Update2:
I'm not sure how I feel about this one. It's the exact settings as before, just with the new base 1.0 checkpoint of Anima. You'll have to see for yourself, it's definitely a different result compared to the previous one. Moreover, the version trained on preview3 still seems to work well, so it's up to personal preference.
Update1:
So, I suppose since it worked well on the first run (epoch 10) I never thought about resuming the training... well, good thing I kept the save-state, because it definitely wasn't finished training there.
So for the first update to this model I'm uploading epoch 18. It feels closer to the actual style now (I think), though I suspect it's not without minor flaws (it likes to use depth of field, same as the original model, maybe even more).
Also, I just noticed that the princess peach preview image I uploaded to the first version had wrong metadata, and I couldn't really find what prompt I used with it exactly, thus there won't be a 1-1 comparison with the new epoch. Aside from that one I tried generating the same images, but with the new model, so you can have a general feel on how it's different.
General:
(original description)
Model trained on 3d renders. As an experiment I put 64 dim, and it turns out it's 350 megabytes, so it's considerably larger than usual. Maybe I'll make an alternative one with lower value just to see how much of a difference it makes.
I'm not sure what to make of it. It's the same as with other anime focused models like Il and Noob, where it kind of looks 3d, but it also feels a bit off, especially due to lower resolution.
Higher resolutions do yield a more detailed result and improve the 3d look.
Tags and information
Use tags:
@mdrthrd, 3d, realistic
+ optionally
depth of field, blurry background
Support:
If you like what you find here, please consider making a donation https://ko-fi.com/hat0ri, thanks!
Description
Continued training
FAQ
Comments (8)
This model is so amazing, I absolute love the images it can make, and it even plays well with other artist loras!
Nicely done, I noticed this too about anime-oriented models. They're quite good at illustrative styles but they get this sloppy look for more semi-realistic illustration styles.
Looks awesome, perfect for Life is Strange generations :). Would you mind sharing the training data?
Thanks! I've seen some, good work on the character lora too
I'm afraid I shouldn't share any of the works. I didn't do much with them too, besides picking the most sharp ones, cropping, and resizing all of them to 768 max dimensions - so they would also be a bit useless if someone has better hardware or trains online with 1024 or even 1536 mixed resolution
@hat0ri Thanks. I was just curious about your captioning but alright.
@faye_nsfw Oh, okay. Well, I can tell you the gist of it: in the case of this model I used the wd-eva02-large-tagger-v3 to tag everything, deleted some obvious detection errors.
Tags of importance were at the beginning, so:
"@mdrthrd, 3d, realistic, depth of field, blurry background" would be how all caption files started (same way with character tags). The rest is essentially just shuffled randomly. I'm not sure if the "@" helps or not, but I think it's beneficial to put it with styles
Can you upload more previews like you upload one with peach then the rest is the same character
I made a new post, it should show up on the page soon