conradlocke

/

krea2-identity-edit

like 602

Thank you very much!

The training was done with various open datasets as well as datasets created by myself in different stages. The general editing capability was achieved via the Omniedit dataset. And then refined and more capabilities were added on top via my own smaller datasets. Higher quality definitely beats quantity at this stage of the Lora. Adding new features now takes very little example data. Krea 2 also really just learns very well. V1.1 has a image restoration example in the model card, while not perfect, it's noteworthy that only a single image restoration triple is in the entire dataset, and it's a pretty bad one.

If you have data. I would greatly appreciate it if you were willing to share. That way I can include it in the next training run :)

Yep, krea 2 is a great learner. Best model I ever trained. My editing dataset is rather small & might not be that effective. I'm thinking about other ways to create editing datasets more efficiently, but not there yet.

Do you know this dataset of mine? https://huggingface.co/datasets/stablellama/change_eye_face_head_person

When not, you might want to use it for further training as well

Do you know this dataset of mine? https://huggingface.co/datasets/stablellama/change_eye_face_head_person

When not, you might want to use it for further training as well

I wasn't aware of it and it definitely looks highly useful. Thank you!

Training code would be nice.

Since using your LoRA with the official AI-Toolkit node (https://github.com/ostris/ComfyUI-Krea2-Ostris-Edit) doesn't really work well - one can only conclude you didn't train on plain AI-Toolkit code.

Long overdue answer, but a good one: the training code is now public — https://github.com/lbouaraba/krea2edit-trainer — with measured consumer-GPU VRAM numbers you'll find familiar :)

Thanks bro.

I gave it a try on my 5060Ti and its basically impossible to train at 16Gb atm - I went all the way down to: rank32 ; 512 bucket ; 100% layer offloading ; cached text embeddings ; single control dataset ; fp8 quant - it started fine then had a spike and overflew at around step 5 or 6. EDIT: hell I even tried convrot4 and judging by the log it seemed to have been successfully quantized with triton but I didn't notice any VRAM difference - still overflew

I know you are working on it but the level of VRAM consumption I saw makes no sense.

Qwen-Image-Edit is much larger. I was able to train it with basically the same settings but with 1024 buckets on Musubi-Tuner (high blocks to swap).
On AI-Toolkit - can't do it on a much smaller model and at half the resolution.

Sure you can improve VRAM usage by making the text encoder caching much better but that won't be enough to make it work on 16Gb.

Anyways, thank you for everything.

Yep, 16 GB VRAM is not enough. You need at least 32 GB VRAM. But even then you will only be able to train 768px with cached embeddings and reduced lora size. I don't think 1024px is possible (although conradlocke said we will try to make it work) without 48 GB VRAM at least. Maybe works with offloading but I never use offloading because it is so damn slow.
I'm not able to train editing on Qwen Edit with 32 GB VRAM (without offloading). Only single image training works.