Dart Prompt V2 🍌

This is Dart Prompt's successor, built for a newer Dart checkpoint that understands a couple of control axes the original didn't. Same job - assemble the structured seed string DartGenerate expects out of separate fields instead of a hand-typed blob - but with more knobs to steer what comes back.

What changed from V1

Side by side with the original Dart Prompt, three fields are new and one is gone:

• aspect_ratio - a combo, defaulting to tall. Lets you hint the target image's proportions, which matters because tag frequency correlates with framing - a "tall" portrait pulls different pose and crop tags than a wide landscape composition would.
• length - replaces V1's long combo with the same idea under a clearer name, defaulting to medium.
• identity - a combo defaulting to none. Reads as a control over how strongly the model should lock onto a specific character's known visual identity versus generating more loosely around the character tag - useful if you want Dart to stay conservative on an established character rather than drifting.

The core fields carry over unchanged: rating (Danbooru's content-rating axis), copyright (series, default original), character (a character tag), and general (your actual seed tags, multiline). Output is a single STRING, same as V1, meant to feed straight into DartGenerate's prompt input.

Why a V2 exists at all rather than just upgrading V1

This is the same pattern the pack repeats elsewhere: rather than breaking existing workflows when a newer, more capable Dart checkpoint ships, the author added a second node alongside the first. If your workflow was built around Dart Prompt and works fine, there's no obligation to switch - V2 is for when you specifically want aspect-ratio or identity control and are willing to pair it with a Dart checkpoint that actually understands those tokens (check what Load Dart is pointed at; not every Dart-format model on Hugging Face necessarily supports the newer control axes).

Installing it

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

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

Restart ComfyUI. No model download tied to this node specifically - it's string assembly, and the actual weights load back at Load Dart.

Common issues

The one trap specific to V2: if aspect_ratio or identity do nothing to your generated tags, the most likely explanation is a mismatch between this node and the Dart checkpoint you've loaded - the original p1atdev/dart-v1-sft default in Load Dart predates these control axes, so if you haven't pointed the tokenizer/model fields at a newer Dart release, V2's extra fields may just be silently ignored by a model that never learned to condition on them. Beyond that, the same general-purpose caveats as V1 apply: obscure characters produce thin completions because the model can only extrapolate from co-occurrence it actually saw in training, and rating still quietly shapes the whole tag distribution before your general tags get a say, so check it first if output feels off-target.