ComfyUI Outfit Caption

Two ComfyUI nodes that each send one outfit reference image to a vision provider and return:

• raw_caption: evidence-first outfit QA text.
• vton_prompt: the same caption adapted for a generic virtual try-on prompt.

Node	Provider	Env var	Source
NVIDIA NIM Outfit Caption	NVIDIA NIM	NVIDIA_API_KEY_COMFYUI	nodes/nvidia_nim.py
Gemini Outfit Caption	Google Gemini	GEMINI_API_KEY_COMFYUI	nodes/gemini.py

Both nodes reject image batches and add themselves under image/captioning. This package supersedes the two single-provider repos comfyui-nvidia-nim-outfit-caption and comfyui-google-gemini-outfit-caption, which now redirect here.

Install

Clone this repository into ComfyUI/custom_nodes, restart ComfyUI, then add either node from image/captioning. It is also published on the Comfy Registry as comfyui-outfit-caption.

No additional Python packages are required beyond ComfyUI's existing Pillow, NumPy, and Torch runtime.

Inputs and API key handling

• image: one ComfyUI IMAGE.
• model: native pull-down list built from the live provider catalog and the models available to the current API account (NVIDIA also probes each candidate with a real request; Gemini intersects the live list with nodes/models_gemini.json).
• custom model (custom_model): explicit escape hatch for an ID not present in the catalog.
• API key environment variable (api_key_env): environment fallback, see table above for the per-node default.
• timeout (seconds), max. image size (MB), and max. output tokens.
• remove person accessories from VTON prompt: keeps raw_caption complete but removes personal/carried accessories such as bags, loose bag straps, eyewear, and jewelry from vton_prompt. Mixed sentences retain their garment colors, materials, construction, and footwear details.
• vton_prompt preserves the described layer order: outer garments cover overlapping underlayers, including sleeves, except where the reference visibly exposes the underlayer.
• Session API key (not saved): masked browser widget. It is sent only to the local ComfyUI server, retained in process memory, excluded from workflow serialization, and never logged. Re-enter it after a restart or node clone.

Set the environment fallback before starting ComfyUI if preferred:

$env:NVIDIA_API_KEY_COMFYUI = "..."
$env:GEMINI_API_KEY_COMFYUI = "..."

Model list and ratings

Each node keeps its own benchmark catalog (nodes/models_nvidia.json, nodes/models_gemini.json) with local ratings, scores, and compatibility exclusions. The last successful verified list is cached under the user's local application cache; after a provider/catalog/probe failure, the cached list keeps stable model labels while the refresh status reports stale/error.

Enter the session key and click Refresh models; the existing model pull-down is updated in place with the live list.

Order is fixed by rating, then caption-quality score only:

1. 🟢 tested, very good (80–100)
2. 🟡 tested, usable (60–79)
3. 🔴 tested, poor (0–59, or no usable final text)
4. ⚪ untested

Benchmark v1.0

Use five distinct outfit images with the exact same built-in prompt. Score each image from 0–4 in each category:

• garment coverage
• color and material
• construction
• footwear and accessories
• hallucination avoidance

The five images produce a maximum of 100 points. Store model ID/version, date, benchmark version, total score, rating, and a short reason in the node's models_*.json.

Checks

python -m unittest discover -s tests -v
python -m py_compile nodes/nvidia_nim.py nodes/gemini.py
node --check web/nvidia_secret.js
node --check web/gemini_secret.js

Publishing to the Comfy Registry

pyproject.toml carries the registry metadata (PublisherId = "delcado"). .github/workflows/publish_action.yml publishes automatically whenever pyproject.toml's version is bumped on main, using the COMFY_REGISTRY repository secret (a Comfy Registry API key for the delcado publisher).

API references: NVIDIA NIM API, NVIDIA model catalog, Gemini generateContent, Gemini models.list.