--- title: Krea-2-Lucas-Impasto canonical_url: "https://www.modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Impasto" md_url: "https://www.modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Impasto.md" repository: Sunset20050214/Krea-2-Lucas-Impasto last_updated: 2026-10-25 parameters: 12.8B tensor_type: - BF16 library_name: - pytorch - safetensors frameworks: - Pytorch downloads: 2 stars: 0 license: other license_name: krea-2-community-license license_link: LICENSE.pdf tasks: - text-to-image-synthesis base_model: - krea/Krea-2-Raw base_model_relation: finetune language: - en - zh tags: - krea2 - krea-2 - text-to-image - full-finetune - anime - illustration - comfyui - diffsynth --- # Krea-2-Lucas-Impasto > Krea-2-Lucas-Impasto - Sunset20050214 在 ModelScope 开源的模型。Krea 2 Lucas Impasto - **Repository**: Sunset20050214/Krea-2-Lucas-Impasto - **Parameters**: 12.8B - **Downloads**: 2 - **Stars**: 0 - **Last updated**: 2026-10-25 Source: https://www.modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Impasto --- # Krea 2 Lucas Impasto **[English](#english) | [中文](#中文)** · Unofficial community model · 非官方社区模型 ## English **Krea 2 Lucas Impasto** is a **full-parameter fine-tune** (all 12.8 B parameters) of [Krea 2 Raw](https://huggingface.co/krea/Krea-2-Raw) toward a loose anime illustration style: bold ink lines, flat colour blocks, brush-stroke shadows and generous white space. Compared with the [Wash LoRA](https://modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Wash) it gives the heaviest, most coherent brushwork, at the cost of Raw-style sampling (CFG, 30–52 steps). ![comparison](images/impasto-compare.jpg) *Left: Krea 2 Raw · Right: Krea 2 Lucas Impasto. DiffSynth-Studio, 52 steps, CFG 4.5, same prompt and seed. (The text in the left images is Krea 2 Raw rendering the unknown trigger word as a signature.)* ## Samples ComfyUI, bf16, 30 steps, CFG 3.0, euler / simple, 1024×1024, seed 42. | Image | Prompt (seed 42) | |---|---| | ![](images/impasto-1.jpg) | `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.` | | ![](images/impasto-2.jpg) | `sqmstyle, A young man with messy black hair swinging a katana toward the viewer, intense expression, motion streaks and sparks, dark background.` | | ![](images/impasto-3.jpg) | `sqmstyle, An old fisherman with a white beard and a straw hat, smoking a pipe, sitting on a wooden boat, warm evening light.` | | ![](images/impasto-4.jpg) | `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.` | | ![](images/impasto-5.jpg) | `sqmstyle, A black cat sitting on a windowsill, looking at the viewer, sunlight and soft shadows.` | | ![](images/impasto-6.jpg) | `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, ...` | | Type | Stand-alone diffusion transformer (replaces Krea 2 Raw); needs the Krea 2 text encoder + VAE | | Format | single-file safetensors, bf16, 430 tensors, same key layout as Krea's `raw.safetensors` | | File | `Krea-2-Lucas-Impasto.safetensors` (25.6 GB) | | Sampling | ComfyUI: 30 steps, **CFG 3.0**, euler/simple · DiffSynth: 52 steps, CFG 4.5 | | Lower-VRAM versions | [Q8_0 GGUF](https://modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Impasto-Q8_0-GGUF) (14.6 GB) · [Q5_1 GGUF](https://modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Impasto-Q5_1-GGUF) (10.7 GB) | ## Usage ### ComfyUI (tested with ComfyUI 0.42.0) 1. Put `Krea-2-Lucas-Impasto.safetensors` in `models/diffusion_models/` (it is a drop-in replacement for `krea2_raw_bf16.safetensors`: identical 430-tensor layout). 2. Add the text encoder and VAE: | Component | File (ComfyUI folder) | Source | |---|---|---| | Text encoder | `qwen3vl_4b_bf16.safetensors` or `qwen3vl_4b_fp8_scaled.safetensors` → `models/text_encoders/` | [Comfy-Org/Krea-2](https://huggingface.co/Comfy-Org/Krea-2) · [ModelScope](https://modelscope.cn/models/Comfy-Org/Krea-2) | | VAE | `qwen_image_vae.safetensors` → `models/vae/` | same | 3. Load [`comfyui_workflow_api.json`](comfyui_workflow_api.json) (API format) or build: `Load Diffusion Model` → `KSampler` (**30 steps, CFG 3.0**, euler, simple), positive `CLIP Text Encode` with `CLIPLoader` type **`krea2`**, negative `CLIP Text Encode` with an empty prompt, `Empty Latent Image` 1024×1024. > ⚠️ **Use CFG ≈ 3.0 in ComfyUI.** In our tests CFG 4.5 produced colour fringing / glitch artefacts in ComfyUI — also with the *unmodified* Krea 2 Raw — while 3.0 was clean. ### DiffSynth-Studio (tested with 2.1.12) ```python import torch from diffsynth import load_state_dict 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-Raw", origin_file_pattern="raw.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.dit.load_state_dict(load_state_dict("Krea-2-Lucas-Impasto.safetensors", torch_dtype=torch.bfloat16, device="cuda")) image = pipe("sqmstyle, An old fisherman with a white beard and a straw hat, smoking a pipe, sitting on a wooden boat.", seed=42, height=1024, width=1024, num_inference_steps=52, cfg_scale=4.5) image.save("impasto.png") ``` (DiffSynth's CFG 4.5 with 52 steps is clean; the ComfyUI artefact above did not occur there.) ## Recommended deployment | Your GPU | Suggested setup | |---|---| | ≥ 48 GB | This bf16 file + text encoder bf16. Measured in ComfyUI: **36 GB** peak device memory, **~35 s / image** (30 steps, 1024²) | | 24–40 GB | This bf16 file with ComfyUI's automatic offloading (slower), or the [Q8_0 GGUF](https://modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Impasto-Q8_0-GGUF) | | 16 GB | [Q5_1 GGUF](https://modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Impasto-Q5_1-GGUF) — measured to fit entirely under a 16 GB-class cap | 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. System RAM: ≥ 64 GB recommended for the bf16 file (ComfyUI peaked at 50 GB host RSS). ## Notes & limitations - **Do not load these weights into a Krea 2 Turbo workflow.** Although the architecture is identical, replacing Turbo's weights with this checkpoint produced broken images in our tests; this model needs Raw-style sampling (CFG, 30+ steps). For fast 8-step use, take the [Wash LoRA](https://modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Wash) on Turbo instead. - Best on **anime characters**. Off-domain prompts (photographic animals, armour) are less stable: with high CFG they can flip into neon / glitch looks. Keep CFG ≈ 3 in ComfyUI. - The change from Raw is spread across the whole network: a rank-128 LoRA extracted from it keeps only ~9 % of the weight delta, so it cannot be compressed into a small LoRA without losing most of the