Dart Generate 🍌

This is where Dart actually earns its keep. You feed it a sparse seed - a character name, maybe a rating, a couple of tags you care about - and it runs the loaded language model to predict a fuller tag list around them, the same way a text LLM completes a sentence. It's the sibling idea to KBlueLeaf's DanTagGen, just from a different author (p1atdev), and both exist because hand-writing 20+ well-chosen Danbooru tags every single generation gets old fast.

The mechanism, briefly

Dart was trained on real Danbooru tag co-occurrence data, so it's learned which tags actually show up together - which clothing goes with which pose, which quality tags cluster with which content. DartGenerate runs it as a proper language model: tokenize your seed prompt, sample forward token by token (or tag by tag) according to whatever DartConfig you hand it, and stop once it's produced a coherent tag sequence. Unlike a diffusion sampler, there's no image involved yet - this is purely text in, text out, and the output is meant to be dropped straight into your CLIP Text Encode node afterward.

Inputs and outputs

Required:

• tokenizer / model - the DART_TOKENIZER and DART_MODEL from Load Dart.
• prompt - your seed string, e.g. 1girl, hatsune_miku, blue hair.
• batch_size - how many independent tag-list variations to generate at once, up to 4096.
• seed - for reproducibility, same idea as any diffusion seed.

Optional:

• config - a DartConfig node; skip it and Dart falls back to its own defaults.
• negative - a string of tags you want the model to steer away from during generation.
• ban_tags - tags that should never appear in the output at all, a harder constraint than negative.

Two outputs: BATCH_STRING (all generated variants, useful if batch_size > 1 and you want to feed a batch of prompts downstream) and a single STRING (handy when you just want one result to plug straight into a text encode node).

Installing it

ComfyUI Manager: search "cgem156-ComfyUI". Manual clone:

cd ComfyUI/custom_nodes
git clone https://github.com/laksjdjf/cgem156-ComfyUI

Restart after. This node itself needs nothing extra - all the weight-downloading happens back at Load Dart, the first time you run that node.

Where this goes sideways

Because there's no image encoder in this loop, low-VRAM setups shouldn't have trouble here - Dart is a small model compared to anything doing actual diffusion. The more common snags: if negative and ban_tags overlap heavily with your prompt, you're fighting the model against itself and you'll get thin, generic output; keep the seed and the exclusions on genuinely different axes. If output reads as incoherent tag soup, that's almost always a DartConfig sampling issue (temperature/top_p too loose) rather than this node - check that first before assuming Dart itself is broken. And remember this pack's README flags that some of its nodes assume you also have pythongosssss's ComfyUI-Custom-Scripts installed for its ShowText node - if you want to actually read the generated tag string in the UI without routing it all the way to a CLIP encode first, wiring the STRING output into ShowText is the easiest way to sanity-check what Dart produced before you commit to a full render.