VNCCS Dataset Generator

Everything else in VNCCS is about using a character consistently. This node is about the step after that: taking a character you've already built a full sprite set for and packaging it as a dataset - images plus captions, on disk, ready to hand to a LoRA trainer if you want that character baked into a checkpoint rather than regenerated through VNCCS every time.

What it does and why it matters

The community consensus on character LoRA training (covered at length elsewhere in this ecosystem) is that 15–25 varied images with decent captions is the minimum viable dataset, and that the tedious part is rarely the training itself - it's assembling and captioning a clean set of images in the first place. VNCCS already has a folder full of consistent, pose-varied, background-removed sprites for every character it's generated, which is exactly what a good training dataset looks like. Dataset Generator is the bridge: pick a character, name the project, optionally add a caption fragment to apply across the set, and it writes out a ready-to-use dataset folder rather than making you manually zip and caption a hundred PNGs by hand.

Inputs and outputs that matter

• character (enum, populated from your existing VNCCS characters) - which character's sprite set to export. This list is empty until you've actually generated at least one character.
• game_name (STRING, default "VN") - a label for the export, likely used as a project/folder identifier for the generated dataset.
• Optional additional_caption (STRING, multiline) - extra caption text applied on top of whatever VNCCS already knows about the character (name, tags from creation), useful for adding a trigger word or a style note that should appear in every caption.
• Output: dataset_path (STRING) - where the generated dataset landed on disk.

Installing it

Bundled with the base pack. ComfyUI Manager: search "VNCCS - Visual Novel Character Creation Suite", install latest, restart. Manual: cd ComfyUI/custom_nodes && git clone https://github.com/AHEKOT/ComfyUI_VNCCS.git, restart. No extra model downloads for this node specifically - it operates on sprites you've already generated, so it's only useful once a character actually has output to draw from.

Common issues & troubleshooting

The character dropdown is empty or missing a character you expect. Dataset Generator only knows about characters that have already been through generation - if you just created a character in Character Creator but haven't run it through a generator node yet, there's no sprite set to export. Run the character through Character Generator or Clothes Generator first.

Not sure what to put in additional_caption. If you're planning to train a LoRA from the export, the standard practice elsewhere in the ecosystem is a short, consistent trigger word or phrase that doesn't collide with common vocabulary - something the trainer can reliably associate with this specific character rather than a generic word the base model already has strong opinions about.

Training on the exported set produces a LoRA that doesn't generalize well. This traces back to how varied your original VNCCS sprite set actually is, not to Dataset Generator itself - a character generated in only one or two poses and one costume will export a dataset with the same limited variety. Run the character through more poses and at least a couple of outfits before exporting if you want a LoRA that holds up outside the exact scenarios you generated.