Anima seems to have support for natural language for prompting, as they used it for training. It seems that JoyCaption seems to be as quite good at achieving captioning, but I've heard that qwen and gemma work well too, but may be censored.
I was curious because the LoRA example suggested by Anima used captioning with Gemma 4 31b. Results seem good?
I don't know how impactful captions would be compared to tags for training on Anima, more testing will have to be needed.
The other problem is that JoyCaption is rather large model and cannot fit even on my 16GB VRAM. I had to use quantized models. But once I got it running, I guess it's running, albeit slowly with my vibe-codded frakencode to set up trials.
So with the release of 1.0 Anima and its support for natural language, do you think having natural-language captioning would fit imgutils?
Anima seems to have support for natural language for prompting, as they used it for training. It seems that JoyCaption seems to be as quite good at achieving captioning, but I've heard that qwen and gemma work well too, but may be censored.
I was curious because the LoRA example suggested by Anima used captioning with Gemma 4 31b. Results seem good?
I don't know how impactful captions would be compared to tags for training on Anima, more testing will have to be needed.
The other problem is that JoyCaption is rather large model and cannot fit even on my 16GB VRAM. I had to use quantized models. But once I got it running, I guess it's running, albeit slowly with my vibe-codded frakencode to set up trials.
So with the release of 1.0 Anima and its support for natural language, do you think having natural-language captioning would fit imgutils?