sd_hassan
patreon
Stable Diffusion 2.1 NSFW training update
🕑 Added 2023-01-18 23:10:15 +0000 UTCCONTEXT
So as you know from a previous update, I've run a test of training NSFW content into SD2.1 and it worked well on a small dataset of 300 images across 6 different types of content.
Since the last update I have been gathering more datasets, captioning and here's where I'm at:
CURRENT STATE
- I've started training SD2.1 again, this time the "real" training, not a test run.
- I'm training one dataset at a time, ensuring it's coming through properly before then moving to the next dataset. This is to ensure better quality. The reason is, I found during some tests that training multiple datasets at the same time can work well but there can be a bunch of them that seem to not be trained as "evenly" as others in the batch. As such putting a focus on one dataset at a time will spread the quality better overall.
- I will train each dataset, download the model as a backup, then start the next training run immediately.
- In parallel to this, I am continuing to grab more datasets and setting them to 768 resolution and manually captioning. I think this process will continue even when the model is released I think it will continue to be finetuned with more and more datasets to make it an awesome NSFW model.
Datasets:
I have the same datasets as the testrun, just additional amounts in each dataset:
- BSDM male/female
- BJ (Male and female performing)
- Female body/anatomy - Various body types, breast types, nipple types, muff types etc
- Male body/anatomy - various body types, penis shapes/states/muff types etc
- Buttplugs
- Dildos
- Men on Men
- Various types of penetration M2M, M2F, F2F etc
- Upskirts
If you want to recommend more datasets or have datasets at a higher than 768 resolution, feel free to reach out or share
- Sharing a source website is good
- Sharing manually captioned images is very helpful as I need to manually caption each image, I'm not satisfied with CLIP captioning
FAQ:
- I expect the first release in 1-2 weeks if all goes well in training, this is based on 24/7 cloud training
- Patreons will gain first access, maybe 1 week early access to test it - provide feedback before I open it up, may limit this to tiers to prevent leakage and allow us to make sure it's ready
- You can recommend datasets
- The model will be released free like all my models, once testing and early access is complete for supporters
- This model will not be a merge of any sorts, only trained content but other users can freely merge once it's out in the wild. I want to ensure I fully control the quality of what's going into the base model which I cannot do when it's merged
Thanks for your continued support, the 24/7 cloud training adds up, at some points I've multiple clouds in parallel testing various settings to try get this right so your support really goes a long way for this!
Comments
I'll share an update in a few days but the summary right now is training is still going on 24/7, it's taking a long time due to the dataset size vs the resolution vs the learning rate. I'll share samples in the coming days as it progresses in a new update
Hey any updates on this?
i've created a new post explaining the process a little more
Terrific, thanks for the update! It would be great if we could see a sample of the images you are using and the captions that you're writing to go with them--just curious to understand the format/detail.
+1 this!
William Tatum
do you have a guide anywhere on how to appropriately caption things - is it just a text file that follows the image file with the same name as the image file and very literal words, phrases in it?
