Dart Load

Dart Load is the entry point for the whole ComfyUI-Dart pack, and it does exactly one job: it hands you a language model and its tokenizer, tuned specifically to autocomplete Danbooru-style tag prompts. Not a checkpoint, not a captioner - a small text model that was trained on millions of Danbooru tag lists and learned what tags tend to show up together. Feed it a partial prompt ("1girl, this rating, this character") and it fills in the rest with plausible tags. Everything downstream in this pack - Compose Prompt, Generate, Decode - needs the model and tokenizer this node produces, so it's always the first box on the canvas.

What it actually loads

The model input is a single dropdown with three choices: dart-v1-sft, dart-v2-moe-sft, and dart-v2-sft. These map straight to p1atdev's Dart (Danbooru Tags Transformer) checkpoints on Hugging Face - the underlying research model this pack wraps. v1 is the original release; the v2 line adds an aspect-ratio-aware prompt format and a mixture-of-experts variant (moe). Which one you pick matters more than it looks, because it decides which "Compose Prompt" node downstream actually understands your output: v1 models want the plain Dart Compose Prompt node, v2 models want Dart Compose Prompt V2. Mixing them - say, loading dart-v2-sft but feeding it a v1-composed prompt string - gets you a prompt format the model wasn't trained to parse, and generation quality falls off fast.

Two outputs come off this node: DART_MODEL and DART_TOKENIZER. Both are custom types unique to this pack - they won't plug into anything outside of ComfyUI-Dart's own nodes. DART_MODEL goes to Dart Generate or Dart Generate(Advanced); DART_TOKENIZER fans out much wider, since almost every downstream node (Ban Tags From Regex, Decode, Decode By Splited Parts, Remove Tag Token, Rearranged By Animagine) needs the tokenizer to translate between tag text and token IDs.

Installing it

Two ways in, both from the README:

• ComfyUI Manager - search for "ComfyUI-Dart" and install.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/nkchocoai/ComfyUI-Dart.git, then restart ComfyUI.

There's no separate model-download step to do by hand. The model dropdown lists Hugging Face repo IDs directly, and the underlying library pulls the weights the first time you run the node with a given selection. These are small language models, not diffusion checkpoints - nowhere near the size of an SDXL or Flux download - but the first run with a fresh choice still needs a working connection out to Hugging Face. If you're offline, or on a network that blocks huggingface.co, that first execution is where it'll stall.

Where it trips people up

The recurring one is picking the wrong version pairing downstream. If you loaded a dart-v2-* model, use Dart Compose Prompt V2 (which has the aspect_ratio and identity fields v1 doesn't understand); if you loaded dart-v1-sft, use the plain Compose Prompt. Swap them and you'll get either an error from the tokenizer choking on unexpected special tokens, or - worse - a "successful" generation that's just noise, because the model is trying to complete a prompt shape it never saw in training.

The other one is just the first-run pause: switching the model dropdown to a checkpoint you haven't used yet triggers a fresh download, which can look like ComfyUI has hung if you're not expecting it. Give it a minute before you assume something's broken, especially on a slower connection.

Once loaded, DART_MODEL and DART_TOKENIZER stay resident for the rest of your session - you only pay the load (and possibly the download) cost once per model choice, not per generation.