Dart Generate(Advanced)

Dart Generate(Advanced) is the same generation engine as the regular Dart Generate node, minus the part that hands you a finished string. Where Generate decodes its output straight to text, this one stops one step earlier and gives you the raw DART_TOKEN_IDS - the token sequence the model actually produced, before anyone turns it back into readable tags. If you just want tags for your prompt box, you don't need this node; it exists for people who want to do something to the generation before it becomes text.

Why you'd want raw tokens instead of a string

The rest of the pack has a small toolkit that only operates on DART_TOKEN_IDS: Dart Remove Tag Token strips specific tags out of the sequence, Dart Rearranged By Animagine reorders it to match Animagine XL's documented prompt structure, and Dart Decode By Splited Parts breaks it into separate rating/copyright/character/people/other-general strings instead of one blob. All three need token IDs as input, not a finished string - so if your workflow wants to inspect or edit the generation before committing to text, Generate(Advanced) is the node that opens that door. It's the same reason the plain Dart Generate node has an animagine_order toggle built in and a remove_tags filter: this Advanced node is what those conveniences are built on top of, just exposed as separate steps you control yourself.

Inputs and outputs that matter

Required: model (DART_MODEL) and tokenizer (DART_TOKENIZER), both from Dart Load; prompt, a string built by Dart Compose Prompt or Dart Compose Prompt V2 depending on which model you loaded - the default value is the same special-token soup as the regular Generate node (<|bos|><rating>...), so build it with a Compose Prompt node rather than typing it; and seed, which works like an image-gen seed - same seed and prompt reproduce the same output.

Notice what's not here compared to the regular Generate node: no animagine_order toggle and no remove_tags input. Those conveniences are exactly what you'd otherwise get from Dart Rearranged By Animagine and Dart Remove Tag Token downstream - this node assumes you'll wire those in yourself if you want them. The optional setting (DART SETTING from Dart Generation Config) and ban_tags (a string, usually from Dart Ban Tags From Regex) work the same as on the regular node - sampling behavior and generation-time suppression, respectively.

The single output is DART_TOKEN_IDS, which won't plug into a CLIP Text Encode directly. You need Dart Decode (or Dart Decode By Splited Parts) somewhere downstream to turn it into actual text.

Installing it

Install via ComfyUI Manager (search "ComfyUI-Dart") or clone manually: cd ComfyUI/custom_nodes && git clone https://github.com/nkchocoai/ComfyUI-Dart.git, then restart. No extra install step for this specific node - it ships with the rest of the pack, and the model itself is pulled automatically the first time Dart Load runs.

Where people get stuck

The most common trip-up is forgetting there's no text output here and wondering why nothing shows up in the prompt box - this node needs a Dart Decode downstream to become readable. The second is the same version mismatch that hits the regular Generate node: build the prompt with the Compose Prompt node that matches the model you loaded (v1 model → plain Compose Prompt, v2 model → Compose Prompt V2), or the token sequence you get back won't correspond to anything sensible. If you don't actually need to edit the raw tokens - no tag removal, no reordering, no split-part decode - just use the regular Dart Generate node instead; it does the decode for you in one step.