Currently trained about 600k new samples into it, which is nearly the core itself, slow cooked.
Runpod is a bit all over the place in terms of updates, the average loss is fixed pretty firmly at 0.330-0.335 an has been for days now.
Here's some teasers.
The core model is still intact; so realism and real characters can still be generated by excluding booru tags, but thanks to CLIP_L's training, we can shoehorn booru into anything due to convergence and regular language use. It's a bit unpredictable for now, but that finetuning is for stage 5. I'd say the CLIP_L is pretty much cooked to perfection, since the core CLIP_L is already trained with so much data thanks to stability.
CLIP_L's never had a problem figuring out what is there, it's always derped on the layering process of identifying what is behind or in front of something anime contextually. So training that to this point seems to have done the trick for now. We'll see in the future if it needs more. Which is why so many of the core images have layered states and multiple characters overlapping.
Conceptually, I think I've beaten the apple fixation problem, but time will tell.
As long as we're careful, the base woman and man will stay perfectly intact, making sure their structure and bodies are fully formed and functional is integral to this process.
'
I have in fact returned to base Flux1D instead of DeDistilled. The outcomes from DeDistilled were showing considerably worse outcomes the more divergent I trained Flux1D2pro's training model, and they are getting much closer to what I expected to see from Flux to begin with.
Epoch 30 be hitting hard.
Epoch 12 i think?
He only has about 200 images.
Ichigo does not have very many at all, yet he's coming through with just his name.
You fucked up my big mac for the last time McDonalds.
There is about 50 characters or so that got sourced automatically for this one. So it should be pretty jammed full of fun stuff.
Soon, very soon. It needs a full round of testing and I'll release it.
Spoopy.
Also spoopy. Love this sort of thing.
This is how I feel when trying to lasso this beast.
Brooding as always Sasuke?
The dreaded grey sushi.
The new "1boy" base character at seed 420. He's definitely got clothes on.
AH SHIT who pooped my pants? Was it you?
Absolute chad bud.
Total beast.
OH SHIT WHO GAVE HIM THAT!?
Realism still works just fine.
Inpainting works now.
I mean it's not perfect, but I didn't give it timestep training to make this easier. It just kind of does it.
Not bad considering I haven't given it direct organized style training yet. (version 7)
Took a few tries, but poogie be upscaled and spiffy looking.
Not quite there yet, but getting closer.
Post version 8 is going to be training the inpainting model, but for now this is what we get to work with. We are currently at version 3.