A lot of coffee for sure. Regarding the training cost, it's hard to give a good estimate because we used a shared kubernetes cluster with inference + research workloads.
The highest quality finetuning data was hand curated internally.
I would say our post training pipeline is quite similar to SeedDream 2.0 ~ 3.0 series from ByteDance. Similar to them, we use extensive quality filters and internal models to get the highest quality possible. Even from there, we still hand curate a hand-picked subset.
I actually tried a few experiments in early exploration stages! I trained a small classifier to judge AI vs non-AI images. Use it as a reward model to do small RL / post training experiments. Sadly, was not too successful. We found that directly finetuning the model on high quality photorealistic image was most reliable.
Another note about preference optimisation and RL is that it has really high quality ceiling but needs to be very carefully tuned. It's easy to get perfect anatomy and structure if you decide to completely "collapse" the model. For instance, ChatGPT images are collapsed to have slight yellow color palette. FLUX images always have this glossy, plastic texture with overly blurry background. It's similar to reward hacking behavior you see in LLMs where they sound overly nice and chatty.
I had to make a few compromises to balance between "stable, collapsed, boring model" and "unstable, diverse, explorative" model.
The architecture is the same so we found that some LoRAs work out-of-the box, but some LoRAs don't. In those cases, I would expect people to re-run their LoRA finetuning with the trainer they've used.
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.
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.
FLUX.1 is one of the most popular open weights text-to-image models. We distilled Krea-1 to FLUX.1 [dev] model so that the community can adopt it seamlessly into existing ecosystem. Any finetuning code, workflows, etc that was built on top of FLUX.1 [dev] can be reused with our model :)
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.
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.
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.
Hello HackerNews. My name is Sangwu Lee . I work for Krea and I led the research efforts around the post-training for this model. I'll try to answer any questions you may have, but I recommend you read the technical report I wrote on our site (https://www.krea.ai/blog/flux-krea-open-source-release).
I also see that my colleagues already commented here, but I'll try to answer questions you may have.
Hi! I'm lead researcher on Krea-1. FLUX.1 Krea is a 12B rectified flow model distilled from Krea-1, designed to be compatible with FLUX architecture. Happy to answer any technical questions :)