The FLUX.2 [klein] model family are our fastest image models to date. FLUX.2 [klein] unifies generation and editing in a single compact architecture, delivering state-of-the-art quality with end-to-end inference in as low as under a second. Built for applications that require real-time image generation without sacrificing quality.
Originally Posted;
Description
9B-KV is an optimized variant of FLUX.2 [klein] 9B with KV-cache support for accelerated multi-reference editing
FAQ
Comments (20)
If I understand correctly, in order for this kv-cache to work, you need to update comfyui to the latest version (0.17) and use a new node. This is probably worth mentioning in the description of the KV model. However, judging by the issues section on github, version 0.17 brings new bugs in addition to features. Therefore, if stability is more important, then it makes sense to wait. P.S. Thanks for adding the fp8 version too.
Whats up with KV? Wht is it?
this is what i could find: "This version uses "KV Cache" to accelerate image editing."
think of it as a turbo version of the other kleins, only requires 8 steps usually.
kv cache is a saving VRAM technique for LLM. it compress the context more making the reference images of klein less VRAM hungry
I dig this one, responds really well to even basic prompts
wf please?
on my gtx 1080ti it works very well..
tried image gen, image to image workflow and work bestttt
why the 9b and 9b base have exactly the same size ?
Why wouldn't they be?
Everyone is working so hard to show us all how hard they ALL fucked up
Question for everyone using this model, would you rather use the distilled model or the base model with a turbo LoRA when running locally? Would like to know why
base is better imo
@New_Name Ok, can you tell me how you think Base is better then the distilled model?
@elevendr if you want to fine tune your prompt and images I say go with the base. Distilled is also good for general use case, if you use loras then it can produce some amazing results. TL;DR - Depends on your use case.
I'd call it the new generation SDXL. It boasts both incredibly fast image generation speeds and decent image quality. Even better, it's smaller than Flux2 and still offers decent image editing capabilities. I believe it can replace the SDXL. However, it seems to have higher VRAM requirements; apparently, the minimum requirement for the 9B is 16GB, but in reality, my 12GB graphics card still runs it smoothly. Fortunately, there's a gguf version available for more users.
For first-time users, here's a little tip: for text-to-image generation, Flux 2 Klein 9B Turbo is actually the best choice. Flux 2 Klein 9B Base is not as good as Flux 2 Klein 9B Turbo. Flux 2 Klein 9B Base is used for training LoRa and fine-tuning. The Turbo version has better image quality.
when you say Flux 2 Klein 9B Turbo, do you mean Flux 2 Klein 9B Kv?
@kaddzie What I call Flux2Klein Turbo is actually the Flux2Klein 9B. The Flux2Klein KV version is specifically designed to accelerate image editing. If you have low VRAM or just want to speed up the generation process, you can choose the Flux2Klein 9B Int8 version
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.