aleksa-codes

/

flux-ghibsky-illustration

like 235

Hi there @ftopal ,

Thank you for your interest! The images I used for training were sourced from an Instagram page that focuses on AI-generated content. The owner of that page has kept their methods quite secretive, and I believe they may have used Leonardo AI to create those images. While I’d love to share more details, the page itself is a core part of my approach, so I can’t disclose its exact name.

The training was primarily an experiment to see if I could replicate the style of that page. In the future, I plan to fine-tune the model further by selectively retraining it on its best outputs, which I’ll choose manually.

Thanks again for reaching out, and feel free to ask if you have any more questions!

Hi @aleksa-codes

Thanks for your reply! I was wondering if you are planning to provide this SD 3.5 as well. That was my reason for asking. I do like image styles and would love to see it in other architectures as well.

Hi @ftopal ,

Yes, I'm considering it now that I see people having interest in this LoRA/style. I haven't tested SD 3.5 myself yet, so I’m curious about its performance compared to Flux. Have you had the chance to experiment with SD 3.5? I’d love to hear if you think it’s noticeably better.

Also, I noticed that a new LoRA trainer for SD 3.5 was recently released on Replicate, which could streamline this process: SD 3.5 fine-tuner by lucataco.

Really depends on the use case, i know some people are into exceptional realistic images and having better human anatomy and there FLUX seems to be better but for this image style, it's hard to say. That's why i wanted to like compare and see how both of them look.

Hi @aleksa-codes ,

Did you have person images in your dataset as well? I am asking because sometimes faces are not coming out that great. I was wondering if that has something to do with the lora itself.

Hi @ftopal ,

There were no person images in the dataset. I was getting decent results of person faces during my tests, the only problem I was noticing with people was that skin usually had wounds and was bruised up, which I think comes from the style of images it was trained on. Here are some of my results with people:

Yes, close up shots/portraits work fine for faces but natural environments ends up very less detailed faces and mostly eyes are not detectable. Let me share some results:

In most cases, when the people are bit away, their face quality shrink significantly. I thought maybe it's FLUX related but trying out other LoRas showed me it's not.
One example of what I mean used with the same prompt/same seed:

With this LoRa:

with another LoRa:

I think this is tied to having little to no human in the training dataset. This also affects the quality of hands/fingers. Again mostly for medium to long shots. I sort of think it would be improved if the dataset contains more human samples. Do you plan to have like v2 version of this by any chance? :)

Hey @ftopal , just wanted to let you know that I released v2 on CivitAI, the one that I mentioned that had more face data in it, so you can try it out: CivitAI

I am not sure if you generate locally, or on Hugging Face/Replicate, if you can test it out. But versioning is a bit harder on HF and Replicate I would need to override the current one or release a separate repo.

Hey @aleksa-codes ,

Thanks for doing the new version of this. I'll have another look. It works fine on HF as well, just give a different file name. lora.safetensors is the default name but i can download if you had lora_v2.safetensors as well. It just people need to refer to the new version. You can have a new repo as well but i don't think it's needed.

I'll share some updates if I think the new results are looking better.

Hey @ftopal , just added lora_v2.safetensorsand config_v2 like you said.

Hey @aleksa-codes ,

Thanks, I have tested it couple of days ago on set of prompts, I don't see any improvements. Sky/clouds structure got even worse to my taste. :/

@ftopal thanks for the feedback, sorry to hear that..

@aleksa-codes I am also sharing my experiment results here so you can perhaps see it. I think there is a slight improvement but at least to me it wasn't that much. If you include perhaps various angles of people (like wide/medium/close ups) than first one could get better. I don't know much about the training set but that's all I could do for now.

Ghibsky v1

Ghibsky v2