--- title: Krea-2-Lucas-Impasto-Q5_1-GGUF canonical_url: "https://www.modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Impasto-Q5_1-GGUF" md_url: "https://www.modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Impasto-Q5_1-GGUF.md" repository: Sunset20050214/Krea-2-Lucas-Impasto-Q5_1-GGUF last_updated: 2026-10-25 license: other base_model: - Sunset20050214/Krea-2-Lucas-Impasto base_model_relation: quantized library_name: - lora - gguf - pytorch frameworks: - Pytorch language: - en - zh downloads: 6 stars: 0 tags: - krea2 - krea-2 - gguf - text-to-image - anime - illustration - comfyui - quantized --- # Krea-2-Lucas-Impasto-Q5_1-GGUF > Krea-2-Lucas-Impasto-Q5_1-GGUF - Sunset20050214 在 ModelScope 开源的模型。Krea 2 Lucas Impasto Q51 GGUF Sunset20050214/Krea-2-Lucas-Impasto-Q5_1-GGUF 是 ModelScope 魔搭社区上的机器学习模型，采用 other 许可，基于 Sunset20050214/Krea-2-Lucas-Impasto 构建。 - **Repository**: Sunset20050214/Krea-2-Lucas-Impasto-Q5_1-GGUF - **License**: other - **Base model**: Sunset20050214/Krea-2-Lucas-Impasto - **Tags**: krea2, krea-2, gguf, text-to-image, anime, illustration, comfyui, quantized - **Downloads**: 6 - **Stars**: 0 - **Last updated**: 2026-10-25 Source: https://www.modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Impasto-Q5_1-GGUF --- # Krea 2 Lucas Impasto Q5_1 GGUF **[English](#english) | [中文](#中文)** · Unofficial community model · 非官方社区模型 ## English GGUF **Q5_1** quantization of [Krea 2 Lucas Impasto](https://modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Impasto), a full-parameter fine-tune of Krea 2 Raw toward a loose anime illustration style (bold ink lines, flat colour blocks, brush-stroke shadows). Q5_1 is the recommended file for 16 GB cards: it stayed fully resident under a 16 GB-class cap. If you have 20 GB+ you can also use the [Q8_0 GGUF](https://modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Impasto-Q8_0-GGUF). ## Samples ComfyUI + GGUF loader, this Q5_1 file, 30 steps, CFG 3.0, euler / simple, 1024×1024, seed 42, run under a 16 GB-class VRAM cap. | Image | Prompt (seed 42) | |---|---| | ![](images/q5-1.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/q5-2.jpg) | `sqmstyle, A young man with messy black hair swinging a katana toward the viewer, intense expression, motion streaks and sparks, dark background.` | | ![](images/q5-3.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/q5-4.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/q5-5.jpg) | `sqmstyle, A black cat sitting on a windowsill, looking at the viewer, sunlight and soft shadows.` | | ![](images/q5-6.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, ...` | | File | `Krea-2-Lucas-Impasto-Q5_1.gguf` (10.7 GB, 10,657,633,312 bytes) | | Layout | `general.architecture = krea2`; 258 large weight matrices in Q5_1, 172 small / sensitive tensors kept in F32 — identical layout to the community Krea 2 GGUFs ([molbal/krea2-gguf](https://huggingface.co/molbal/krea2-gguf)) | | Quantization error | relative weight error 3.9–4.5 % on sampled layers | | Recommended GPU | 16 GB | ## Usage (ComfyUI, tested with ComfyUI 0.42.0) 1. Install the GGUF loader with Krea 2 support: [molbal/ComfyUI-GGUF](https://github.com/molbal/ComfyUI-GGUF) (tested at commit `61843db`) into `custom_nodes/`. The original city96/ComfyUI-GGUF does not know the `krea2` architecture. 