Ideogram 4 LoRA Tagger

Ideogram 4 wants a very particular kind of dataset caption. Not a sentence, not a Danbooru tag list - a JSON object with per-character bounding boxes, hex color palettes, lighting notes, camera info, and a strict rule that the trigger token does all the identity work and you never describe a known character's hair, eyes, or face. Writing that by hand for hundreds of images is misery. This node builds the instruction that makes a vision model do it for you.

Ideogram 4 LoRA Tagger is part of the Consistent Character Creator (CCC) toolset inside Mickmumpitz-Nodes. It doesn't caption anything itself - it's a prompt assembler. It produces one string, final_prompt, which is the complete system prompt you hand to a vision-language model (Qwen-VL, JoyCaption, whatever you caption with) along with each training image. The node's whole job is to package the captioning spec and your character roster into something the VLM can follow.

What it does

Think of it as a fill-in-the-blank template with two parts:

• base_instruction - a long, editable JSON-spec prompt (the default is dense and opinionated: describe left to right, order elements by x_min, one object per bounding box, boxes normalized to 0–1000, hex palettes, the "forbidden rule" about permanent features, and the clothing rule).
• trigger_1 through trigger_8, each with a matching desc_1 through desc_8 - your known-character roster. The trigger is the exact token the LoRA must learn; the desc is a recognition aid so the VLM can tell which face is which character.

The output final_prompt slots into the roster section of the spec and goes straight into your VLM's prompt input. The descs are what let a model look at a photo and know "that's Ma3hwaKang, not MagnusSt3rn."

Why it matters for your dataset

The LoRA-training rule that survives across every architecture: caption what you want to stay variable (pose, clothing, lighting) and leave what should be permanent undescribed, so the trigger token absorbs it. This node encodes that as a hard rule - the "forbidden rule" forbids the VLM from describing permanent physical features of any known character. That's the difference between an Ideogram LoRA that actually holds identity and one that drifts.

Inputs that matter

• trigger_N / desc_N - fill as many pairs as you have known characters (up to 8). More than 8 known characters and you'll need to prioritize.
• base_instruction - leave it alone until you've seen what the VLM produces; tweak it when captions come back with boxes in the wrong place or too much meta-commentary.

Where it fits

Typical CCC pipeline: batch-load your training images, run Prompt Studio (or this tagger) → VLM → caption output, review, train. If your downstream tooling expects x-first boxes in absolute pixels instead of Ideogram's normalized y-first format, run the caption through the pack's BBox Converter node first - it rewrites every [a, b, c, d] box between the two conventions.

Installing

ComfyUI Manager, search "Mickmumpitz", install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/mickmumpitz/ComfyUI-Mickmumpitz-Nodes

Dependencies are just numpy, Pillow, opencv-python. The tagger itself is pure string assembly - no API, no key, no model download. The VLM you pair it with is the only heavy thing, and that's your choice.

One honest warning: this produces structured JSON captions, which are great for training consistency but only if your captioner reliably emits valid JSON. If you get malformed output, tightening the desc fields and the base instruction usually fixes it faster than retraining anything.