Pairs naturally with kijai/ComfyUI-Florence2: use Florence-2 in od, dense_region_caption, or caption_to_phrase_grounding mode to produce per-object bounding boxes and English labels, then feed them into this node to render a study-ready image.

What it does

Inputs:

An image (e.g. a scene generated by SDXL / Flux / Z-Image-Turbo)
A list of bounding boxes [[x1, y1, x2, y2], ...]
A list of English labels ["apple", "chair", ...]

Output:

The same image with a colored rectangle around each detected object and the English word printed in a colored label box (above / inside / below the bbox)

A built-in 8-color palette keeps adjacent labels visually distinct. The node is permissive about input shape — it accepts flat lists, batch-wrapped lists, or lists of {"bbox":..., "label":...} dicts, so it works across Florence-2 wrapper versions.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/aadebuger/ComfyUI-EnglishAnnotator.git
cd ComfyUI-EnglishAnnotator

# inside ComfyUI's venv
uv pip install -r requirements.txt

Restart ComfyUI. The node appears under category 📚 English Annotator as 📚 English Word Annotator.

See INSTALL_NOTES.md for environment-specific notes (Ubuntu + uv, HF cache redirect, etc.).

Node reference

Input	Type	Default	Notes
`image`	IMAGE	—	The source image
`bboxes`	any	—	Bounding boxes (typically from Florence-2)
`labels`	any	—	Labels (typically from Florence-2)
`font_size`	INT	36	Label text size in pixels
`box_width`	INT	4	Bounding box outline width
`label_position`	enum	`above`	`above` / `inside_top` / `below`, auto-fallback when out of bounds
`font_path`	STRING (optional)	`""`	Explicit TTF/OTF path (overrides defaults)
`text_color`	STRING (optional)	`#FFFFFF`	Hex color for label text
`max_labels`	INT (optional)	`0`	Cap the number of objects to annotate (0 = no limit)

Returns: IMAGE.

Fonts: the node walks a built-in candidate list — DejaVu (Linux), Helvetica/Arial (macOS), Arial (Windows). On a fresh Ubuntu box install fonts-dejavu-core if labels render as boxes.

Example workflow

The flow is:

CheckpointLoader → KSampler → VAEDecode ──┐
                                          ├→ EnglishWordAnnotator → SaveImage
              Florence2ModelLoader → Florence2Run ──┘

See example_workflows/classroom.json for an API-format workflow you can submit via /prompt, or generate one with the CLI's --dump-workflow flag.

Florence-2 task suggestions

Task	When to use
`od`	Clean COCO-style category names — best for beginner-level vocab
`caption_to_phrase_grounding`	You supply your own word list; Florence-2 locates exactly those words
`dense_region_caption`	Natural-language regions like "red apple on table"

For a fixed vocabulary pipeline (e.g. a daily 16-word list), caption_to_phrase_grounding is by far the most reliable.

CLI

A self-contained CLI is bundled under cli/. It uses ComfyUI's WebSocket API to submit a workflow end-to-end and download the annotated image:

uv run cli/english_card_cli.py \
    "flat vector illustration of a tidy classroom: desk, chair, blackboard, books, pencil, backpack, clock, globe, computer" \
    --ip 10.8.0.12 --steps 25

With a fixed vocabulary (uses caption_to_phrase_grounding):

uv run cli/english_card_cli.py \
    "a sunny kitchen scene, flat vector illustration" \
    --words "apple,banana,knife,cutting board,bowl,kettle,faucet,refrigerator" \
    --ip 10.8.0.12

See cli/README.md for full argument reference.

Use case: daily vocabulary card pipeline

This node was built for an automated daily English vocabulary card workflow:

A daily JSON of 16 target words (with definitions and example sentences)
Scene generation via SDXL / Z-Image-Turbo
Florence-2 caption_to_phrase_grounding locks bboxes onto exactly those 16 words
This node renders the annotated image
VoxCPM2 TTS for per-word audio
ffmpeg assembles the final card video

License

MIT