ToriiGate Grounding Builder Reforged

This node never touches a model. It's a string formatter, and that's exactly why you want it. ToriiGate Caption Reforged will happily caption an image all on its own, but with the Grounding Builder wired into its prompt input you get to steer what the model looks for - tags, character names, per-character details - instead of getting a generic "describe this picture" every time.

It works by assembling the caption instruction in ToriiGate's own prompt format. The model is a fine-tuned Qwen3.5-4B whose captioning quality depends heavily on the structure of the instruction it receives, so this node is the structured front-end to the pack's unstructured back-end. Set your switches, type your tags, and it emits one prompt STRING that plugs straight into the Caption node's optional prompt pin.

The inputs that actually matter

• caption_type - the format and detail level, ten modes total. The README's own recommendation: long for general use (2–5 paragraphs of natural text). short is fastest. long_thoughts_v2 is the max-detail mode: character identification, key details, a long description, then per-character breakdowns. There are also structured modes - json, min_structured_md, min_structured_json - and comic-specific ones (json_comic, md_comic) if you caption pages.
• add_tags + tags - comma-separated booru tags. These match how the anime-tag lineage is trained, so if you know the image's tags, this is where they go.
• add_character_list + character_names + character_count - a general list of who's in the frame, up to five. Per-character fields inherit from this list by position if you leave their names empty.
• add_character_tags and add_character_descriptions - reveal the char1_name through char5_name, charN_tags, and charN_description fields. The toggles exist so you only see what you're actually filling in; the fields are all there in the node regardless.

Two honest warnings. First, use_names (on by default) tells the model it may try to recognize character names - and the README is upfront that identification accuracy "may not always be ideal," with both quantized and original weights. That's the same multi-subject attribution weakness every VLM captioner has; don't expect it to reliably know who a character is, even with the name handed to it. Second, these toggles only matter when a mode actually uses them. add_tags gates the tags field, add_character_list gates the character list - flip a mode and ignore a section and the instruction simply won't contain it.

Where it slots in

Build the grounding, then caption:

Load CLIP ──────────────> ToriiGate Caption Reforged.clip
Load Image ─────────────> ToriiGate Caption Reforged.image
Grounding Builder.prompt ─> ToriiGate Caption Reforged.prompt
ToriiGate Caption.caption ─> Preview Any.source

For a training set, this is the tag-versus-natural-language decision played out in one node: short/long for natural-language captioning on an LLM-encoded base, the structured modes when you want machine-parseable output. It's the piece that turns a generic VLM captioner into something you can point at a dataset and trust to follow the same recipe every time.

Install

It ships in the ComfyUI-ToriiGate-Reforged pack, so it installs with the pack: clone https://github.com/CocyNoric/ComfyUI-ToriiGate-Reforged into ComfyUI/custom_nodes/ and restart, or grab it via ComfyUI Manager. No dependencies of its own, and no model - the model belongs to the Caption node you connect it to. The one failure mode worth naming: this node alone produces text that does nothing. If your workflow is just a Grounding Builder floating in space, nothing broke - you're just missing the other half of the pack.