Hello folks - version 3.5 is something different - READ THE FOLLOWING INSTRUCTIONS
What I have improved thanks to your feedback:
Hands quality, has been improved
You have now a much wider control on the camera, pushing the perspectives a lot further than any other model I've done.
Even more "Western"
Unfortunately, through my testing phase, I've noticed that this model - that I've created and tuned with ComfyUI is not as good with Forge. Feel free to download my images, they are all carrying a very simple workflow for Comfy that you can use if you're not used to that UI.
As always, follow the Quick Start instructions in the model description to have a reliable starting point for your generations.
WHY AM I DOING CHARACTERS WITH A WHITE BACKGROUND?
AI has a "limited attention" and by asking for a white background I force the AI to focus on the character design, I do this because I prefere to generate the backgrounds in different ways and then Photoshop the setting behind the characters for a better final output.
If you like this model, consider bidding for it to keep it available for anyone to play on Civitai.
Thank you to NMackelie to have sponsored this model!
The version 3.0 is a collaborations with Bridgewalker who trained a LoRA specifically to fix issues in this model.
✦ Quick Start ✦
1_____Size: Around 1024 x 1344
2_____Prompt: Start from the prompts present here and tune the keywords based on your needs:
Quick Prompt
best quality, absurdres, dynamic pose, solo, (toon (style)) or best quality, absurdres, dynamic pose, solo, (western comics (style))
Quick Negative Prompt
sketch, lowres, low quality, worst quality, 3D, greyscale
3_____CFG: 3.5 to 7.0 (low for realism, high for cartoons)
4_____Sampler: Euler a / Karras
5_____Steps: 25 to 50
(Higher the Steps, between these two values, the better the results)
Description
Major update:
My journey into fine-tuning illustrious models continues. This is an overall improvement based on the Illustrious v2.0 that was lacking in the details and background departments.
FAQ
Comments (7)
3.0 is really solid so far! I always look forward to your models.
I really do enjoy this alot. It's super flexible and nets me consistent results even though it doesn't seen to have many innate artist styles as far as i have tried. Is there any list of artist tag this checkpoint does understand?
Hello! I'm glad you're having fun!
This model has been done by recalibrating the weights manually, training specific LoRAs to teach the model concepts over style and artist based LoRAs with negative values in order to actually remove many of their signature aesthetic in order to have a "plain" starting point that doesn't try to replicate them.
I've done this intentionally, because I'm not a fan of using AI to replicate someone else's style ^_^
@Arthemy Understandable. It's a slightly other workflow than with WAI that somehow became the standard for so many generations here. But the results speak for themselfes.
Again, great work and i hope you keep it up. Excited to try out more with your creations ^^
Really great stuff, as usual. I love how your models keep moving away from the mainstream manga styles. Like somebody else said in another one of the comments all your updates are instant downloads. Just a quick comment/question: I find it sort of hard to represent different ethnicities. As in, "ebony" or "black" almost always render a "white" character. Playing with skin-related prompts sometimes work, but does not seem super reliable. Are you planning to add this kind of feature in the future? Thanks!!!! :):):)
This is a very interesting tip, I might have missed that in my benchmark.
Just to make sure, try again avoiding "Black" or "Ebony" which are tags usually related to a plain color or a material, and try with "Dark_skin" + "Dark_skinned_female" / "Dark_skinned_male".
My models are highly keywords sensitive so, sometime, they need the exact tag they are expecting in the prompt. I use extensions like "TagAutocomplete" to make sure I'm writing the most specific and popular tags.
Let me know if that helps!
@Arthemy Agreed, thanks! What your suggest is what I do when I want a "black" character. I can also think of workarounds for Mediterranean or a Slavic looks---though I haven't played that much with them. But things get harder if you want Asian-looking characters (Indian, Chinese, Japanese, and so on). I fully understand this is harder to do, but it would also make the models a lot more flexible. Again, your work is excellent :). Just my two cents.