You need to agree to share your Hugging Face information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This model is a Derivative of Krea 2 and is distributed under the Krea 2 Community License Agreement (LICENSE.pdf in this repository), including Krea's Acceptable Use Policy. By requesting access you agree to be bound by those terms. Commercial use is only permitted below USD 1M annual revenue.
Log in or Sign Up to review the conditions and access this model content.
Krea 2 Lucas Wash
English | 中文 · Unofficial community model · 非官方社区模型
English
Krea 2 Lucas Wash is a style LoRA for Krea 2: loose ink lines, flat colour blocks and brush-stroke shadows in an anime illustration look. It was trained on Krea 2 Raw and, following Krea's recommended workflow, is meant to be used on Krea 2 Turbo (8 steps, no CFG), where it expresses strongly and renders in ~5 s per 1024² image.
Left: Krea 2 Turbo without LoRA · Right: with Krea 2 Lucas Wash (strength 0.8). Same prompt and seed.
Samples
ComfyUI, Krea 2 Turbo bf16, LoRA strength 0.8, 8 steps, CFG 1.0, euler / simple, 1024×1024, seed 42.
| Image | Prompt (seed 42) |
|---|---|
 |
sqmstyle, A girl with long silver hair and red eyes, wearing a black sailor uniform, looking back over her shoulder, wind blowing her hair, white background. |
 |
sqmstyle, A young man with messy black hair swinging a katana toward the viewer, intense expression, motion streaks and sparks, dark background. |
 |
sqmstyle, An old fisherman with a white beard and a straw hat, smoking a pipe, sitting on a wooden boat, warm evening light. |
 |
sqmstyle, A girl with short pink hair and green eyes, smiling and making a peace sign near her face, wearing a yellow hoodie, pale background. |
 |
sqmstyle, A black cat sitting on a windowsill, looking at the viewer, sunlight and soft shadows. |
 |
sqmstyle, A knight in dented silver armor holding a broken sword in the rain, low angle, dramatic lighting. |
Quick facts
| Trigger word | sqmstyle — start the prompt with sqmstyle, ... |
| Recommended base | Krea 2 Turbo (also works on Krea 2 Raw, weaker effect) |
| Recommended strength | 0.8 (range 0.6–1.0) |
| Format | LoRA, rank 32, bf16, 528 tensors, transformer.* keys (same layout as Comfy-Org's official Krea 2 LoRAs) |
| File | Krea-2-Lucas-Wash.safetensors (235 MB) |
Usage
ComfyUI (tested with ComfyUI 0.42.0)
- Put
Krea-2-Lucas-Wash.safetensorsinmodels/loras/. - Get Krea 2 Turbo (
krea2_turbo_bf16.safetensors, orkrea2_turbo_fp8_scaled.safetensorsfor less VRAM) intomodels/diffusion_models/, plus:
| Component | File (ComfyUI folder) | Source |
|---|---|---|
| Text encoder | qwen3vl_4b_bf16.safetensors or qwen3vl_4b_fp8_scaled.safetensors → models/text_encoders/ |
Comfy-Org/Krea-2 · ModelScope |
| VAE | qwen_image_vae.safetensors → models/vae/ |
same |
- Load
comfyui_workflow_api.json(API format) or build:Load Diffusion Model→LoraLoaderModelOnly(strength 0.8) →KSampler(steps 8, CFG 1.0, euler, simple), positiveCLIP Text EncodewithCLIPLoadertypekrea2, negativeConditioningZeroOut,Empty Latent Image1024×1024.
In our test all 528 LoRA tensors loaded (no "lora key not loaded" warnings).
