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
Flux2-Klein-9B-True-Q8_0
FAQ
Comments (50)
Why would one use this over just base klein 9b? is it faster? I'm asking because the point of the distill is the 4 step generation, if you now have to go up to 25 steps, it's just like using base?
You can take a close look at the explanation on the official model page. The Base model is suitable for training but not ideal for direct use. As for the 4-step distilled model, due to the small number of steps, it is over-distilled, making the details prone to collapse, and it isn’t particularly friendly for LoRA adaptation either. Moreover, the clarity and realism of the 4-step model are also not very good. Of course, which model you choose to use is entirely a matter of personal preference, so just pick the one you like.
this one is faster by design ;)
Distilled is faster (by less steps) but reasonable quality as it has many processing LORAs and effects already baked in. Base you need more steps as the "speed-up" lighting and other LORAs arent already in the model so it needs more steps to do things the loras generally take care of but (in theory) provides better generations.
In theory, base at more steps will be the best quality and allow the most control as you can add it whatever LORAs you like - but it will take much longer as the step count needs to be much higher.
Try using base with this turbo lora, its been working great for me at 6 steps and providing much better image quality (I've found base easier to make do NSFW content). I'll usually do a quick preview run at 6 steps and then when I'm happy I do a really long running 20 step run (or 50 if I decide not to use the turbo lora) for production. https://civitai.com/models/2324315/klein-4b9b-base-to-turbo-lora
It won't work with distilled tho FYI.
@MiddlingMaker Thanks for the sharing. 👍
Just tried the fp8 model for a bit. Seems a lot better at retaining pale skin which is important for me. Thank you for posting this!
By the way, it would be super helpful if you recorded your screen in the image portion while doing the comparison. I've done this for one of my workflows and it really helps people understand how the model differs from base Klein.
Here's what I mean:
https://civitai.com/images/114069179
Thank you a lot for sharing your positive experience. 👍✌️
Life is too complicated You do not need t o complicate it even more just to show every one how You are .Make it simple - make You the Best.
Thanks a lot.
Hi most of your models only focus on Asians. Is this also same or is it friendly for all races?
Haha, where did you hear that? My model has no restrictions in this regard, you just need to specify the area in the prompt, try it yourself, pls.
@wikeeyang Thanks for the updates. I'll surely try this, base on the gallery i see other races like western and europeans. We must embrace all races and colors : )
@haidensd58757 Yeah！Of course！
Every preview image is a female. What is the percentage of young males in the training data?
how to run on forge neo? what encoder is used?
Forge Neo should support Klein model, text_encoders for 9B is qwen3-8b, for 4B is qwen3-4b
I know I am 15+ days late, but you need Qwen3 8B text encoder and FLUX.2 VAE
"Mixed" is not supported on Forge Neo.
@Hyokkuda Thanks a lot!
Thank you so much for sharing this and the workflows 🙏
I personally find bf16 and fp8 quality almost the same when doing image edit in wuli's workflow
Yes, Thanks a lot for your feedback.
for some reason my flux 9b base gguf is deep fried meanwhile a gguf for this model actually works, so hell yeah!
Haha, what is the tools to converted your GGUF version?
Any way to get breasts to not be gigantic? Always get "big naturals" lol.
What happened? I don't know... Haha... 😂😂😂
oh..I get it all the mammary glands are very large.
@vqxss73266679 This may be cause by some LoRA? 🤣🤣🤣
I found using the turbo lora(both r64/r128 at strength 0.6 or 1.0) results in fried images in i2i(tried 4/8/12 steps, cfg 1/2/4, sampler euler/euler a/res_2s/res_2m/res_multistep), other loras don't have this issue. Seems like the turbo lora is not compatible with this checkpoint.
I tested it and there were no any issues. Please refer to the recommended workflow (https://www.bilibili.com/video/BV1arkgBWE2p/) on the model description page, and simply add turbo LoRA load node behind the model load node.
