Base model trained over z image de-distilled z image turbo on 8 gpus for about a month.
It is designed to make transformation sequences, before and after images, transformation comics, whatever, trained on a dataset of 40,000 images.
Was originally going to use this as a base to merge with z images edit model, but turns out you can't merge it back in without frying the weights so here's this.
To run this you use 30 steps and and a cfg of 6 with around 1024x1024 or higher (i usually go higher)
Description
FAQ
Comments (4)
I love this concept its pretty hard to find too. Out of curiosity have you tried to replicate this concept to illustrious? I might gather a dataset and give a Lora a try. Thanks for your hard work
The original model was trained over sdxl although it was a bit fried due to my inexperience tuning hyperparams at the time, i did actually run a small training test on my current dataset with modern training techniques to see if training an sdxl finetune would be something i could do for cheap, but it came out that it cost a only a bit less than training one over z-image.
That being said, my results were with a huge 40k images that were typically relevant; it's probably much faster to do it with a much smaller number, if less generalizable, easy to overfit with that.
Hi RA,
Your Checkpoint works surprisingly good.
In my opinion, it functions better than your Qwen image edit transformation lora.
Is there a chance that you might expand your work to Z image base as Lora or Checkpoint?
Thank you for your work and effort on this!
Thanks!
I tried, merged base models are completely fried, tried to train it back into being normal and £200 later it's still relatively fried, :( .
This has 4x the training data than the qwen image edit one did and is way cheaper to train so was trained a lot more, qwen image edit is a rather big model.
It's quite annoying the merge didn't work because the whole reason i trained this was so i could use it as a base for the edit model, in the same way qwen image and qwen image edit tunes are cross compatible, so i guess i'll just train one over flux klein in a few months when i have some additional spare income, this was eyewateringly expensive to train.