Flux Kontext Dev (local)

[Legolas18]

Well, finally Flux.1 Kontext [Dev] has hit the open source for local use and it is...really good if used in the proper way (it is an AI tool, btw).

You basically feed one or more images with a description of what you want the model to do.

I find it very useful to:
- Change style (and there are many, but you need to experiment to find the one that suits you)
- Quickly add backgrounds
- Character consistency

Personally, I like anime or Pixar with a Pastel color palette (easier to edit in Photoshop than photos, tbh).

NOTE: I don't think it does photography too well (EDIT: see later), but I am not interested in that style, so I haven't tested it too much.

Don't expect to get everything perfect in one go, though. Using an output image as a new input to get a new output, is a valid strategy.

Also, it seems to be good at complex character interactions too (if prepared in Daz)

Here are a few example transformations (first daz image, then style + background change):

EXAMPLE 1:

EXAMPLE 2:

EXAMPLE 3:

[Legolas18]

UPDATE: Of course, you can also use it with style loras, for even more styles:

[Frank21]

Yeh I think if you just want to change the style a style-lora will work equally well. I also see people making loras that could be easily done with prompting.... my rule with loras is a few as possible. Yesterday I spent hours just trying to get a character to lie flat on his back, then gave up. There are some concepts that Flux just doesn't understand, apart from anything to do with rumpy-pumpy.

https://medium.com/@joe.richardson.iii/the-curious-case-of-the-gen-ai-wine-glass-problem-f43778bbe890

[Legolas18]

@frank22 here i am talking about Kontext specific loras, I don't think they are compatible with flux-dev.

And also, Kontext is more of an image editing model, i would use it to generate limited stuff on its own. I think it could be used to make visual novels, product placements, etc.

Today, the company also released lightweight variants of the model to be compatible with lower end gpus too (on huggingface).

As for the glass of wine thing...this model should be able to handle it, since you can input in it several images. So just put as an input a full glass of wine and it should work.

Kontext is certainly fun to play with, especially useful if you use it with Daz, since you can then control poses and controlnet kinda becomes useless.

[Frank21]

Yeh I went off on a tangent there, sorry. I've tried Kontext and it does work well for changing clothing. Are you saying that Kontext is now available at Huggingface for DL and to use locally?

[Legolas18]

@frank22 No worries, sometimes I am the tangent 🙂

And yes, it is available for local use on huggingface:

FP16 (quite heavy, but accurate):
https://huggingface.co/black-forest-labs/FLUX.1-Kontext-dev/tree/main

FP8 Scaled:
https://huggingface.co/Comfy-Org/flux1-kontext-dev_ComfyUI/tree/main/split_files/diffusion_models

Quantized version for less VRAM usage:
https://huggingface.co/bullerwins/FLUX.1-Kontext-dev-GGUF/tree/main

Also, I installed SAGE Attention 2++ on my pc and, with a ComfyUI node, it goes faster at a marginal loss (well, kontext looks better with it, while the regular Flux, a bit worse).

[Author]

Posts