DiffSynth-Studio (tested with 2.1.12)
import torch
from diffsynth.pipelines.krea2 import Krea2Pipeline, ModelConfig
pipe = Krea2Pipeline.from_pretrained(
torch_dtype=torch.bfloat16, device="cuda",
model_configs=[
ModelConfig(model_id="krea/Krea-2-Turbo", origin_file_pattern="turbo.safetensors"),
ModelConfig(model_id="Qwen/Qwen3-VL-4B-Instruct", origin_file_pattern="*.safetensors"),
ModelConfig(model_id="Qwen/Qwen-Image", origin_file_pattern="vae/diffusion_pytorch_model.safetensors"),
],
tokenizer_config=ModelConfig(model_id="Qwen/Qwen3-VL-4B-Instruct", origin_file_pattern=""),
)
pipe.load_lora(pipe.dit, "Krea-2-Lucas-Wash.safetensors", alpha=0.8)
image = pipe("sqmstyle, A girl with short pink hair and green eyes, smiling, pale background.",
seed=42, height=1024, width=1024, num_inference_steps=8, cfg_scale=1, mu=1.15)
image.save("wash.png")
The transformer.* key layout is the common diffusers/PEFT LoRA layout, so Krea2Pipeline.load_lora_weights in 🤗 diffusers is expected to work as well, but we have not tested it.
Recommended deployment
| Your GPU | Suggested setup |
|---|---|
| ≥ 40 GB | Turbo bf16 + text encoder bf16, everything resident; ~5 s / image (8 steps, 1024²) |
| 24 GB | Turbo fp8_scaled (13 GB) + text encoder fp8_scaled (5 GB); ComfyUI swaps the text encoder out before sampling |
| 12–16 GB | A community Turbo GGUF (Q5/Q6, 9–11 GB, ComfyUI-GGUF fork with Krea 2 support) + this LoRA |
Timings were measured on one NVIDIA RTX PRO 6000 Blackwell Max-Q (96 GB, ~277 bf16 TFLOPS measured); consumer GPUs with less compute will be proportionally slower.
Notes & limitations
- Strongest on anime characters; photographic subjects (animals, old people, armour) are also restyled into flat colour blocks.
- At strength ≥ 1.0 the LoRA starts to change content (faces look younger, compositions shift). Lower the strength if you need prompt fidelity.
- Training images were small (mostly 0.33–0.57 MP), so very fine texture detail at 2K is limited.
- Without the LoRA, Krea 2 renders the unknown word
sqmstyleas text (a signature-like scribble); with the LoRA it acts as a style trigger instead.
Training details
| Base | Krea 2 Raw (12.8 B single-stream MMDiT, Qwen3-VL-4B text encoder, Qwen-Image VAE) |
| Data | 54 anime-style illustrations with English captions prefixed by sqmstyle; dynamic resolution ≤ 1 MP, no upscaling |
| Method | LoRA r=32 (alpha 32) on attention/MLP/text-fusion/embedding projections (264 layers), AdamW, lr 1e-4, batch 1 |
| Schedule | 10 repeats × 5 epochs = 2 700 steps; released checkpoint: final epoch |
| Hardware / time | 1× RTX PRO 6000 Blackwell (96 GB), 51 min, 30 GB peak VRAM, DiffSynth-Studio 2.1.12 |
Model family
| Model | Type | Base for inference | Hugging Face | ModelScope |
|---|---|---|---|---|
| Krea 2 Lucas Wash | LoRA (r=32), 235 MB | Krea 2 Turbo (8 steps) | link | link |
| Krea 2 Lucas Impasto | full fine-tune, bf16, 25.6 GB | stand-alone (Raw-type, CFG) | link | link |
| Krea 2 Lucas Impasto Q8_0 GGUF | GGUF Q8_0, 14.6 GB | stand-alone, ComfyUI | link | link |
| Krea 2 Lucas Impasto Q5_1 GGUF | GGUF Q5_1, 10.7 GB | stand-alone, ComfyUI | link | link |
License & attribution
- License: Krea 2 Community License Agreement (a copy is included in this repository). By downloading or using this model you agree to be bound by it, including Krea's Acceptable Use Policy.
- Attribution notice (also in
NOTICE): Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit https://krea.ai/krea-2-licensing. - This model is a Derivative of the Krea Model and has been modified by Lucas (LoRA adapter trained on Krea 2 Raw).
- This is not an official Krea product and is not endorsed by Krea.
- Commercial use is only permitted if you (with your affiliates) have less than USD 1,000,000 annual revenue; otherwise you need an enterprise license from Krea (opensource@krea.ai).
- If you deploy this model, the license requires reasonable content filters (e.g. an NSFW / safety classifier, moderation, human review) and AI-generated content disclosure where required by law or platform policy.
