Tao Generate Referring Expressions

NVIDIAOfficialCreated by NVIDIA Corporation

Skill
Workflow
Data and AI
Apache-2.0
Medium risk

From the author

Four-step image referring-expression pipeline: turns images plus KITTI bounding-box labels into region descriptions, scene captions, grounded referring expressions, and (optionally) verified expressions via VLM distillation. Use when the user wants to generate referring-expression annotations from images with KITTI labels, build region descriptions, produce grouped grounding phrases tied to bboxes, run a double-check verification pass on grounding expressions, auto-label traffic / scene images for referring datasets, or run the image_referring_expression pipeline. Triggers include 'referring expression', 'region description', 'KITTI labels', 'spatial relationship annotation', 'auto-label image referring expression', 'image_referring_expression'.

— the description by NVIDIA Corporation

AutoGineer summary

Written by AutoGineer from the skill’s own files, not by the author

The tao-generate-referring-expressions skill provides an automated pipeline for transforming images and KITTI-format bounding box labels into rich, grounded referring expressions. By leveraging a Vision-Language Model (VLM) as a teacher, the pipeline performs four distinct steps: generating region-specific descriptions, creating holistic scene captions, producing grouped grounding phrases tied to bounding boxes, and optionally verifying the accuracy of these expressions. This tool is designed to assist in auto-labeling traffic or scene images for training datasets, offering a structured approach to data annotation that supports both Gemini and OpenAI-compatible VLM endpoints.

What it does

Generates discriminative region descriptions for objects defined by KITTI bounding boxes.
Produces holistic, location-agnostic scene captions for input images.
Groups objects into semantic phrases with associated bounding box lists.
Performs an optional verification pass to tighten or remove inaccurate bounding boxes.
Supports multiple VLM backends including Gemini and self-hosted OpenAI-compatible services.
Outputs unified annotations in JSONL format for downstream machine learning tasks.

When to use it

When you need to auto-label traffic or scene images for referring expression datasets.
When you have existing KITTI-format bounding box labels and want to add descriptive text annotations.
When you need to verify and refine existing grounding expressions for higher accuracy.
When you want to build a dataset of grounded referring expressions tied to specific image regions.

Advantages

Automates complex annotation tasks using powerful VLM capabilities.
Provides a modular workflow where steps can be enabled or disabled based on requirements.
Includes a built-in verification step to improve the quality of generated labels.
Flexible integration with various VLM endpoints including local self-hosted servers.
Supports both modern JSONL output and legacy formats for compatibility with existing tools.

Usage examples

“Run the referring expression pipeline on my traffic images.”: The agent will guide you through setting up the image and label directories, selecting a VLM endpoint, and executing the four-step annotation process.
“How can I verify my existing grounding expressions?”: The agent will configure the pipeline to run the Step 3 double-check verification pass on your existing annotations to identify and correct mismatches.

Install

Download the skill as a ZIP:
https://autogineer.com/agent-skills/nvidia/tao-generate-referring-expressions/download

Unzip it into the cross-client skills folder of your project:

mkdir -p .agents/skills && unzip tao-generate-referring-expressions.zip -d .agents/skills

Documentation by NVIDIA Corporation

Image Referring Expression Pipeline

Standalone install? If this session was not initialized by the TAO skill bank plugin, run the tao-setup skill first (host preflight, credentials, cross-skill discovery).

Generate referring-expression and grounding annotations from images with KITTI-format bounding box labels. A single VLM (Gemini or any OpenAI-compatible endpoint) runs four steps: per-object region descriptions, holistic image captions, grouped grounding expressions tied to bboxes, and an optional double-check verification pass.

Purpose

Transform (image, KITTI labels) pairs into a unified annotations.jsonl containing rich, grounded referring expressions. The VLM acts as a "teacher" annotator: Steps 0-1 see the image; Step 2 groups Step 0 outputs into grouping phrases with bbox lists; Step 3 (optional) re-examines those bboxes against the image and corrects mismatches.

