sd_hassan
patreon
SD2.1 Update
🕑 Added 2023-04-22 13:04:37 +0000 UTCHey all,
It's been a while since we've had a progress update on SD2.1
What's been going on?
I've been training various concepts (12 different types of datasets) over the last couple of months. SD2.1 has proven quite difficult to train and I've been experimenting across many different training configs to try find the sweet spot of training parameters and convergence of my datasets.
How was the training results?
Some concepts like female body/BJ/Cameltoe / breasts etc went quite well
Penetration, Gay, Kissing, BDSM , Dildo didn't go quite well and SD2.1 was struggling to learn these concepts.
What methods did I try to resolve these?
- I first started out training all my concepts together across multiple training sessions. Each dataset concept has a different varying amount of images in the dataset so I found that some concepts may become overfitted / overtrained and others not enough. Again, it struggled with some concepts and not so much with others.
- Then I split out my datasets into individual training runs so I can monitor each training session individually to ensure individual quality. This took a huge amount of time because the training parameters are quite different for each dataset I used. There was a tonne of trial and error and this is what has taken me most of the time.
- When I thought I had a good result, I kept that cktp and basically tried block merging each individual model back together to form a complete model. This merging process didn't work so well so I ditched this effort and went to the next method.
- Next I took the best version of the hardest concept to train on base SD2.1, then I finetuned that with one concept at a time. In other words, I trained concept A over SD2.1 base until I was happy with it. Then got my Concept B dataset ready and tested various parameters finetuning concept A with the Concept B dataset. This process I rinsed and repeated for every dataset with different parameters and I locally tested each model along the way.
The last method has been the most successful so far. Yes there are still concepts that have not been made perfect. My goal is not to get stuck on them but to progress to a release and then finetune those over time with newer versions of this same model to enhance them. For now I want the good working concepts to become the base.
What is next?
- I am testing the model locally, testing it individually and testing it merged too to find the optimized result. So far, the individual finetuned model performs the concepts best which is expected as it's less diluted but there is value in merging with models like WD to gain some concepts that were trained into WD while avoiding some of the strong bias WD has such as only asian women. So far I am not re-training any concepts any more just yet and leaving some of the flaws as they are for now, no such thing as perfect. My hope is this release will allow others to have a good starting base to continue bringing NSFW to SD2.1
- Training hands - I am taking a run at training thousands of hands into the model as an experiment to see if it will improve
- Once I am satisfied, I will release this as an early version to patreons only for a period of time. This is because of your patience I want to give you the benefit to have a good early access period to play around with the model and give feedback etc
- Then the model will be released free to all as I have done with all my previous models.
How long is left?
I'm excited about this update because this is the closest I've come yet and I think the early access release is only around the corner. I think 1-2 weeks left then the release will take place.
How will early access work?
I will release to the patreons at each tier first. The higher tier, the longer access you have. ie Ultimate supporters will likely have it a week before the other tiers.
Enough of the news, show us examples
Alright, there are flaws but here are both. Bear in mind these prompts are super basic, only 3 - 4 word negatives in there and run at 20 steps between 5-7 CFG. There's a lot more tweaking that can happen at the prompt level so I'm being basic as I can on purpose so you can use prompts to enhance your results in many ways
Somewhat decent examples: https://imgur.com/a/OYILQSm
Failure examples: https://imgur.com/a/PC9xz7X
Comments
Hi there, I no longer have a commercial licensing agreement for my models with Fantasy
PirateDiffusion.com
Hi Hassan, love your work! Can you clarify if the 2.1 release is also property of Fantasy.ai or can it be shared on our Telegram bot?