effect. - Training images were small (mostly 0.33–0.57 MP), so fine texture detail at high resolution is limited. ## 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 | full-parameter fine-tune in bf16 with **Adafactor + stochastic rounding** (fp32 update, stochastic write-back so 1e-5 updates are not rounded away), lr 1e-5, no weight decay, batch 1, gradient checkpointing | | Schedule | 10 repeats × 4 epochs = 2 160 steps; released checkpoint: final epoch | | Hardware / time | 1× RTX PRO 6000 Blackwell (96 GB), 88 min, **54.2 GiB** 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](https://huggingface.co/KeweiXu/Krea-2-Lucas-Wash) | [link](https://modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Wash) | | Krea 2 Lucas Impasto | full fine-tune, bf16, 25.6 GB | stand-alone (Raw-type, CFG) | [link](https://huggingface.co/KeweiXu/Krea-2-Lucas-Impasto) | [link](https://modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Impasto) | | Krea 2 Lucas Impasto Q8_0 GGUF | GGUF Q8_0, 14.6 GB | stand-alone, ComfyUI | [link](https://huggingface.co/KeweiXu/Krea-2-Lucas-Impasto-Q8_0-GGUF) | [link](https://modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Impasto-Q8_0-GGUF) | | Krea 2 Lucas Impasto Q5_1 GGUF | GGUF Q5_1, 10.7 GB | stand-alone, ComfyUI | [link](https://huggingface.co/KeweiXu/Krea-2-Lucas-Impasto-Q5_1-GGUF) | [link](https://modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Impasto-Q5_1-GGUF) | ## License & attribution - **License:** [Krea 2 Community License Agreement](LICENSE.pdf) (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`](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 (full-parameter fine-tune of 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 厚涂 **Krea 2 Lucas Impasto（厚涂）**是对 [Krea 2 Raw](https://huggingface.co/krea/Krea-2-Raw) 的**全参数微调**（全部 128 亿参数），画风为松弛的动漫插画：粗墨线、平涂色块、笔触式阴影和大量留白。与 [Wash LoRA](https://modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Wash) 相比，它的笔触最重、最统一，代价是需要 Raw 式采样（开 CFG，30–52 步）。 ![对比](images/impasto-compare.jpg) *左：Krea 2 Raw · 右：Krea 2 Lucas Impasto。DiffSynth-Studio，52 步，CFG 4.5，提示词与种子相同。（左图中的文字是 Krea 2 Raw 把陌生的触发词当成签名画了出来。）* ## 效果图 ComfyUI，bf16，30 步，CFG 3.0，euler / simple，1024×1024，种子 42。 | 图像 | 提示词（种子 42） | |---|---| | ![](images/impasto-1.jpg) | 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.` | | ![](images/impasto-2.jpg) | 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.` | | ![](images/impasto-3.jpg) | 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.` | | ![](images/impasto-4.jpg) | 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.` | | ![](images/impasto-5.jpg) | 窗台上的黑猫：`sqmstyle, A black cat sitting on a windowsill, looking at the viewer, sunlight and soft shadows.` | | ![](images/impasto-6.jpg) | 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 | | 类型 | 独立的扩散 Transformer（替换 Krea 2 Raw）；需配合 Krea 2 的文本编码器与 VAE | | 格式 | 单文件 safetensors，bf16，430 个张量，键名与 Krea 官方 `raw.safetensors` 一致 | | 文件 | `Krea-2-Lucas-Impasto.safetensors`（25.6 GB） | | 采样 | ComfyUI：30 步，**CFG 3.0**，euler/simple · DiffSynth：52 步，CFG 4.5 | | 低显存版本 | [Q8_0 GGUF](https://modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Impasto-Q8_0-GGUF)（14.6 GB）· [Q5_1 GGUF](https://modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Impasto-Q5_1-GGUF)（10.7 GB） | ## 使用方法 ### ComfyUI（已在 ComfyUI 0.42.0 实测） 1. 