2. Put `Krea-2-Lucas-Impasto-Q5_1.gguf` in `models/diffusion_models/` (or `models/unet/`). 3. 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 | 4. Load [`comfyui_workflow_api.json`](comfyui_workflow_api.json) (API format) or build: `Unet Loader (GGUF)` → `KSampler` (**30 steps, CFG 3.0**, euler, simple), `CLIPLoader` type **`krea2`**, negative = empty `CLIP Text Encode`, `Empty Latent Image` 1024×1024. > ⚠️ **Use CFG ≈ 3.0.** CFG 4.5 caused colour fringing / glitch artefacts in ComfyUI in our tests, even with the unmodified bf16 Krea 2 Raw. ## Measured on a 16 GB-class budget The PyTorch allocator was hard-capped at 15 GiB (+ CUDA context ≈ a 16 GB card) and ComfyUI was told the same budget, so it planned offloading as it would on a real 16 GB GPU. | | bf16 reference (no cap) | **this Q5_1 file (16 GB cap)** | |---|---|---| | Peak device memory (incl. CUDA context) | 36.3 GB | **15.6 GiB** (fully resident, no offloading) | | Host RAM peak | 50 GB | 20 GB | | Time / image (30 steps, 1024²) | ~35 s | ~70 s | | OOM errors | — | none | | PSNR / SSIM vs bf16 (mean of 6 prompts) | — | 20.2 dB / 0.82 | Part of the extra time is ComfyUI swapping the 8.5 GB text encoder in and out of the 16 GB budget for every new prompt. 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. ![bf16 vs Q8_0 vs Q5_1](images/quant-compare.jpg) *Columns: bf16 · Q8_0 · Q5_1 (same prompt and seed, CFG 3.0). The anime prompts (rows 1–4) are visually near-identical. Rows 5–6 are off-domain prompts on which this fine-tune is sensitive: tiny weight perturbations can change the composition or flip into a glitch look (here Q8_0 flipped, Q5_1 did not), which is why the pixel metrics are low on those rows. One seed per prompt, so this does not mean Q5_1 is generally better than Q8_0.* ## Notes & limitations - Do not use with a Krea 2 Turbo workflow / Turbo sampling settings; this is a Raw-type model (CFG, 30+ steps). - Best on anime characters; off-domain prompts are less stable (see above). - Plan for ≥ 32 GB system RAM: the text encoder and offloaded weights live in RAM while not on the GPU. ## How it was made Quantized from the bf16 `Krea-2-Lucas-Impasto.safetensors` with `gguf-py` 0.19.0, reproducing the per-tensor layout of the community Krea 2 GGUFs; the written file was read back and checked tensor by tensor (names, order, shapes, types) against that layout, and sampled tensors were dequantized and compared with the bf16 source. ## 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 (GGUF Q5_1 quantization of Krea 2 Lucas Impasto, itself a 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 Q5_1 GGUF [Krea 2 Lucas Impasto（厚涂）](https://modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Impasto) 的 GGUF **Q5_1** 量化版。