Any plans to get into working with the Flux 'Kontext' version, the editing models? I think the use cases of such prompted image editing is just wildly huge. Their demo blew my mind, although I haven't seen the quality of the open weight version yet. It is also a 12B distill.

Hi. Thanks for this. What is your goal of doing so? From a business standpoint. Or is it purely altruistic?

Regarding the P(.|photo) vs P(.|minimal) example, how do you actually decide this conflict? It seems to me that photorealism should be a strong default "bias".

My reasoning: If the user types in "a cat reading a book" then it seems obvious that the result should look like a real cat which is actually reading a book. So it obviously shouldn't have an "AI style", but it also shouldn't produce something that looks like an illustration or painting or otherwise unrealistic. Without further context, a "cat" is a photorealistic cat, not an illustration or painting or cartoon of a cat.

In short, it seems that users who want something other than realism should be expected to mention it in the prompt. Or am I missing some other nuances here?

I need model for other language than english

The model looks incredible!

Regarding this part: > Since flux-dev-raw is a guidance distilled model, we devise a custom loss to finetune the model directly on a classifier-free guided distribution.

Could you go more into detail on the specific loss used for this and any other possible tips for finetuning this that you might have? I remember the general open source ai art community had a hard time with finetuning the original distilled flux-dev so I'm very curious about that.

From a traditional media production background, where media is produced in separate layers, which are then composited together to create a final deliverable still image, motion clip, and/or audio clip - this type of media production through the creation of elements that are then combined is an essential aspect of expense management, and quality control. Current AI image, video and audio generation methods do not support any of that. ForgeUI did briefly, but that went away, which I suspect because few understand large scale media production requirements.

I guess my point being: do you have any (real) experienced media production people working with you? People that have experience working in actual feature film VFX, animated commercial, and multi-million dollar budget productions?

If you really want to make your efforts a wild success, simply support traditional media production. None of the other AI image/video/audio providers seem to understand this, and it is gargantuan: if your tools plugged into traditional media production, it will be adopted immediately. Currently, they are tentatively and not adopted because they do not integrate with production tools or expectations at all.

I recently ran a training experiment using the same dataset, number of steps, and epochs on both Flux Dev and Flux Krea models.

What stood out to me was that Flux Dev followed the text prompts more accurately, whereas Krea’s generations were more loosely aligned or "off" in terms of prompt fidelity with deformations in body type and the architecture.

Does this suggest that Flux Krea requires more training to achieve strong text-to-image alignment compared to Flux Dev? Or is it possible that Krea is optimized differently (e.g. for style, detail, or artistic variation rather than strict prompt adherence)?

Curious if anyone else has experienced this or has any insight into the differences between these two. Would love to hear your thoughts

thanks for doing this!

what does " designed to be compatible with FLUX architecture" mean and why is that important?

Can you point to a URL with the tests you’ve done?

Also, FWIW, this model focus was around aesthetics rather than strict prompt adherence. Not to excuse the bad samples, but to emphasize what was one of the research goals.

It’s a thorny trade-off, but an important one if one wants to get rid of what’s sometimes known as “the flux look”.

Re: Wan 2.2 I’ve also been reading of people commenting about using Wan 2.2 for base generation and Krea for the refiner pass which I thought was interesting.

Quick napkin math assuming bfloat16 format : 1B * 16 bits = 16B bits = 2GB. Since it's a 12B parameter model, you get around ~24GB. Downcasting to bfloat16 from float32 comes with pretty minimal performance degradation, so we uploaded the weights in bfloat16 format.

A parameter can be any size float. Lots of downloadable models are FP8 (8 bits per parameter), but it appears this model is FP16 (16 bits per parameter)

Often, the training is done in FP16 then quantized down to FP8 or FP4 for distribution.

That's a good ballpark for something quantized to 8 bits per parameter. But you can 2x/4x that for 16 and 32 bit.

Off topic but did you really hide scroll bars on the website? Why...?

