Joy_caption_alpha_prompt

If you've ever stared at the JoyCaption Alpha Two demo and wished you could grab those preset captioning modes without hand-typing them, this is the node for you. Joy_caption_alpha_prompt is a glorified prompt builder: it takes a caption style, a length, and a bunch of optional instruction toggles, and outputs one ready-to-use instruction string for Joy_caption_alpha_run or the batch nodes.

It's the difference between captioning your training set as a generic "describe this image" and telling the model exactly how you want it to write - which is a much bigger deal than it sounds for dataset quality. The same caption model produces very different training captions depending on whether you ask for a "Training Prompt", "Booru tag list", or "Product Listing".

How it works

This is a straight port of the Alpha Two mode system: nine caption types map to template sentences, and the length picker slots into a word-count phrase. The magic is the toggle list - seventeen optional instructions (each one also carries a Chinese translation, a nod to the author's audience) that get appended to the base prompt when you flip them on. Want the captioner to mention lighting and camera angle but never mention resolution? Tick those boxes. The node strips the brackets from the option text and stitches everything together, then substitutes any *name* placeholder with whatever you put in the name field.

The inputs you'll actually touch

• caption_type - nine modes: Descriptive, Descriptive (Informal), Training Prompt, MidJourney, Booru tag list, Booru-like tag list, Art Critic, Product Listing, Social Media Post. For training datasets, "Training Prompt" is the crowd favorite; Booru modes give you tag-style output.
• caption_length - "any", the adjectives (very short → very long), or exact word counts (20/50/80/100/120/250/500). Numeric lengths enforce a word count; adjectives ask for a vibe.
• name - a character name to substitute in. Several toggles only make sense if you set this.

Then the ~17 toggles. The ones that matter most for training: "Do NOT include information about people/characters that cannot be changed (like ethnicity, gender)" and "If there is a person/character in the image you must refer to them as name". For a character LoRA you generally want the former on and the person's changeable attributes described, exactly like the KB's captioning advice: describe what you want to stay variable, leave permanent traits undescribed.

• Output: a single STRING - your built prompt. Wire it into Joy_caption_alpha_run's prompt input or the batch nodes' prompt.

Installing

Same as the rest of the pack - ComfyUI Manager (search "Comfyui_CXH_joy_caption") or:

cd ComfyUI/custom_nodes
git clone https://github.com/StartHua/Comfyui_CXH_joy_caption
cd Comfyui_CXH_joy_caption
python -m pip install -r requirements.txt

Restart and you're done. This node has no model dependencies at all - it's pure string assembly - so it works even before you've downloaded the caption weights.

Notes

• The mode + instruction combo is the same system that shipped in the real Alpha Two release, so results track what you'd see in the fancyfeast demo space.
• One honest caveat: Alpha Two occasionally ignores an instruction or loops in the more exotic modes. That's the model's known quirk (later Beta One releases were trained specifically to fix it), not a bug in this node - re-roll the seed if a toggle seems ignored.