Pipeline Architecture

Step 0: Region expression  ──┐
                              ├──▶  Step 2: Grounding expression  ──▶  [Step 3: Double check]
Step 1: Image caption  ──────┘                                                   (optional)

Step 0 (region_expr) — VLM emits one short discriminative phrase per KITTI bbox (bbox_2d, type, color, description).
Step 1 (image_caption) — VLM emits a holistic, location-agnostic scene caption.
Step 2 (grounding_expr) — VLM groups Step 0 objects into grouping phrases and returns one bbox list per group, optionally using Step 1's caption as extra context.
Step 3 (double_check) — VLM re-checks each Step 2 bbox against the image; bad matches are removed, slightly-off boxes get tightened.

Steps 0 and 1 run in parallel within a single thread pool (they only depend on the seed records). Each step writes its own step_<N>_*/annotations.jsonl and skips already-processed images on re-run unless workflow.force_reprocess: true.

Instructions

Initial setup

When a user wants to run this pipeline, walk through these steps:

Images: Ask for data.image_dir, the directory containing .jpg, .jpeg, or .png images.
KITTI labels: Ask for data.kitti_label_dir, the directory containing one .txt label file per image. Each label line must use KITTI format: <type> <truncated> <occluded> <alpha> <bbox_left> <bbox_top> <bbox_right> <bbox_bottom> .... Lines with fewer than 8 fields are silently skipped. Set this even for Step 1-only runs because Steps 0 and 2 require it.
Resume from existing annotations: If the user already has a unified annotations.jsonl from a previous run, set data.input_annotations_jsonl to that file instead of seeding from data.image_dir and data.kitti_label_dir.
API access: Ask the user which VLM endpoint they want to use. Present these five options and act on the choice:
1. Gemini — set vlm.backend: "gemini"; require GOOGLE_API_KEY (env var or vlm.gemini.api_key).
2. NIM (e.g. https://inference-api.nvidia.com/v1) — set vlm.backend: "openai"; collect base_url, model_name, and api_key.
3. TAO inference microservice (self-hosted, OpenAI-compatible). Confirm whether the server is already running:
  - Running — collect base_url, model_name, and (optionally) api_key; set vlm.backend: "openai".
  - Not running — guide the user through the skills/applications/tao-run-inference-service skill, which stands up a local TAO inference microservice with an OpenAI-compatible API. Before promising a specific model, check skills/applications/tao-run-inference-service/references/service.yaml for valid_network_arch_config_basenames. Once the server is up, collect base_url, model_name, and (optionally) api_key; set vlm.backend: "openai".
4. vLLM (self-hosted, OpenAI-compatible). Confirm whether the server is already running:
  - Running — collect base_url, model_name, and (optionally) api_key; set vlm.backend: "openai".
  - Not running — follow references/vllm_server.md to install and launch a vLLM server, then collect base_url, model_name, and (optionally) api_key; set vlm.backend: "openai".
5. Custom (any other OpenAI-compatible endpoint) — set vlm.backend: "openai"; collect base_url, model_name, and (optionally) api_key.
If the user has no endpoint and does not want to set one up, stop and help resolve API access first.
Workflow steps: Choose one of:
- Full pipeline: ["0", "1", "2", "3"]
- No caption generation: ["0", "2", "3"], where Step 2 falls back to image-only context
- No verification: ["0", "1", "2"]
- Custom subset: any supported subset of steps
Output format: Choose one of:
- jsonl: unified schema only
- legacy: byte-compatible .txt.stepN files only
- both: writes both formats and is the default for downstream tooling

Running the pipeline

The pipeline runs inside the TAO Toolkit container via the auto_label CLI:

auto_label generate -e /path/to/spec.yaml \
    results_dir=/results \
    image_referring_expression.data.image_dir=/data/images \
    image_referring_expression.data.kitti_label_dir=/data/labels \
    image_referring_expression.vlm.gemini.api_key=$GOOGLE_API_KEY

Generate a default spec: auto_label default_specs results_dir=/results module_name=auto_label, then set autolabel_type: "image_referring_expression". All fields support Hydra dot-notation overrides on the command line.

See references/configuration.md for the full YAML structure, all parameters, model/endpoint setup, and error patterns.

Recommended pilot workflow

Run on 5-10 images with all four steps.
Inspect step_0_region_expr/annotations.jsonl — are object types, colors, and discriminating phrases accurate?
Inspect step_2_grounding_expr/annotations.jsonl — are objects grouped sensibly, and do bbox coordinates match the described groups?
Inspect step_3_double_check/annotations.jsonl — were mismatched bboxes removed or tightened? Are any new errors introduced (rare)?
If quality is insufficient, switch the VLM to a stronger model (e.g. gemini-2.5-pro or a larger Qwen3-VL endpoint), raise media_resolution / max_output_tokens, then re-run with workflow.force_reprocess=true.
Scale to the full dataset once satisfied.