原模型是 Krea 2 Raw 的全参数微调，画风为松弛的动漫插画（粗墨线、平涂色块、笔触式阴影）。Q5_1 是 16 GB 显卡的推荐文件：在 16 GB 级显存上限下完整驻留。若显存 ≥ 20 GB，也可以用 [Q8_0 GGUF](https://modelscope.cn/models/Sunset20050214/Krea-2-Lucas-Impasto-Q8_0-GGUF)。 ## 效果图 ComfyUI + GGUF 加载器，本 Q5_1 文件，30 步，CFG 3.0，euler / simple，1024×1024，种子 42，在 16 GB 级显存上限下运行。 | 图像 | 提示词（种子 42） | |---|---| | ![](images/q5-1.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/q5-2.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/q5-3.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/q5-4.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/q5-5.jpg) | 窗台上的黑猫：`sqmstyle, A black cat sitting on a windowsill, looking at the viewer, sunlight and soft shadows.` | | ![](images/q5-6.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 | | 文件 | `Krea-2-Lucas-Impasto-Q5_1.gguf`（10.7 GB，10,657,633,312 字节） | | 布局 | `general.architecture = krea2`；258 个大权重矩阵为 Q5_1，172 个小/敏感张量保留 F32——与社区 Krea 2 GGUF（[molbal/krea2-gguf](https://huggingface.co/molbal/krea2-gguf)）布局完全一致 | | 量化误差 | 抽样层的相对权重误差 3.9–4.5 % | | 推荐显卡 | 16 GB | ## 使用方法（ComfyUI，已在 0.42.0 实测） 1. 安装支持 Krea 2 的 GGUF 加载器：[molbal/ComfyUI-GGUF](https://github.com/molbal/ComfyUI-GGUF)（实测 commit `61843db`），放进 `custom_nodes/`。原版 city96/ComfyUI-GGUF 不认识 `krea2` 架构。 2. 把 `Krea-2-Lucas-Impasto-Q5_1.gguf` 放进 `models/diffusion_models/`（或 `models/unet/`）。 3. 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 | 4. 加载 [`comfyui_workflow_api.json`](comfyui_workflow_api.json)（API 格式），或自行搭建：`Unet Loader (GGUF)` → `KSampler`（**30 步，CFG 3.0**，euler，simple），`CLIPLoader` 类型选 **`krea2`**，负向为空提示词的 `CLIP Text Encode`，`Empty Latent Image` 1024×1024。 > ⚠️ **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 bf16 Krea 2 Raw 也一样。 ## 16 GB 级显存实测 PyTorch 显存分配器被硬性限制在 15 GiB（加上 CUDA 上下文约等于一张 16 GB 卡），同时告知 ComfyUI 同样的预算，让它像在真实 16 GB 卡上一样规划卸载。 | | bf16 基准（不限显存） | **本 Q5_1 文件（16 GB 上限）** | |---|---|---| | 显存峰值（含 CUDA 上下文） | 36.3 GB | **15.6 GiB**（完整驻留，无卸载） | | 内存峰值 | 50 GB | 20 GB | | 每张耗时（30 步，1024²） | 约 35 秒 | 约 70 秒 | | OOM | — | 无 | | 与 bf16 的 PSNR / SSIM（6 条提示词均值） | — | 20.2 dB / 0.82 | 多出来的时间有一部分来自 ComfyUI 每换一条提示词都要把 8.5 GB 的文本编码器在 16 GB 预算内换入换出。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。 ![bf16 vs Q8_0 vs Q5_1](images/quant-compare.jpg) *三列依次为 bf16 · Q8_0 · Q5_1（提示词与种子相同，CFG 3.0）。动漫提示词（第 1–4 行）肉眼几乎一致。第 5–6 行是偏离训练题材的提示词，本微调模型对它们比较敏感：权重的微小扰动就可能改变构图或翻成故障风（这里 Q8_0 翻了、Q5_1 没翻），因此这两行的像素指标偏低。每条提示词只有一个种子，不能据此认为 Q5_1 普遍优于 Q8_0。* ## 注意事项与局限 - 不要用于 Krea 2 Turbo 的工作流或 Turbo 的采样设置；这是 Raw 类模型（开 CFG，30 步以上）。 - 对动漫人物效果最好；偏离训练题材的提示词稳定性较差（见上）。 - 建议系统内存 ≥ 32 GB：文本编码器和被卸载的权重不在显卡上时会占用内存。 ## 制作方式 由 bf16 的 `Krea-2-Lucas-Impasto.safetensors` 使用 `gguf-py` 0.19.0 量化，逐张量复刻社区 Krea 2 GGUF 的布局；写出后回读文件，逐个核对张量名称、顺序、形状与类型，并抽样反量化与 bf16 原值对比。 ## 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 Lucas Impasto（Krea 2 Raw 的全参数微调）进行 GGUF Q5_1 量化）。 - 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。