Finenude

UPDATE V0.2 :

As promised, here is the V0.2 trained for 3 more epochs. If you look at the comparison images I've uploaded with the version, you'll see a woman_64_prodigy-step00002184 and woman_64_prodigy-step00002324.

This 0.2 version corresponds to the 2184 steps because I found that it seemed to be better and the 2324 started to overfit on other concepts (like man) for example.

I've also tested this LoRA on the DreamShaper base model and it's sometimes producing better images so you might want to try that !

I've started working on a full finetuning model. I'll give you more informations this week-end and I might even have an alpha version for donators :)

🤗 Here is the link for donations https://ko-fi.com/fineudesd ☕

Here is the POC version of the LoRA I'm currently training called Finenude specialized on women . It's using a huge dataset of around 30 000 high quality photos and trained for ~9 epochs on a A100 with a batch size of 128.

I'm not very proud of this current version but it's very encouraging and it's more of a POC (Proof of concept) than a true release.

Why is it encouraging ?

Because the model is sometimes capable of producing good quality images

Even though :

1. The dataset has not been filtered

2. The dataset has not been captioned, it's only using the class token "woman" for every images

3. Text encoder has not been trained (because I thought it was too risky to train it without captions right now)

The V1.0 version I'm currently working on will be retrained from scratch and include all these corrections. It will also the texts that can appear in the bottom of the generations and hopefully will be using a VAE that has not the current identified problems.

I'm also considering truely finetuning the base model (not through a LoRA), tell me if you think it can be useful.

Here are my current plans for next releases (in the order of priority) :

• V0.2 : Same dataset, same training but just trained longer... Expect this to be released tomorrow.

• Captioning dataset

• Filtering/modifying the dataset

• More closeups of parts (to make the model better generate them, I think this one is pretty crucial)

• Adding some "regularization" images to prevent some catastrophic forgetting of the model

• Adding "action photos" into the dataset

• Making a better generalized model that will include male photos

Tell me what the limitations are on the current version and what you are expecting for the next versions !

...I'm a very passionate person about SD and I want to give you the best models I can but this passion costs a lot (especially on cloud because I don't have a local GPU strong enough and it also takes A LOT of time to gather the dataset, try the different finetuning methods and test the resulting models so if you want to help me you can here : https://ko-fi.com/fineudesd ☕

So... To answer the question "Is SDXL capable of producing photos ?" ... Well we can't say for sure right now but it seems promising.

Thanks for reading and see you soon for the next versions !