Get the pro version of my checkpoints and workflows on my Patreon.
Onsite generations are permanently available on these models:
👉 Realism_Illustrious_By_Stable_Yogi : https://civarchive.com/models/974693?modelVersionId=2091367
Read "About this version" Section for Model details and settings.
Important Usage Tips to get similar results like my images.
Add Realism_Illustrious_Negative_Embedding at the beginning of your negative prompt.
Add Realism_Illustrious_Positive_Embedding at the beginning of your prompt.
Join me on my Patreon for exclusive perks and early access to unique resources.
To discuss custom Lora's or models, feel free to connect on Discord.
đź‘Ť Like this model to keep me motivated and inspired to create more!
đź’¬ Drop a comment and let me know what you'd love to see next.
🌟 Review this model to help me improve and make even better creations.
đź”” Hit that notification bell to stay updated with my latest models and updates!
Description
Prompt Adherence intact
Focus on Increasing realism, detail and lighting.
Next Update i will fix the faces and diversity.
FAQ
Comments (11)
I've tested the version 2.2 and have good realistic results even using low steps like 25, anyway, amazing new version :)
Wait 25 is considered low? What's the normal amount you're supposed to use with Illustrious? Just now switching over to it from Pony so I don't know a whole lot about it lol
Thanks for the feedback.
I usually consider 27 steps as normal. 36+ steps as good.
This mostly works with all checkpoints, (Flux excluded)
@Zuriel89Â yes, for realistic models 25 steps is low, I recommend you try the dev consideration to create a good image, but don't worry it's just for realistic models, for some illustrious models who isn't realistic like anime 15-20 steps is good :)
What is the difference between FP16 and BF16 ?
For example, Prompt adherence vs details ?
As per Chat Gpts explanation it better in both departments, lets test it out, this is my first attempt at BF16.
FP and BF both represent fractions (floating point numbers) in different ways. Neither is "better", they just have different levels of precision where sometimes BF is more 'accurate' and sometimes FP. Something you could try is create a prompt in one, fix the seed, then generate with the other and observe how it changes.
As far as I understand, bf16 is better, but only supported in RTX 30-series GPUs and above.
Specifically:
fp16: 1 sign bit, 5 bits for the exponent, and 10 bits for the fraction (mantissa)
bf16: 1 sign bit, 8 bits for the exponent, and 7 bits for the mantissa
I might be wrong in that it might be better, as it's a difference of how the precision is calculated. For example:
Original value: 1e-08 (0.00000001)
FP16: 0.00000000000000 (Binary: 0|00000|0000000000, Hex: 0000)
BF16: 0.00000001001172 (Binary: 0|01100100|0101100, Hex: 322C)
In this case fp16 makes a very small value 0, while bf16 manages to represent it. Does this matter? Maybe not. If it breaks LoRAs or something I don't know either.
@Dead_Internet_Theory This is the technical explanation, but my question was more functional (hence the second line), knowing that the theory does not necessarily allow to determine the quality of the image that we will obtain (detail, depth, interpretation of the prompt, shadows...). In short, as said @Stable_Yogi , we have to test.
Awesome model and was wondering is there a place to read changelogs ?
Pls check "About this version" Section.