Krea2-Turbo-HD-V1

Krea2-HD-V1

Krea2-Turbo 高清优化版，同步微调 VAE，更多细节、更好质感。

Krea2-Turbo HD optimized version, with synchronized fine-tuned VAE, more details, better texture.

---

✌️✌️✌️ 0Day Support "int4_convrot" quantized format with Comfy-Org, Krea2-Turbo-HD-V1-int4_convrot.safetensors

---

针对低显存的用户，推荐 smthemex/ComfyUI_UniBlockSwap，只需 4-6GB 显存，即可跑 bf16 全量模型。 Recommend smthemex/ComfyUI_UniBlockSwap plugin for LOW VRAM users, it only requires 4-6GB of VRAM to run the full bf16 model.

您可以尝试一下全新的 ComfyUI GGUF 模型加载插件 smthemex/ComfyUI_Dif_GGUF，它能适配更多的 GGUF 文件格式，并且将很快集成低显存(4-8GB)显卡的 GGUF 模型块卸载管理能力。 Recommend to try the new GGUF ComfyUI loader plugin, smthemex/ComfyUI_Dif_GGUF, it compatible with more GGUF format, and will add low VRAM (4-8GB) management for GGUF soon.

---

采用极致优化的分层加混合精度量化方式，确保 4bit 以上都能获得良好的输出质感，把量化损失降到最低。不同量化精度出图品质对比如下(Using an extremely optimized layered mixed-precision quantization method, it ensures good output quality for 4-bit and above while minimizing quantization loss. Here's a comparison of image quality at different quantization levels:)：

mxfp4 和 nvfp4 量化的 gguf 格式，目前属于试验阶段，供有试用经验的爱好者下载测试，目前 ComfyUI-GGUF 插件还未提供 torch dequant 代码，跑的话会非常慢。本仓库提供的所有 gguf 格式模型，目前 ComfyUI 环境均可直接加载。(The mxfp4 and nvfp4 quantized gguf formats are currently in the experimental stage, meant for enthusiasts with some trial experience to download and test. Right now, the ComfyUI-GGUF plugin doesn’t provide torch dequant code, so running them will be very slow. All gguf format models provided in this repository can be directly loaded in the ComfyUI environment.)

本模型为基础模型，对各种 LoRA 兼容性良好，本模型不直接支持 NSFW 内容，想要尝试 NSFW 内容的，可以适配 alcaitiff‘s MysticXXX LoRA ，同步微调的 HD-VAE 可以适配所有使用 Qwen-Image / Wan 2.1 VAE 的模型，能提高细节渲染、清晰度和对比度，色彩 gamma 方面，与官方原 VAE 接近。(This model is a base model and is highly compatible with various LoRAs. It doesn't natively support NSFW content, but if you want to try NSFW content, you can adapt alcaitiff‘s MysticXXX LoRA. The HD-VAE fine-tuned alongside it can be apply with any model that used Qwen-Image / Wan 2.1 VAE, improving detail rendering, clarity, contrast, and color gamma, staying close to the official original VAE.)

下载模型前，请先参考右侧或下方展示的图例，如果觉得有尝试的价值再下载，免得耽误您的流量和时间。同时，请遵守 Krea 官方的社区使用许可。(Before downloading the model, please first check out the sample widget on the right area or below gallery. Only download it if you think it's worth to trying, so you don't waste your network traffic and time. Also, please follow Krea's official community usage guidelines.)

<Gallery />

本模型唯一发布渠道为 Huggingface 平台，未经本人允许，谢绝 <u>搬运</u> 和 <u>转发</u>，谢谢！

The only official release channel for this model is the Huggingface platform. Without my permission, please <u>do not copy</u> or <u>share it</u>. Thanks!

特别致谢，Special thanks：

感谢：@亮亮rayne，@GuangyuanSD 在模型调制优化过程一如既往的给与技术、构图、美学、社群等方面的指导和支持。右侧例图的大部分提示词，来自 @亮亮rayne 大佬。

Thanks to @亮亮rayne and @GuangyuanSD for their usual guidance and support in the model tuning and optimization process, covering technology, composition, aesthetics, community, and more. Most of the prompts in the example image on the right come from @亮亮rayne .

如果您喜欢我的模型，请支持 ❤️ 我以获得更多的算力来调优模型。

If you like my models, please support ❤️ me for more computing resources to do finetuning work.