Updates?
I've pivoted to another project, sorry!!
I'm mostly getting what I need out of it as is, especially with Hyperfusion now being available.
I may revisit it at some point, but frankly Krea 2 has stolen a lot of Anima's thunder for me. I can't exactly train it at home but i have a bunch of buzz on Civit that I'm not doing anything with, so I'm liable to take some stabs at it over there.
I have ADHD, I tend to drift from thing to thing. No doubt I'll come back to this if Tigerpillow never does :P
I did a run for Krea 2, and it turned out okay, so I might end up redoing this version of the dataset on Anima with Civit, since it's far more convenient for me.
Sad. Any way I can see the dataset?
Oh, Hyperfusion? Is it good?
I'm not giving anybody my dataset unfortunately. I don't own the material, some of it belongs to people who I respect and I could get into trouble for redistributing it if this identity gets connected to my old one. It was pulled from E621 and captioned with Gemini 3.5 Flash, this is not hard to do.
And yeah Hyperfusion's good!
Any prompting tips for Hyperfusion?
Hello again @inflatebot . Sorry to bother you again, but I've been having trouble with my dataset since I've decided to try making a similar LoRA.
I decided to make my own LoRA because hyperfusion is not good at different compound shapes, and yours only works well for upright characters (from what i've tried, if you can get it to work, please tell me how to prompt that).
To start off, my two taggers, JoyCaption and Gemma-4-12B-Heretic, are NOT good at this. They constantly hallucinate facts and shapes. For example, they think there are multiple characters when there is only one, that breasts are buttocks, bellies are...also buttocks, feet are hands, etc.
So I'd like to try it with Gemini to see how good it is.
Also, what prompt are you using for Gemini to make it not refuse? Can I see it verbatim? I'd hope it's not as confidential like your dataset.
Oh, yeah, sure! I'll actually just upload the scripts I used, as they differ from Anima's reference captioning scripts; basically rather than loading a model itself, it calls an OpenAI-Compatible endpoint (I used OpenRouter.)
They're easiest to use with uv, otherwise make sure the pillow and openai packages are installed from PyPI.
The model was Gemini 3.5 Flash, which seems to be the most open to this type of work.
The captioning script's system prompt was edited by hand from the reference Anima captioning system prompt.
The caption expansion/rewriting prompt was made for the Krea 2 offshoot, so its system prompt follows the shape of Krea 2's prompt enhancer.
Tysm!