Flux 2 Klein Support #2279
Replies: 15 comments 16 replies
-
|
Thanks for your amazing work , does it support inpaint or simulated seamless inpaint using masked edit? |
Beta Was this translation helpful? Give feedback.
-
|
I tried using edit for inpaint, but it either failed completely or the result was not better than without the reference. So it's just regular inpaint without special handling. It works well enough for fill. To add/remove objects and similar stuff it's better to use edit mode with specific instructions. |
Beta Was this translation helpful? Give feedback.
-
|
Very nice, thank you. (edit, deleted question, I was not paying attention, the predefined style is of course in the Styles list) |
Beta Was this translation helpful? Give feedback.
-
|
The results are quite convincing! LOL You must be kidding, the results of 9b are amazing. In major of cases much better than Qwen edit... |
Beta Was this translation helpful? Give feedback.
-
|
9b really is wonderful |
Beta Was this translation helpful? Give feedback.
-
|
Base or distilled? |
Beta Was this translation helpful? Give feedback.
-
|
i use distilled but base is also good |
Beta Was this translation helpful? Give feedback.
-
|
Does Flux require newer graphic cards? FP8 etc only runs on RTX cards right? |
Beta Was this translation helpful? Give feedback.
-
|
If your graphic card does not support fp8 comfy will automatically upcast precision to fp32 for non-rtx cards , in my case I'm running klein-4b-fp8 on a gtx1660ti and it runs fine , it's very slow for sure but the model being a 4-step distilled model helps a lot with it. |
Beta Was this translation helpful? Give feedback.
-
|
ah ok, ty! |
Beta Was this translation helpful? Give feedback.
-
|
These models are the first ones that I've run on my Apple Mini that produce good output in a reasonable amount of time, really impressive. And thanks again to Acly for this plugin, using this inside Krita makes for a natural workflow without having to learn ComfyUI. It's a great project, but I'm interested in creating images, not learning how to wire up a processing toolchain. |
Beta Was this translation helpful? Give feedback.
-
|
First, thank you for your amazing work! Truly. A question regarding custom sampler settings. Clicking "edit custom presets" opens up a .json file with the current sampler settings presumably. And yet, that list does not match what I see in the Quality preset drop-down inside Krita. Working with Flux Klein, I find myself switching between Flux 2 Euler and Euler Beta a lot. And each time I have to manually adjust the CFG and steps sliders down from their default settings which is a bit of a pain. Ideally I'd want to type in new default values for the samplers, but currently I can't see how that can be done. Unless I'm missing something significant here. |
Beta Was this translation helpful? Give feedback.
-
|
You can create new samplers and append them to the json file with a custom name and desired defaults you need, you don't have to edit existing ones every time. |
Beta Was this translation helpful? Give feedback.
-
|
Yes, that's essentially what I'm doing. Clicking "Edit custom presets" in the interface opens the json-file. But that file (residing in the C:\Users...\AppData\Roaming\krita\ai_diffusion\presets) is not corresponding with what is visible in the interface, different values. Could one delete the old json-file and create a new one? |
Beta Was this translation helpful? Give feedback.
-
|
you wouldn't need to delete the file and create a new file for it to work, one thing to keep in mind that you must restart krita after modifying the |
Beta Was this translation helpful? Give feedback.
-
|
Has anyone had success using this model for multi-image editing? I have tried to use the 9b model with a line art layer to add a design to a base image ... this sort of works except if the base image is a color image the edit result also randomly recolors the image and/or changes the color balance, regardless of negative prompting to the contrary. Maybe this functionality only works for generate? |
Beta Was this translation helpful? Give feedback.
-
|
I haven't used negative prompts since I only use the distilled version , but I had decent success with multi image edit in klein-4b, what you need to do is reinforce what you you want to not change using negative instructions in your the positive prompt , for example "do not change the color balance of the image" or" do not change the color of the image" , "change X from image 1 to Y from image 2 and output image 1 without changing anything else, keep color and white balance and composition exactly the same" |
Beta Was this translation helpful? Give feedback.
-
|
Updated the original post for 1.47.0:
|
Beta Was this translation helpful? Give feedback.
-
|
Hello Acly! |
Beta Was this translation helpful? Give feedback.