Responsible use
- The model was trained on a small set of third-party anime-style illustrations (not redistributed). Do not present outputs as the work of any real artist, and respect the rights of artists and of the characters/IP you depict.
- Do not use it to create content that violates the Krea Acceptable Use Policy or applicable law.
中文
Krea 2 Lucas 淡墨
Krea 2 Lucas Wash（淡墨）是 Krea 2 的画风 LoRA：松弛的墨线、平涂色块、笔触式阴影的动漫插画风格。它在 Krea 2 Raw 上训练，并按 Krea 官方推荐的方式在 Krea 2 Turbo 上使用（8 步、不开 CFG）——风格表达最强，1024² 每张约 5 秒。
左：Krea 2 Turbo 原版 · 右：加载 Krea 2 Lucas Wash（强度 0.8）。提示词与种子相同。
效果图
ComfyUI，Krea 2 Turbo bf16，LoRA 强度 0.8，8 步，CFG 1.0，euler / simple，1024×1024，种子 42。
| 图像 | 提示词（种子 42） |
|---|---|
|
银发少女回眸：sqmstyle, A girl with long silver hair and red eyes, wearing a black sailor uniform, looking back over her shoulder, wind blowing her hair, white background. |
|
挥刀少年：sqmstyle, A young man with messy black hair swinging a katana toward the viewer, intense expression, motion streaks and sparks, dark background. |
|
船上抽烟斗的老渔夫：sqmstyle, An old fisherman with a white beard and a straw hat, smoking a pipe, sitting on a wooden boat, warm evening light. |
|
比剪刀手的粉发少女：sqmstyle, A girl with short pink hair and green eyes, smiling and making a peace sign near her face, wearing a yellow hoodie, pale background. |
|
窗台上的黑猫：sqmstyle, A black cat sitting on a windowsill, looking at the viewer, sunlight and soft shadows. |
|
雨中残剑骑士：sqmstyle, A knight in dented silver armor holding a broken sword in the rain, low angle, dramatic lighting. |
基本信息
| 触发词 | **sqmstyle**——提示词以 sqmstyle, ... 开头 |
| 推荐底座 | Krea 2 Turbo（也可用于 Krea 2 Raw，效果较弱） |
| 推荐强度 | 0.8（0.6–1.0） |
| 格式 | LoRA，秩 32，bf16，528 个张量，transformer.* 键名（与 Comfy-Org 官方 Krea 2 LoRA 相同） |
| 文件 | Krea-2-Lucas-Wash.safetensors（235 MB） |
使用方法
ComfyUI（已在 ComfyUI 0.42.0 实测）
- 把
Krea-2-Lucas-Wash.safetensors放进models/loras/。 - 准备 Krea 2 Turbo（
krea2_turbo_bf16.safetensors；显存紧张可用krea2_turbo_fp8_scaled.safetensors）放进models/diffusion_models/，以及：
| 组件 | 文件（ComfyUI 目录） | 来源 |
|---|---|---|
| 文本编码器 | qwen3vl_4b_bf16.safetensors 或 qwen3vl_4b_fp8_scaled.safetensors → models/text_encoders/ |
Comfy-Org/Krea-2 · ModelScope |
| VAE | qwen_image_vae.safetensors → models/vae/ |
同上 |
- 加载
comfyui_workflow_api.json（API 格式），或自行搭建：Load Diffusion Model→LoraLoaderModelOnly（强度 0.8）→KSampler（8 步，CFG 1.0，euler，simple）；正向CLIP Text Encode，CLIPLoader类型选 **krea2**；负向接ConditioningZeroOut；Empty Latent Image1024×1024。
