Flux2-Klein-9B-True-V2
***请不要下载该模型文件，烦请移步 HF 或 ModelScope 下载***
Due to network, Pls Download from : Huggingface.co, Modelscope.cn
V2 版本模型在 FLUX.2-Klein-9B-True-V1 的基础上再次进行了全量微调，相比 V1 模型，在如下方面得到了改善和提升：
图像的真实感和质感进一步改善，基本接近香蕉(Nano Banana)的水平。
提示词遵循和还原能力，参数适配性和LoRA兼容性进一步改善，。
成图结构更归整、合理，美学方面也有一些改善。
图像编辑能方面，提示词遵循更好，偏移和偏色也有一些改善。
具体效果参见如下对比图片。
The V2 version of this model has undergone a full fine-tuning based on FLUX.2-Klein-9B-True-V1. Compared to the V1 version, it has some improvements and enhancements as below:
The realism and texture of images have been further improved, basically reaching the level of Nano Banana.
Prompt adherence and restoration capability, parameter adaptability, and LoRA compatibility have been further enhanced.
The image structure is more organized and reasonable, with some aesthetic improvements as well.
In image editing ability, prompt adherence is better, pixel shift and color bias have also seen some improvements.
For specific results, please refer to the comparison images as below:
使用方法：
The same with V1 vesrion.
for Turbo Lora : anyMODE's rank128 / rank64 Civitai or Huggingface, weight 0.2 - 0.3 for steps 4 - 8.
for Distill LoRA : vafipas663's rank128 / rank256 Huggingface, weight 0.4 - 0.6 for steps: 4 - 8.
for Color and Shape correct of Image, you can use plugin of : ComfyUI-LCS
=======================================
Flux2-Klein-9B-True-V1
本模型为 FLUX.2-klein-9B 的微调模型，相比官方模型，去蒸馏、更清晰、更写实，模型编辑能力更精准，大大减少蒸馏模型过低步数带来的细节崩坏问题。具体效果参见如下示例图片，本模型易于使用、参数适配性和LoRA兼容性良好，如果结合 Real-Qwen-Image-V2 模型的中文字体渲染和多风格图像生成能力，AI生图创作如虎添翼。
This model is a fine-tuned version of FLUX.2-klein-9B. Compared to the official model, it is undistilled, clearer, and more realistic, with more precise editing capabilities, greatly reducing the problem of detail collapse caused by insufficient steps in distilled models. For specific effects, see the example images below. This model is easy to use, has good parameter adaptability, and is compatible with LoRA. If you combined with the Real-Qwen-Image-V2 model's Chinese font rendering and multi-style image generation capabilities, The image creation will greatly enhanced.
使用方法：
basic workflow pls refer to: image
inference cfg: 1.0；
inference steps: 20 -30；
edit or inpainting steps: 10-25;
or add turbo lora about steps: 8-12;
sampler / scheduler: euler / simple ，or any other。
Also on: Huggingface.co, Modelscope.cn
单区域编辑示例：
多区域+参考图编辑示例：
附加提示：
Turbo Lora Link：anyMODE's rank128/rank64 Klein Turbo Lora。
推荐使用 wuli大雄oO 大佬的 Flux2 Klein 超强多功能工作流，功能齐全又方便。
模型许可：
本模型请在遵循 black-forest-labs 公司非商业许可 (FLUX Non-Commercial License) 下使用。
This model falls under the FLUX Non-Commercial License.
Description
请不用下载该权重文件，请移步 HF 或 Modelscope 下载。
Don't use this safetensors, Pls download from HF or ModelScope.
FAQ
Comments (62)
I will test it more but so far I seem to be getting better results with v1 than v2.
May be some different workflow or parameter. haha...😂🤝
I agree , absolutely love your model , but V1 is amazing , V2 is messing up my character loras , no matter what Wf i use . V1 is a beast.
@cristianlyon67811 Thanks a Lot! ❤️💐👍
You're shooting yourself in the foot by putting this under Flux.2 Klein 9B Refiner lol. That's bascially not searchable.
Haha... Thanks...
How do you use this model along side the main model flux2Klein9BTrue_v10Bf16 safetensor checkpoint? coz CFG guider or Ksampler can only handle one diffusion model. Is this lora?
Sorry! Pls don't use V2_bf16 safetensors file, due to network problem, pls download V2 from HF or Modelscope site under my repo. 🤝💐
@wikeeyang Ok bro thanks for suggestion. What do you think, is the image quality realism and prompt understanding better in flux2Klein9BTrue_v10Bf16 17GB or the Flux2-Klein-9B-True-v2-fp8mixed.safetensors 9gb model? Does higher file size means better?
@haidensd58757 Yes, if you have good GPU VRAM, high precision will better, but if your VRAM not so good like above 24GB, then fp8mixed will better, due to this is mixed precision, then it's quality compare with fp16 less then 3-5%, almost the same.
@wikeeyang Alright bro thanks for the clarity. but why is this current Flux2-Klein-9B-True v2 f16 checkpoint only 18MB? it should be 17gb insn't?
@haidensd58757 due to network, I can't upload v2-bf16.safetensors file to Civitai site successfully, but in civitai if not one safetensors weight can't publish version of model, then I just upload a tiny file not the true v2-bf16.safetensors for model version publish, understand? bro...🤣🤣🤣🤝, I commit pls download model weight file from HF or Modelscope site under my repo...
