Turn any image into a prompt. Load an image and this workflow gives you two outputs side by side:

Full prompt: a comma-separated, SD-style prompt covering subject, clothing, pose, expression, setting, lighting, camera angle and style.

Pose only: just body position, limbs, hands, head tilt, gaze and camera angle. Paste it into your own prompt to borrow a pose without the rest of the image.

Uses JoyCaption Beta One, which is uncensored and handles NSFW images without refusing or getting vague.

REQUIREMENTS

- Custom node: joycaption_comfyui by fpgaminer (install via ComfyUI Manager, or git clone https://github.com/fpgaminer/joycaption_comfyui)

- Recent ComfyUI (uses the built-in Preview Any node)

- About 17 GB VRAM by default, or set memory_mode to "Balanced (8-bit)" for about 4-5 GB

MODEL (~16 GB), two options

- Auto: the first run downloads JoyCaption Beta One from Hugging Face. Later runs reuse it.

- Manual: download every file from https://huggingface.co/fancyfeast/llama-joycaption-beta-one-hf-llava into this exact folder:

ComfyUI/models/LLavacheckpoints/llama-joycaption-beta-one-hf-llava

If that folder exists, the node never downloads anything. It must contain all the files; a partial folder errors instead of resuming.

TIPS

- Edit the user_query text in either caption node to change what it describes.

- keep_loaded is on, so captioning more images is fast. Turn it off to get your VRAM back for generating.

- Captions describe what the model sees; they are not the original prompt. Skim for small errors (left/right, counts) before using.

- For an exact pose match, use DWPose/OpenPose + ControlNet instead of text.

All credit for JoyCaption goes to fpgaminer. This is just a ready-to-use workflow around it.