indexed has turned out to be my favorite model I've worked on so far.
I've found it balances strengths expected of an Illustrious model well - and handles high resolution generation with satisfying detail. It's fun to prompt and feels steerable.
I really hope you enjoy it - I have!
Inspect the showcase images to see prompts I've had fun with.
The recommended resolution is 1168x1712 (or other variations of ~1414^2).
Setting for all images in showcase:
All images in showcase are straight from the model - no post-processing.
Steps: 30
CFG: 6
Sampler: DPM++ 3M K or Euler A
Resolution: 1168x1712
Description
further training and more refinement from v1b
enjoy : )
FAQ
Comments (2)
The details in v2 are sharp and crisp, but I feel like it lost some of the versatility of v1b.
For example: I have a hard time generating broad shouldered men with venetian masks (come out as women), pixel art has great coloring but not enough pixels or dithering, and traditional media comes out a little too clean.
That said v2 does a fantastic job with digital art styles, just keep in mind that it is very "clean". I'm experimenting with prompts that will make images less perfect but can't really get the looks I want.
I love indexed and v1b is my go-to model for generating everything so far. I love your work epiTune!
This is excellent feedback! Thank you very much.
You bring up good points - and I think it stems from the v2 training objective being too singular.
I'm right now exploring post-RL training methods and that will (hopefully) results in either new amendments to v2 or even a v3 - but in addition to that I will likely look into training with a broader training set also.