实测 528 个 LoRA 张量全部加载（无 "lora key not loaded" 警告）。
DiffSynth-Studio（已在 2.1.12 实测）
代码见英文部分（与上方相同）。要点：底座用 krea/Krea-2-Turbo 的 turbo.safetensors，pipe.load_lora(pipe.dit, "Krea-2-Lucas-Wash.safetensors", alpha=0.8)，采样 num_inference_steps=8, cfg_scale=1, mu=1.15。
transformer.* 是通用的 diffusers/PEFT LoRA 键名，🤗 diffusers 的 load_lora_weights 理论上可用，但我们未实测。
推荐部署方案
| 显卡 | 建议 |
|---|---|
| ≥ 40 GB | Turbo bf16 + 文本编码器 bf16 全部常驻；8 步 1024² 约 5 秒/张 |
| 24 GB | Turbo fp8_scaled（13 GB）+ 文本编码器 fp8_scaled（5 GB），ComfyUI 会在采样前把文本编码器换出 |
| 12–16 GB | 社区 Turbo GGUF（Q5/Q6，9–11 GB，需 支持 Krea 2 的 ComfyUI-GGUF 分支）+ 本 LoRA |
耗时在单张 NVIDIA RTX PRO 6000 Blackwell Max-Q（96 GB，实测 bf16 约 277 TFLOPS）上测得；算力较低的消费级显卡会相应变慢。
注意事项与局限
- 对动漫人物效果最强；写实题材（动物、老人、盔甲）也会被改画成色块风格。
- 强度 ≥ 1.0 时会开始改变内容（人物显得更年轻、构图变化）；需要更贴合提示词时请调低强度。
- 训练图分辨率较低（多为 0.33–0.57 MP），2K 下的细微纹理有限。
- 不加载 LoRA 时，Krea 2 会把陌生词
sqmstyle当成文字画出来（类似签名）；加载后它才是风格触发词。
训练细节
| 底座 | Krea 2 Raw（12.8B 单流 MMDiT，Qwen3-VL-4B 文本编码器，Qwen-Image VAE） |
| 数据 | 54 张动漫风格插画，英文标注统一以 sqmstyle 开头；动态分辨率 ≤ 1 MP，只缩不放 |
| 方法 | LoRA r=32（alpha 32），覆盖注意力/MLP/文本融合/嵌入投影共 264 层；AdamW，lr 1e-4，batch 1 |
| 训练量 | 10 遍 × 5 epoch = 2 700 步；发布最后一个 epoch |
| 硬件/耗时 | 1× RTX PRO 6000 Blackwell（96 GB），51 分钟，显存峰值 30 GB，DiffSynth-Studio 2.1.12 |
模型家族
| 模型 | 类型 | 推理底座 | Hugging Face | ModelScope |
|---|---|---|---|---|
| Krea 2 Lucas Wash（淡墨） | LoRA（r=32），235 MB | Krea 2 Turbo（8 步） | 链接 | 链接 |
| Krea 2 Lucas Impasto（厚涂） | 全参数微调，bf16，25.6 GB | 独立模型（Raw 类，需 CFG） | 链接 | 链接 |
| Krea 2 Lucas Impasto Q8_0 GGUF | GGUF Q8_0，14.6 GB | 独立模型，ComfyUI | 链接 | 链接 |
| Krea 2 Lucas Impasto Q5_1 GGUF | GGUF Q5_1，10.7 GB | 独立模型，ComfyUI | 链接 | 链接 |
许可证与署名
- 许可证：Krea 2 社区许可协议（本仓库附带全文）。下载或使用本模型即表示你同意受该协议（含 Krea 可接受使用政策）约束。
- 署名声明（亦见
NOTICE）：Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit https://krea.ai/krea-2-licensing. - 本模型是 Krea 模型的衍生模型，由 Lucas 修改（基于 Krea 2 Raw 训练的 LoRA 适配器）。
- 本模型不是 Krea 官方产品，也未获得 Krea 背书。
- 商业使用：仅当你（含关联实体）年收入低于 100 万美元时允许；否则需向 Krea 申请企业许可（opensource@krea.ai）。
- 若对外部署本模型，协议要求配备合理的内容过滤（如 NSFW/安全分类器、审核流程、人工复核），并在法律或平台要求时标注 AI 生成。
负责任使用
- 本模型使用少量第三方动漫风格插画训练（训练数据不随模型分发）。请勿将生成结果冒充任何真实画师的作品，并尊重画师以及你所描绘角色/IP 的相关权利。
- 请勿用于生成违反 Krea 可接受使用政策或适用法律的内容。
- Downloads last month
- -
Model tree for KeweiXu/Krea-2-Lucas-Wash
Base model
krea/Krea-2-Raw