Very good job! But I have a question.
OMG! I'm so excited and looking forward to the release of the full version. This model just amazing. Although it's early access, the quality is already better than Illustrious. It's a pleasure to train styles and characters for this model. But when I do Hires fix, my art becomes either pale or distorted. What kind of upscaler and what settings do I need to set?
https://civitai.com/models/1613840/ultrasharpv2
https://civitai.com/models/1017531/4x-animesharp
Pair with er_sde for cleaner and sharper lines and it's great. With around 30 steps and CFG of 4
My favourite Lora:
https://civitai.com/models/2364703/rdbt-anima
This Lora really improves quality, details, prompt adherence, stability and anatomy.
My favourite checkpoint:
https://civitai.com/models/2385278/animayume
This is an interesting find as well:
https://www.reddit.com/r/StableDiffusion/s/ajX5A4hSVw
Hires fix upscale works on square images or landscape, just not portrait, and 2x upscale is too much. 1.5 works just fine.
My favourite Lora:
https://civitai.com/models/2364703/rdbt-animaThis Lora really improves quality, details, prompt adherence, stability and anatomy.
My favourite checkpoint:
https://civitai.com/models/2385278/animayume
Both of these are bad, don't get baited. Until we get a fix for the catastrophic memory loss, finetune and lora will be dodgy...
My favourite Lora:
https://civitai.com/models/2364703/rdbt-animaThis Lora really improves quality, details, prompt adherence, stability and anatomy.
My favourite checkpoint:
https://civitai.com/models/2385278/animayumeBoth of these are bad, don't get baited. Until we get a fix for the catastrophic memory loss, finetune and lora will be dodgy...
What do you mean memory loss? It's not perfect but it's definitely getting there it's only a preview version but it's still very promising.
Regarding the checkpoint it is by duongve13112002 one of the best Pony creators and they used a custom dataset (according to their CivitAI page) to further fine-tune the base preview anima model. They also have a great fine tune called NetaYume Lumina (Neta Lumina/Lumina Image 2.0) and a Illustrious one so they definitely have great experience with different models.
Regarding the Lora it's by reakaakasky and I'm not sure of the wizardry behind this Lora it drastically improves the quality, stability and prompt adherence of the model and they even have their own checkpoint with this lora baked in.
So they are definitely not bad and again for a preview model it's very impressive and will only get better when the full anima model is released.
My favourite Lora:
https://civitai.com/models/2364703/rdbt-animaThis Lora really improves quality, details, prompt adherence, stability and anatomy.
My favourite checkpoint:
https://civitai.com/models/2385278/animayumeBoth of these are bad, don't get baited. Until we get a fix for the catastrophic memory loss, finetune and lora will be dodgy...
I agree with rdbt-anima, when i used it with my character lora getting much poorer and worst results then without it.
With animayume disagree. I dont trained chara lora on based anima, but Animayume result are much better than Illustrous v0.1 and v2, and easily worked than with ZIT
The former comment is right at least on one of those- the rdbt model is absolutely just snake oil that weakens artist styles and nothing else, the base model is way better off without it.
The author is a complete schizo who is crying about the very permissive license Anima has in his model description and blocks people for asking valid questions about his placebo models on weird ego trips.
I can't speak to the other model though.
As for upscaling, it's a known issue, the full model release will surely resolve the problem.
I haven't been this excited about image gen since early 2024 lol.
+1 on what Mudflap said about the upscaling issue, its at least very easily mitigated with this lora from civitai in the meantime
https://civitai.com/models/2448854/2048-resolution-support-or-anima-preview?modelVersionId=2753537
I've shared numerous examples both sfw and nsfw most should show up on its page, for issues like this we just have to wait for the full release. Every single gen i run uses this lora for the time being. (civitai's metadata scanner is dogshit and may not always show it).