I have performed further testing and it looks like the issue is stemming from generating image larger than 1 megapixels and using the turbo LoRa at strength over 0.6, all the configurations I tested works fine when the output is at 1MP. At lora strength over 0.6 and 2MP the generations starts to have a lot of noise introduced. Example of an completely affected generation(single 1MP reference, 4MP output, 1.0 turbo lora strength, 12 steps, Euler/Flux2Scheduler, cfg 1): https://imgur.com/a/7LDV1me
I also tested Flux 2 Klein 9B Base FP8, it did fine at generations over 1MP. Just seems like this checkpoint does not need that high of a turbo boost than the base checkpoint.
@ehsog Tks a lot for the feedback. 👍
I tried the Turbo LoRA R128 with this Checkpoint (I have the Q8_0 version) , after some testing I found that using a maximum strength value of 0.4 for the Turbo LoRA will make it work nicely & give reasonable results, higher than 0.4 starts making quality worse & gives excessive overbright .. but in general I am happy with this Checkpoint Klein-9B-True , it gives nice results especially for image editing, & it adapts well to a wide range of LoRAs for T2i image generation. 👍🏽👍🏽
@hazzoom82659 Thank you a lot for sharing this, thanks again!
The anyMOD Turbo LoRA was recommended for use at about 0.25 strengths, and I have been successful in using the True-V2 model at 1MP and 4MP with no problem, either with LoRA W=0.0 or W=0.25.
If I understand correctly, there is a Turbo LoRA, possibly with a weight about 0.7, already baked into the checkpoint, or a a process equivalent to this effect has been performed during training and merging of the checkpoint. Is my understanding correct (if you don't mind me asking :-) )? @wikeeyang In any way, because of this or another reason, the Turbo LoRA is not strictly necessary - at least for me! :-) .
I need to do some more trial-errors for image edits and I2I though ....
Thanks again for the great checkpoint!
@informedviewz 👍🤣, It's roughly like this, but the actual process is more complicated. anyMOD Turbo LoRA is extracted from the Klein-9B 4-step model through the Base model, whereas my model is initially based on a mixture of some weights from Klein-9B and 9B-Base, allowing it to generate images in about 20-25 steps at cfg=1.0, and then undergo fine-tuning to improve texture and editing capabilities, reaching a publishable level in the end. These details are thoroughly explained in my Flux1-DedistilledMixTuned series of models. So, if you feel that it seems to contain part of Turbo LoRA's weights, that's normal.
@wikeeyang Thanks so much for the thorough explanation. Really appreciate the time! Sounds interesting, and makes sense :-)
I had major problems creating a draft with few steps and a final image with 25 steps using the untrained Checkpoint, so I haven't yet made the move from Flux 1 Dev to Flux 2 Klein, but now it's really fun to create new images.
Yeah, Flux.2-Klein is a very excellent model. 👍
Hey do u have a workflow with 2 images? I tried to drag the picture u uploaded but it just add an image node.
Pls ref to model description page, I put a link of the workflow by Mr. WuliDaXiong
you can download it from https://www.runninghub.ai/post/2013285336390045697 , pls
There is a lora called realistic-prompt_000008500.safetensors in the workflow, I am not able to find this, can you provide a link? @wikeeyang
@Cerner I don't know this lora should be which one, this may be the author's own naming, pls ask to dx8152. or try to this one: https://huggingface.co/dx8152/Flux2-Klein-9B-Enhanced-Details/resolve/main/realistic.safetensors
Wow, I was sleeping on this. Getting some crazy good outputs compared to the regular Klein model. Thanks for the work, keep it up!
Thanks a lot!
He just uploaded the V2 version. It seems even better.
@nightcall Thanks a lot! 🤝👍
@wikeeyang thanks to you for your amazing work!
@nightcall Thanks for letting me know! I forgot to click like so didn't get the notification. Excited to test!
Can I use the fp8 true version and leave steps as the normal/original flux 2 klein 9b?
Sure, you can add Turbo or Distillation LoRA to reduce steps, for example: anyMODE's rank128/rank64 Klein Turbo Lora, you could try set lora weight 0.2-0.3 for 4 - 8 steps. for Distillation lora, weight about 0.4 - 0.6, you can try it by yourself.