What's the difference between v4 and v3? Were there improvements or just style variation?

WE ARE SO BACK! 🗣️

My personal test result for ver4.0 VS ver2.0 :

*4.0 has better background sharpness & details than 2.0.

*2.0 focus closely on the person in the frame, while 4.0 normally has a wider FOV.

*4.0 has slightly better prompt understanding.

*Lora support is slightly better on 4.0.

Ver4.0 &Ver2.0 seems to have baked in similar 2.5D artist style, while 3.0 clearly has more flat 2D anime style instead. In conclusion, I would prefer using Ver4.0 for wallpaper and Ver2.0 for NSFW.

This upgraded model is really bad. How did it get upgraded? It got worse after the upgrade.

The boobs are really perfect! Thanks

This model is trained to focus only on the girl. If you try to create couples or add more than one character, it won’t work. I recommend going back to version 3.

V4 is the more natural continuation of V1 and V2. While V3 should be considered more of a sidegrade with a stylistic overhaul that allows for a slightly sharper, more 2D approach.

V4 returns to the distinct Prefect look while focusing on its softer, more mature look. The overall lighting, both cascade lighting and direct lighting is a bit warmer and returns more complex images with complicated rather than simpler lighting. Some may like this, but it can feel at times that it's "trying too hard".

I'm having some issues with some gens as previous members have mentioned. The model does feel like it's very girl-centric, and at times even at 2K the males feel lowres.

Personally I'm interested in the further development of V3 as it seems to be a more unique contender in the arena of the "2D, but not quite 2.5D" styles with its high-end backgrounds and lighting, while still allowing for colorful and exaggerated facial expressions that people expect out of an anime girl.

This is a really good model. This model is very faithful to the prompts. It reproduces the fine details. This has always been a high level.

it's follow my prompt well and i love it, thank you!

I am waiting you to upload V4 to tensor

Sometimes the eyes get messed up for me. They are distorted and have multiple colors.

I'm using Euler ancestral, 20 steps, 5 CFG in Comfyui. It happens with long prompts and short prompts. Any idea what I'm doing wrong? I didn't have this issue with V3.

v4 has a worse prompt adherence compared to v3. For example, after prompting for face details, and camera angle like cowboy shot in the positive, also "head out of frame" in the negative, the model insist generating the character's head out of the frame. The issue was solved when I switch to v3 with the same prompts and settings.

A great model, but... butts. Seems to have a strong predisposition to giving any character thick thighs and a fat curvy ass.

In my experience, v3 favors generating younger characters, and v4 favors mature characters. I think it's best to switch between them according to your taste.

good

Just wanted to give you a GIGANTIC thank you for this model! I have not tried V4 yet, but V3 has been a joy to use both with SFW and NSFW generations. Your a genius, this model is rock fucking solid.

v4 really lives up to its name — it’s perfect. The characters come out at just the right age, and the visuals are absolutely beautiful. However, if you want to make anime-style screencaps that look like scenes from a TV anime (not game CGs or illustrations), you’ll need to set very high weights for both “anime screencap” and “anime coloring.” So I’d recommend making normal illustrations instead.

Please make one that has no lineart in training 🙏

Does V4 contain updated Datasets? Like for Characters and the such?

v4 is like v2plus,awesome

what would be the recommended level of Upscale and Hires Steps

Probably the best illustrious model atm but for the next version could you try to make it a bit more compatible with style loras?

very consistent, just a joy to work with

I think v1.0 will forever be the ebst illustrious model? Well for this style. Feel like the later versions never come close to the results I ge twith v1.0 And every time I try a different model I am disappointed in comparison. Except for WAI, that is the other best of model that is great, lately I mostly combine WAI and Prefect v1 ...

For some reason, no matter what I do, I only get black generations with this (intermediate gens too). Other ILL models work just fine. No idea what I'm doing wrong...