  .scrollbar-hide {
    -ms-overflow-style: none;
    scrollbar-width: none;
  }

Describing it as "Octopus DJ with no fingers" got rid of the hands for me, but interestingly, also removed every anthropomorphized element of the octopus, so that it was literally just an octopus spinning turntables.

Have you tried building one irl? I can't find a video of a double one

We have not added a separate RTX accelerated version for FLUX.1 Krea, but the model is fully compatible with existing FLUX.1 dev codebase. I don't think we made a separate onnx export for it though. Doing 4~8 bit quantized version with SVDQuant would be a nice follow up so that the checkpoint is more friendly for consumer grade hardware.

You can find details about the license here: https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/mai...

In a nutshell, it follows the same license as BFL Flux-dev model.

Yup... We have the following: https://www.krea.ai/pricing

Though we wanted to keep this technical blogpost free from marketing fluff, but maybe we over-did it.

However, sometimes it's hard to give an exact price per image, as it depends on resolution, number of steps, whether a LoRA is being used or not, etc.

The license, as I understand, applies only to the model, not to the images produced? Otherwise they should respect the license of images it was trained on.

for the love of the game, not for corporate greed

We actually have a GitHub repository to help with inference code.

Check this out: https://github.com/krea-ai/flux-krea

Let me see if we can add more details on the blog post and thanks for the flag!

I love owls. Photorealism was one of the focus areas for training because "AI look" (e.g. plastic skin) was biggest complaint for FLUX.1 model series. Photorealism was achieved with both careful curation of finetuning and preference dataset.

Thanks!

From a business point of view, there are many use-cases. Here's a list in no particular order:

- You can quickly generate assets that can be used _alongside_ more traditional tools such as Adobe Photoshop, After Effects, or Maya/Blender/3ds Max. I've seen people creating diffuse maps for 3D using a mix of diffusion models and manual tweaking with Photoshop.

- Because this model is compatible with the FLUX architecture, we've also seen people personalizing the model to keep products or characters consistent across shots. This is useful in e-commerce and fashion industry. We allow easy training in our website — we labeled it Krea 1 — to do this, but the idea with this release is to encourage people with local rigs and more powerful GPUs to be able to tweak with LoRAs themselves too.

- Then I've seen fascinating use-cases such as UI/UX designers who prompt the model to create icons, illustrations, and sometimes even whole layouts that then they use as a reference (like Pinterest) to refine their designs on Figma. This reminds me of people who have a raster image and then vectorize it manually using the pen tool in Adobe Illustrator.

We also have seen big companies using it for both internal presentations and external ads across marketing teams and big agencies like Publicis.

EDIT: Then there's a more speculative use-case that I have in mind: Generating realistic pictures of food.

While many restaurants have people who either make illustrations of their menu items and others have photographers, the big tail of restaurants do not have the means/expertise to do this. The idea we have from the company perspective is to make it as easy as snapping a few pictures of all your dishes and being able to turn all your menu (in this case) into a set of professional-looking pictures that accurately represent your menu.

Thank you! Glad you find it helpful. The model is focused on photorealism so it should be able to generate most realistic scenes. Although, I think using 3D engines would be more suitable for typical cases for robotics training since it gives you ground truth data on objects, location, etc.

One interesting use case would be if you are focusing on a robotics task that would require perception of realistic scenes.

Hey thanks! I’ll ask Sangwu to hop here to answer this and give a more research-oriented answer

yeah data really is everything that was the number one lesson from this whole project

We used two types of datasets for post-training. Supervised finetuning data and preference data used for RLHF stage. You can actually use less than < 1M samples to significantly boost the aesthetics. Quality matters A LOT. Quantity helps with generalisation and stability of the checkpoints though.

hello Erwann, great work! I have a very technical question just for you: how are you today?

Copying and pasting Sangwu’s answer:

We used two types of datasets for post-training. Supervised finetuning data and preference data used for RLHF stage. You can actually use less than < 1M samples to significantly boost the aesthetics. Quality matters A LOT. Quantity helps with generalisation and stability of the checkpoints though.

Haha thanks! Do you happen to have a use-case for it?