-
|
If I try to load a 9b model, I get a warning saying that the text encoder for this model is not installed: qwen3:8b, qweb3-8b I tried downloading Qwen3-8B-Q4_K_M.gguf from huggingface, but then, when I tried to create an image Comfyui gave an error |
Beta Was this translation helpful? Give feedback.
-
|
Any idea why using a reference control layer for style with a prompt like: Transfer the style from image 2 Results in both images being combined? The original image does get the style treatment but the bg is filled with image from the control layer. |
Beta Was this translation helpful? Give feedback.
-
|
The model does not deduct noise, it immediately comes with a best guess. The reference is a strong indicator. I haven't tried this, but use language like "the picture should be in the style of the reference image without copying it" in the prompt. |
Beta Was this translation helpful? Give feedback.
-
|
IME Flux 2 can't do style transfer from one image to another. It can do style transfer where you have 1 input image and describe the style in text, or it can generate image based on text in the style of a reference. |
Beta Was this translation helpful? Give feedback.
-
|
I'm not able to set Flux2 Klein as the linked edit model and i can't see why. |
Beta Was this translation helpful? Give feedback.
-
|
can you add selection also for 9b model? |
Beta Was this translation helpful? Give feedback.
-
|
Seconded. It would be nice. Looks like it also uses a different text encoder. |
Beta Was this translation helpful? Give feedback.
-
|
You mean an automated download option? 9B has a different license, you need to go to https://huggingface.co/black-forest-labs/FLUX.2-klein-9B and accept some stuff to download it (and the plugin can't just download it either without a login) |
Beta Was this translation helpful? Give feedback.
-
|
Can Krita AI work with a FLUX.2-small-decoder - a faster, drop-in replacement for their standard decoder. ~1.4x faster, Lower peak VRAM https://huggingface.co/black-forest-labs/FLUX.2-small-decoder Compatible with all open FLUX.2 models: FLUX.2-klein-4B |
Beta Was this translation helpful? Give feedback.
-
|
With latest (custom installed) ComfyUI, yes. |
Beta Was this translation helpful? Give feedback.
-
|
Oh, I thought I can do it directly |
Beta Was this translation helpful? Give feedback.
-
|
Has anyone else noticed that Expand with Flux 2 Klein (distilled) often places random objects from the original image into the expanded area? Objects that don’t belong there? Example: In this case, it’s a detached branch, but it could just as easily be a disconnected tree trunk, a fragment of the man’s head, or any other element from the original image. |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
-
FLUX.2 [klein]
Last updated: 1.47.0
Features
Requirements
The managed install can run Flux 2 Klein models since plugin version 1.46.0, and offers automatic download since 1.47.0. So you can skip manual downloads below, and just install it from the settings connection tab.
For custom ComfyUI installs remember to update:
Model Download
To get started you need to download 3 model files: the diffusion model, the text encoder, and the VAE. Locate the folder where you installed the ComfyUI server, and place the files in the sub-folder specified below.
Diffusion Models
Download one of these. They go into the
/models/diffusion_modelsfolder.Text Encoder
Flux 2 Klein uses Qwen 3 as text encoder. It is the same model as the one used by Z-Image, so if you already have that working, you don't need to download it again.
/models/text_encodersVAE
/models/vaeHow to Use
The plugin comes with a built-in "Flux 2 Klein" style. Select it and start generating images as usual.
Image Editing
Flux 2 Klein has built-in edit capabilities. You can switch the "Generate" button to "Edit" mode without a separate model. In this mode you write instructions on how to modify an existing image, rather than generating a new one. You can also attach individual layers as additional reference image inputs. See Edit Models for more info.
Control Layers
Image editing capabilities also include typical control inputs, like sketches, line art, canny edge and character pose. Selecting any of those control layers in "Generate" mode will setup the appropriate instructions. The results are quite convincing!
9B and Base models
The larger 9B variant and the base models are also supported. You can select the "Flux 2 Klein" style in the style settings, duplicate it, and then switch to the other models. If you are using a "Base" variant the sampler settings should be adjusted to ~20 steps and ~4.0 guidance.
Additional Information
Beta Was this translation helpful? Give feedback.
All reactions