Configuration

Key configuration fields (full reference in references/configuration.md):

Field	Default	Description
`workflow.steps`	`["0","1","2","3"]`	Which steps to execute (`0`=region_expr, `1`=image_caption, `2`=grounding_expr, `3`=double_check)
`workflow.max_workers`	`4`	Parallel threads per step (watch API rate limits)
`workflow.force_reprocess`	`false`	Ignore cached per-step outputs and reprocess from scratch
`workflow.output_format`	`"jsonl"` (set to `"both"` in the default spec)	`"jsonl"`, `"legacy"`, or `"both"`
`vlm.backend`	`"gemini"`	`"gemini"` or `"openai"` (OpenAI-compatible endpoint)
`data.image_dir`	required	Directory of input images (`.jpg` / `.jpeg` / `.png`)
`data.kitti_label_dir`	required (unless resuming)	Directory of KITTI-format `.txt` label files
`data.input_annotations_jsonl`	`""`	Optional pre-seeded `annotations.jsonl` (skips KITTI seeding)

Inputs

Two ways to seed the pipeline:

Image directory + KITTI labels (default). Set data.image_dir and data.kitti_label_dir. The orchestrator walks the image directory, reads the matching <stem>.txt KITTI file, parses bboxes (fields 0 + 4-7), reads each image's width/height via PIL, and writes a seed_annotations.jsonl to results_dir/.
Pre-seeded annotations JSONL (resume / pre-computed regions). Set data.input_annotations_jsonl to a file with one {"image_id", "image_path", "width", "height", "kitti_bboxes": [...]} object per line.

Outputs

All outputs go to results_dir/:

seed_annotations.jsonl — initial per-image records (unless input_annotations_jsonl was supplied).
step_0_region_expr/annotations.jsonl — adds regions[] (each with bbox/bbox_2d, type, color, description).
step_1_image_caption/annotations.jsonl — adds caption (string).
step_2_grounding_expr/annotations.jsonl — adds expressions[] (each {text, instances: [{bbox: [x1,y1,x2,y2]}]}).
step_3_double_check/annotations.jsonl — same shape as Step 2, with bboxes removed/updated.
results_dir/annotations.jsonl — copy of the last completed step's output.
When workflow.output_format is "legacy" or "both", each step also writes byte-compatible step_<N>_*/labels/<stem>.txt.stepN files for the original 2d-data-engine tooling.

Prerequisites

Container: nvcr.io/nvidia/tao/tao-toolkit:7.3.0-pyt
API access: At least one VLM endpoint (Gemini API key or OpenAI-compatible endpoint capable of image input)
PIL / Pillow: Required to read image dimensions during seeding (already present in the TAO container)

Files

ATTRIBUTION.mdAdded by AutoGineer972 B
BENCHMARK.md7.3 KiB
LICENSELicence file added by AutoGineer11.2 KiB
SKILL.md10.2 KiB
config/skillspector-baseline.yaml834 B
evals/evals.json849 B
references/configuration.md12.7 KiB
references/skill_info.yaml1.5 KiB
references/vllm_server.md6.8 KiB
skill-card.md4.8 KiB
skill.oms.sig5.4 KiB

references/vllm_server.md

vLLM Server Setup for Vision-Language Models

Use this reference only when the parent SKILL.md points here for the current task. If this file conflicts with current SKILL.md, skill_info.yaml, schemas, or platform/model skills, the current authoritative source wins.

1. Prerequisites
1. Install vLLM
- Option A — Docker (default, recommended)
- Option B — pip (host install, advanced)
1. Pick a VLM checkpoint
1. Launch the server
- Option A — Docker (default, recommended)
- Option B — pip install (host launch, advanced)
1. Verify the endpoint
1. Wire the server into the pipeline
Common issues

This guide walks the user through standing up a self-hosted vLLM server that exposes an OpenAI-compatible /v1/chat/completions endpoint for vision-language models (VLMs). Once the server is running, point the pipeline at it with vlm.backend: "openai" and the matching base_url / model_name / api_key values.

