TL;DR; tiny FP8 model with quality of FP32; merges allowed (please mention this model); improved single pass eyes and fingers correctness; optimized combination of sharpness and realism of CinEro Illustrious V2 with cinematic style and creativity; precise details and cinematic look; light particles.
Note: Tested in A1111 and ComfyUI. For A1111 I recommend to turn off the "--xformers" as it didn't work with latest PyTorch.
Hey guys and gals!
No EA AGAIN 👇
Thank you SO MUCH for support on V2 and V3. I've nearly 10k of BUZZ on training with new datasets. Some of expanses were compensated by the donation from kind supporters (MANY THANKS AGAIN!!). Some of expenses wasn't recovered.
This is why I need to shyly encourage you to post images (bad and good are welcome), tell your friends about my creations, leave a feedback. That's the only thing fueling my progress.
Just 150+ postings will compensate training costs at CivitAI. You can also collect this model or interact with my images. This is just a TIME for you but IS A GOLD to me.
I don't ask you to donate YELLOW BUZZ, but will be very glad if you spare some. No matter what, all my BUZZ will be spent inside CivitAI on training new models.
[DISCLOSURE] I am using the Post-Maker (3-pass workflow) to produce HQ images using only the released model. You can download it and modify the Steps or resolutions. Steps 30 for 1st pass and 8-10 steps for 2nd and 3rd passes are enough to produce good results 90% of tries. 3rd pass is not so important.
Description
FP16 Unet + BF16 CLIP
Probably more stable and better for merging.
Please mention this model to track the inheritance.
FAQ
Comments (5)
Im new to Illus. Do 0.1 based loras work on 1.0 based models.
I tried a popular lora, and has no likeness on 1.0.
1. @Omega_Xlll reported that even SDXL LORAs are working with CinErl Illustrious V4 FP8.
2. Ill 0.1 is a beta slightly trained on datasets with Booru tags and Natural Language descriptions. Ill 1.0 is a 0.1 with more training with better natural language description (as reported by devs). So, I'm expecting the v0.1 LORAs should work on both versions of Illustrious.
3. LORAs for Ill v1.0 and v0.1 should work on CinEro ILL v4, but I didn't test it and it wasn't a goal of this release.
@homoludens thank you !
so far testing it, this model has a lot more use and function than other illustrious models. While the nsfw works pretty well, I can use the other prompts to get real nice looking illustrated like images. Other illustrious models focus on the girls and less on the object looking good.
That said, the hands are still mangled in some images and I wish the faces emoted better. It seems to only have happy and sad and nothing in between. That said I haven't fined tuned it yet.
it could also use more attention to steampunk, it seems to ignore my commands for pipes, gauges etc. Like I wanted a coffee maker steampunk style and all it gave me was a coffee pot. Even the cartoon versions can do that.
Yep. 3 / 6 fingers sometimes. HiRes Fix or ADetailer solve that most of the time, but hands can be better. Agree.
Didn't train for emotions. Main focus is on sharpness, cinematic colors, anatomy and creativeness. Will work on emotions in future.
I can recommend putting the "(neutral_expression:1.35)" into Negative Prompt. It helps me sometimes.