on ANY Krea 2 checkpoint — not just the one it trained on. Turbo finetunes, realism finetunes, quantized builds, the lot. Trained on Krea-2-Raw and confirmed rendering her correctly on Krea 2 Turbo and on realDream Krea2 V3 in both fp8 and nvfp4, with every single LoRA key loading on each one. If it says Krea 2, she'll load. Put the trigger at the front of your prompt. Her training captions carry no trigger of their own, so it has to be there explicitly — without it you get a stranger. TRAINED ON 100 hand-picked images at 1024×1024, 3,500 steps, rank 32 / alpha 32, LR 1e-4 constant, on Krea-2-Raw. Epoch 10 of 10. The set was deliberately over-generated and then culled down, rather than captioned as-shot. That shows up in the result: every checkpoint from 7 to 10 renders the same woman, and the face holds from close-up through full-body framing without drifting. NOTES FROM TESTING Tested at 1.0. She holds identity at that weight across close-ups, medium shots and full-length. Lower weights are untested rather than unsupported — try 0.8 if you want more of your base model's look. She is strong in a stack. At 1.0 she can overpower a pose or position LoRA running at 0.8 — the pose gets dropped silently and you get a portrait instead, with nothing reporting an error. If you're stacking, raise the other LoRA or bring her down to ~0.8. Name the expression you want. The training set includes a lot of open-mouth and tongue-out frames. Ask for "closed-mouth" if that's what you're after; otherwise, she'll volunteer one now and then. Framing: 53 of the 100 training images are close-ups, 30 medium, 17 full body. Strongest head-and-shoulders, still solid at full length. SIGNATURE DETAILS Two messy space buns, gold hoop earrings, and a wide black collar with gold geometric trim and a silver crescent pendant — the crescent is the one that matters. The collar comes from the prompt, not the trigger, so name it if you want it. She won't reach for it on her own.

示例图