If you're still using wd14, there's nothing wrong with it, it's good, but obviously not good enough,

ollama llava&Joy caption's accompanying problems are obvious, especially the Joy caption's VRAM cache usage,

So.. This is the just-released PromptGen v1.5

PromptGen is trained based on Microsoft Florence2 base model, thus the model size is only 1G and can generate captions in light speed and uses much less VRAM, And I paired it with FLUX GGUF Q4、CLIP-GmP-ViT-L-14,Because I heard it handles the long-clip better.

PromptGen v1.5 can handle image caption in 5 different modes all under 1 model: danbooru style tags, one line image description, structured caption, detailed caption and mixed caption, each of which handles a specific scenario in doing prompting jobs.

[

   Let's see the difference>>>>

   pull image is here

]

Workflow plugins & models are mainly from MiaoshouAI and xlabs

Huggingface Link：https://huggingface.co/MiaoshouAI/Florence-2-base-PromptGen-v1.5
ComfyUI MiaoshouAI Tagger  Github  Link：https: //github.com/miaoshouai/ComfyUI-Miaoshouai-Tagger