@wikeeyang Oh i got it and understand your struggle. yes HF link is the rescue. And thank you once again, really appreciate your hard work man. 🫡🏆
@haidensd58757 you are welcome, bro. 🤝✌️
Very impressive! I tested 9b fp8: Klein base / v1 / v2. v2 is my new main model.
Thanks a Lot! 👍💐❤️
I downloaded the BF16 quant of the V2 model from huggingface, but when I run it I get this error:
ERROR: Could not detect model type of: F:\AI Models\diffusion_models\Flux2\flux2Klein9BTrue_v20Bf16.safetensors
The original FLux.2 Klein 9B models by Black Forest Labs work perfectly fine. I'm using ComfyUI.
Haha...That's so strange! All model file had tested in ComfyUI newest version.
@wikeeyang This is strange because I've just downloaded V1 and it works fine. :/ I will try downloading V2 again. By the way, the model description says that the model is undistilled, but this is not the case. The model works properly only when CFG is set to 1.0 which means that this is a distilled model. Did you upload a wrong file by mistake? If not and this was intended, the description should be corrected.
@mmdd2543 De-Distilled of the models is my model feature.
The solution to this problem is to fully update your ComfyUI and all node plugins. Click Update All in the Manager.
Hello, i can't download the checkpoint, v2 bf16 is now 18MB, not 16GB
@hoyjooh170 Thanks a lot.
sorry for this absolutely noob comment, i didn't took time to read the whole post... but if it helps lazy folks like me, thank you @hoyjooh170
UP your model is absolutely amazing! Truly impressive — I'm really enjoying it.大佬您的模型太棒了！很赞，正在体验。
Tks a lot! your likes is my greatest motivation. 谢谢！大家喜欢，就是我继续调制和优化模型的最大动力。
The Flux series models, produced by Black Forest, are open-source but not for commercial use. have Flux.1 / Flux.2 / Flux.2-Klein etc, pls ref to https://huggingface.co/black-forest-labs, it's Pro and Ultra or Max model not open source.
V2到底是有没带加速？为什么我没加加速lora 4步或者25步都可以
Besides the model, what do I need to install? VAE, Text Encoder, etc.???
you need the Text Encoder and VAE part from Flux2-Klein-9B, you can get it from comfy-org repo, and pls ref to workflow which had provided under my repo.
Thanks for making it! It's an amazing model.
I have found CFG=2.0, Steps=16, Euler_Ancestral (or Euler), and 0.25xTurboLoRA to give excellent results at 4K (2048x2048).
The same settings at 1024x1024 and 4 Steps give a reasonable fast preview for draft-prompting.
It clearly outperforms the original F2K9B (Base or distilled) with or without Turbo LoRA. Also, it does actually compete with and even wins the competition against NanoBanana 1/2/Pro, Imagen 4, and GPT Image 1.5 in terms of realism and quality of generation. Only the world knowledge (limited by the text encoder and original model size) is behind these commercial models.
Much thanks for your recognition of the model, and also much thanks for sharing your valuable experience. Bro. 🤝💐✌️👍
大佬，能更新Q6_K和更小文件的GGUF版本吗？
A preview: Today my friend lrzjason's consistency LoRA, https://huggingface.co/lrzjason/Consistance_Edit_Lora/resolve/main/f2k_9B_lcs_consist_20260415.safetensors, will be released. When it combined with my V1 or V2 models, it can greatly improve the Klein model's drift and color bias issues in I2I.
I watched his personal test results, and there is a significant improvement compared to f2k_9B_lcs_consist_preview_20260328 version.
@wikeeyang Thanks but The file is not found, Can you update the link or tell us where the final version will be available ?
@hazzoom82659 Wait a bit, he is making a video now, and will upload the model later. 🤣🤣🤣
Can anyone help me by answering this?
This is the first Klein model I've tried, and it's more than incredible, especially version 2. I couldn't generate NSFW content beyond bare breasts, so I wanted to know if it's possible to do it with version 2 while maintaining the same quality?
Thank you so much for taking the time to work on this model and for sharing it! I've been doing lots of experiments using it in different workflows and it's simply incredible! 再谢了，朋友.
You're welcome and thanks a lot for your appreciation. Your liking is my greatest motivation. ✌️🤝❤️
bf16 v2 file is 19mb...
I comment don't use it, please download from https://huggingface.co/wikeeyang/Flux2-Klein-9B-True-V2
Amazing model, but I've noticed that V2 has less prompt and edit adherance compared to V1 (prompt can be fixed with high cfg and negative turbo lora), but edits like face swaps and keeping body shape from a reference is less consistent on V2
EDIT: Acutally V2 seems more realistic on the face swaps but you need custom nodes to control the weight of the latents
Thank you for your feedback and sharing, thank you again. Yes, there are some differences between V1 and V2.
This works great with turning anime, 3d images into realistic one, but my only issue is the faces. Turning cartoon looking images into realism usually outputs Asian faces even if I tell it to use European, English looking people with TextEncodeQwenImageEditPlusCustom.. Is there a way to get it to fix this?
Pls try my Flux2-Klein-9B-True-V1, V2 has made some optimizations for Asian faces.