把 `Krea-2-Lucas-Impasto.safetensors` 放进 `models/diffusion_models/`（可直接替换 `krea2_raw_bf16.safetensors`：430 个张量布局完全一致）。 2. 准备文本编码器与 VAE： | 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](https://huggingface.co/Comfy-Org/Krea-2) · [ModelScope](https://modelscope.cn/models/Comfy-Org/Krea-2) | | VAE | `qwen_image_vae.safetensors` → `models/vae/` | 同上 | 3. 加载 [`comfyui_workflow_api.json`](comfyui_workflow_api.json)（API 格式），或自行搭建：`Load Diffusion Model` → `KSampler`（**30 步，CFG 3.0**，euler，simple）；正向 `CLIP Text Encode`，`CLIPLoader` 类型选 **`krea2`**；负向接一个空提示词的 `CLIP Text Encode`；`Empty Latent Image` 1024×1024。 > ⚠️ **ComfyUI 里请用 CFG ≈ 3.0。** 我们实测 ComfyUI 中 CFG 4.5 会出现彩边/故障伪影——*未修改的* Krea 2 Raw 也一样；3.0 则干净。 ### DiffSynth-Studio（已在 2.1.12 实测） 代码见英文部分。要点：先加载官方 Raw 管线，再 `pipe.dit.load_state_dict(load_state_dict("Krea-2-Lucas-Impasto.safetensors", ...))`，采样 `num_inference_steps=52, cfg_scale=4.5`（DiffSynth 下 4.5 是干净的，不会出现上述 ComfyUI 伪影）。 ## 推荐部署方案 | 显卡 | 建议 | |---|---| | ≥ 48 GB | 本 bf16 文件 + bf16 文本编码器。ComfyUI 实测显存峰值 **36 GB**，**约 35 秒/张**（30 步，1024²） | | 24–40 GB | 本 bf16 文件配合 ComfyUI 自动卸载（较慢），或 [Q8_0 GGUF](https://modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Impasto-Q8_0-GGUF) | | 16 GB | [Q5_1 GGUF](https://modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Impasto-Q5_1-GGUF)——实测在 16 GB 级显存上限内完整驻留 | 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。 内存：使用 bf16 文件建议 ≥ 64 GB（ComfyUI 进程内存峰值 50 GB）。 ## 注意事项与局限 - **不要把这份权重装进 Krea 2 Turbo 的工作流。** 虽然架构相同，但我们实测把 Turbo 的权重换成本模型会出坏图；本模型需要 Raw 式采样（开 CFG、30 步以上）。想要 8 步快速出图，请在 Turbo 上使用 [Wash LoRA](https://modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Wash)。 - 对**动漫人物**效果最好。偏离训练题材的提示词（写实动物、盔甲）稳定性较差，CFG 偏高时可能变成霓虹/故障风。ComfyUI 中请保持 CFG ≈ 3。 - 相对 Raw 的改动分布在整个网络中：从中提取的秩 128 LoRA 只保留约 9% 的权重变化，无法压缩成小 LoRA 而不损失大部分效果。 - 训练图分辨率较低（多为 0.33–0.57 MP），高分辨率下的细微纹理有限。 ## 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 | bf16 全参数微调，使用 **Adafactor + 随机舍入**（fp32 计算更新、随机舍入写回，避免 1e-5 的更新被 bf16 舍入吃掉），lr 1e-5，无权重衰减，batch 1，梯度检查点 | | 训练量 | 10 遍 × 4 epoch = 2 160 步；发布最后一个 epoch | | 硬件/耗时 | 1× RTX PRO 6000 Blackwell（96 GB），88 分钟，显存峰值 **54.2 GiB**，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](https://huggingface.co/KeweiXu/Krea-2-Lucas-Wash) | [链接](https://modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Wash) | | 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](https://huggingface.co/KeweiXu/Krea-2-Lucas-Impasto) | [链接](https://modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Impasto) | | Krea 2 Lucas Impasto Q8_0 GGUF | GGUF Q8_0，14.6 GB | 独立模型，ComfyUI | [链接](https://huggingface.co/KeweiXu/Krea-2-Lucas-Impasto-Q8_0-GGUF) | [链接](https://modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Impasto-Q8_0-GGUF) | | Krea 2 Lucas Impasto Q5_1 GGUF | GGUF Q5_1，10.7 GB | 独立模型，ComfyUI | [链接](https://huggingface.co/KeweiXu/Krea-2-Lucas-Impasto-Q5_1-GGUF) | [链接](https://modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Impasto-Q5_1-GGUF) | ## 许可证与署名 - **许可证：**[Krea 2 社区许可协议](LICENSE.pdf)（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`](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**（对 Krea 2 Raw 进行全参数微调）。 - 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。