Experiment - test of CivitAI trainer for a Flux LoRA using some of the training images from my Decorative Leather TI SDXL

https://civarchive.com/models/569267/decorative-leather-and-mixed-media-tixl

So far, the TI produces better results than the LoRA.

I hope to be able to train a better version locally once kohya_ss allows Flux LoRA training.

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

The LoRA works on CivitAI generator using Flux DEV.

The LoRA has a strong effect as expected.

It cost me 163 Buzz for four 1024x1024 images - see the showcase images above for the 4 parrot images.

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

The LoRA works on ComfyUI with Flux DEV.

The LoRA has a strong effect as expected.

Thanks to @osikubi for the info. See image post below.

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

Working on Forge vf2.0.1v1.10.1-previous-350-g0b70b728 (or greater I assume) with

Flux.1-Dev GGUF Q4.0

https://civarchive.com/models/647237?modelVersionId=724099

I also needed:

t5xxl_fp16.safetensors and clip_l.safetensors in Forge’s \webui\models\text_encoder folder

+

ae.safetensors in Forge’s \webui\models\VAE folder.

Speed was a couple of minutes per 1024x1024 image.

However, in Forge the effect of the LoRA was not as strong as I expected.

It is possible that future changes in Forge will sort this “subtlety” issue.

INCREASING THE LoRA WEIGHT FROM DEFAULT 1 ---> 1.5 GIVES FULL EFFECT

Also in Forge:

With flux1 Dev BNB NF4 V1 model:

https://civarchive.com/models/638187?modelVersionId=713648

Simple prompt works fine, taking a couple of minutes.

Adding this LoRA to prompt does complete, but it took around 34 minutes for 1 image :-(

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -