illustrious ULtraReal (Beta)
Full FP32 Model (Full Checkpoint)
Using FP32 CLIP is recommended but not required
Forcing FP32 UNET is not recommended for most users (BF16 UNET is recommend)
If using GGUF use these CLIP models
BF16 UNET and FP32 CLIP will see no slow down other then model load time for most users.
Note: I have multiple request to do illustrious, I always thought of it as a artist/art style model. I feel most of the function is gone when forcing the model back to realism.
Description
FAQ
Comments (8)
I will second that, I would love a detailed reason for realistic illustrious. Does it work well with current Illustrious LoRAs, that might be a cool reason.
I'm not sure on that, I would think that character lora's would effect clothing, hair, eye color.
Ornaments as well, I did some work with a anime Flux Checkpoint and it did a pretty decent realistic Bao Seth when I asked for it.
Realism_Illustrious_By_Stable_Yogi v2.2: https://civitai.com/models/974693?modelVersionId=1265485
Is a good example of a realistic illustrious model that kept its prompt adherence, character knowledge, etc.
His later model v3 seem to have traded that for more realism. So you might ask why use a realistic illustrious model over a realistic pony model? It's because a realistic illustrious model that's not overtrained on realism has a lot better prompt adherence, a lot better character knowledge and a lot better image coherency than pony models. Basically, all of the reasons that illustrious is better than pony also apply to its realistic versions. His v2.5 is still good but his v3 has lost a lot of the prompt adherence that makes illustrious so special. I personally use v2.2.
Both look pretty good I think the newest version is the most life like but painting like semi-realistic of 2.2 would still be acceptable if it came at the cost of prompt adherence. I notice this checkpoint is a full 6 gigabits bigger then Stable_Yogis is that the cost of keeping both?
@devon_rex18406 I think this Ultra Real checkpoint is 12 GB because it’s in FP32, whereas the Stable Yogi checkpoint is FP16. That’s just my guess, though—I could be wrong.
@devon_rex18406 I think, if you have enough RAM a good way to keep prompt adherence and maximum realism is to use a prompt adherent model with a extremely realistic refiner, so maybe v3 would make a good refiner for v2.2.
@QH96 You are correct, though most people (Me also) are downcasting the UNET, if they read any of my articles they wouldn't be downcasting the CLIP though