TagRAG Image

TagRAG Image is the pack's bridge from "I have a picture" to "I have a better prompt." Feed it Danbooru-style tags from an external tagger node (or type them), and it retrieves related tags from your CSV dataset. Flip refine on and a local LLM turns the whole combined set into a structured prompt. It's the node that connects the tag-to-prompt world to the image-to-tag world without you copy-pasting anything.

The key design decision, and the one to read first: it's tag-string first. If you provide the tags input, that wins - no image tagging happens at all. That's deliberate: it keeps the node deterministic and composable with whatever tagger you already run, since the pack author isn't betting on you switching taggers to use this. The built-in built_in_wd14 backend exists for when you want tagging in-node, but it's strictly opt-in.

How it works

If tags is blank and an IMAGE tensor is wired in with image_tagging_backend=built_in_wd14, the node runs a WD14-style ONNX tagger - the same Danbooru-vocabulary tagger lineage the anime-model crowd has used for years, and the right tool for models trained on booru tags (the KB's danbooru-tags essay covers why: a correct tag beats a sentence on those checkpoints). threshold (0.35) and character_threshold (0.85) control which tags survive; exclude_tags and replace_underscores are the usual cleanups.

Then the retrieval half kicks in regardless of backend: query = tags + optional idea, matched against your index_ref's CSV via lexical/semantic/hybrid search, and retrieved_tags comes back with related tags your dataset knows. If refine is true it calls the generation backend with the image-prompt template to assemble a polished prompt; with refine off it returns original tags plus related tags with zero LLM calls - the cheap, deterministic path.

Inputs that matter

• index_ref - the JSON string from TagRAG Index.
• tags - comma-separated tags (multiline). Required unless you're using the WD14 path.
• image - IMAGE tensor, only used when image_tagging_backend=built_in_wd14.
• image_tagging_backend - existing_tags (default) or built_in_wd14.
• threshold / character_threshold - WD14 tag cutoffs.
• refine - whether to run the LLM at all.
• model_path - GGUF for native refinement.

Outputs

tags (the tag set actually used), refined_prompt (empty unless refine is on), retrieved_tags (the related tags from your CSV), and status.

Install

ComfyUI Manager - search comfyui-tag-rag - or:

cd ComfyUI/custom_nodes
git clone https://github.com/FrostySDXL/comfyui-tag-rag
cd comfyui-tag-rag
pip install -r requirements.txt

Restart after installing. For the built-in WD14 path you also need image extras (pip install -r requirements-image.txt - onnxruntime + Pillow) and model files placed manually at ComfyUI/models/tag_rag/wd14/<model_name>/model.onnx plus the tag metadata CSV. The node will not download any of that for you - missing deps or files just return an actionable status string. The default tagger_model name it looks for is wd-v1-4-moat-tagger-v2.

Where people get burned

• existing_tags with no tags and no tagger wired up - you get an error telling you exactly that. Either wire an external tagger into tags or switch to the WD14 path.
• WD14 "missing model files" - you have to supply the ONNX model and tag CSV yourself; nothing auto-downloads. Set tagger_model_path if they live somewhere non-standard.
• Both tags and image provided - tags wins, silently. The status string notes it, so glance at that if tagging seems to do nothing.
• Expecting vision. Scope check: the README is explicit that there's no vision-language chat, no Florence2, no ColPali, no llama.cpp vision/mmproj. This node tags or accepts tags, retrieves, and optionally rewrites - that's the whole job. WD14 won't describe a scene either; that's not what it does.

It's pre-alpha (0.1.0a0), so treat it as a tool to poke at. But as a glue node between your existing tagger and your CSV library, it's genuinely useful even with refine left off.