1. Prerequisites

NVIDIA GPU(s) with enough VRAM for the chosen VLM (≥24 GB for 7-8B-class models, ≥80 GB for 32B+ models; tensor-parallel across multiple GPUs is supported via --tensor-parallel-size).
A recent NVIDIA driver and nvidia-container-toolkit (for the recommended Docker path).
Docker (recommended). Python ≥ 3.10 is only needed for the optional host install path.
For gated HuggingFace repos: a HF_TOKEN with access to the model.

2. Install vLLM

Default: Docker (Option A). Use it unless you have a specific reason to install on the host — Docker pins a known-good CUDA/PyTorch/vLLM combination and avoids local environment drift. The pip path (Option B) is provided for advanced users who need to patch vLLM, debug locally, or run on a system where Docker is not available.

Option A — Docker (default, recommended)

docker pull vllm/vllm-openai:latest

Option B — `pip` (host install, advanced)

pip install --upgrade "vllm>=0.7.0"

3. Pick a VLM checkpoint

vLLM supports many vision models; the pipelines just need an OpenAI-compatible chat-completions endpoint that accepts image inputs. Common picks:

Model	HuggingFace repo	Notes
Qwen3-VL-8B-Instruct	`Qwen/Qwen3-VL-8B-Instruct`	Newer Qwen3 family
Qwen3-VL-235B-A22B-Instruct	`Qwen/Qwen3-VL-235B-A22B-Instruct`	MoE; requires serious hardware

If the chosen repo is gated, accept its license on the HuggingFace web UI first, then make HF_TOKEN available in your session before launching.

4. Launch the server

Use the launcher that matches the install path. Both commands listen on 0.0.0.0:8000 and serve an OpenAI-compatible API at /v1. Prefer Option A (Docker) unless you installed via pip in Section 2.

Option A — Docker (default, recommended)

set -a; source /path/to/.env; set +a   # omit if already exported
docker run --runtime nvidia --gpus all \
    -p 8000:8000 \
    -e HF_TOKEN \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    vllm/vllm-openai:latest \
    --model Qwen/Qwen3-VL-8B-Instruct \
    --served-model-name Qwen/Qwen3-VL-8B-Instruct \
    --tensor-parallel-size 1 \
    --gpu-memory-utilization 0.9 \
    --media-io-kwargs '{"video": {"num_frames": -1, "fps": -1}}'

Option B — `pip` install (host launch, advanced)

set -a; source /path/to/.env; set +a   # omit if already exported

vllm serve Qwen/Qwen3-VL-8B-Instruct \
    --host 0.0.0.0 \
    --port 8000 \
    --dtype bfloat16 \
    --served-model-name Qwen/Qwen3-VL-8B-Instruct \
    --tensor-parallel-size 1 \
    --max-model-len 32768 \
    --gpu-memory-utilization 0.9 \
    --media-io-kwargs '{"video": {"num_frames": -1, "fps": -1}}'

Key flags:

--served-model-name <NAME> — value to use later for vlm.openai.model_name. Defaults to the full HF repo path if omitted.
--tensor-parallel-size <N> — number of GPUs to shard the model across.
--max-model-len <N> — context window; image tokens count against this, so leave headroom for multi-image prompts.
--limit-mm-per-prompt image=<N> — max images per request. Bump it if a prompt sends multiple images.
--gpu-memory-utilization <0..1> — lower (e.g. 0.85) if you hit OOM at load time.

The first launch downloads weights to ~/.cache/huggingface; expect several minutes for 7B+ models.

5. Verify the endpoint

Wait for the log line Uvicorn running on http://0.0.0.0:8000, then sanity-check the server.

List served models:

curl http://localhost:8000/v1/models

Send a minimal vision request:

curl http://localhost:8000/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "Qwen3-VL-8B-Instruct",
        "messages": [{
            "role": "user",
            "content": [
                {"type": "image_url", "image_url": {"url": "https://upload.wikimedia.org/wikipedia/commons/thumb/3/3a/Cat03.jpg/640px-Cat03.jpg"}},
                {"type": "text", "text": "Describe this image in one sentence."}
            ]
        }],
        "max_tokens": 128
    }'

A non-empty choices[0].message.content in the response confirms the server is ready.

6. Wire the server into the pipeline

Once verified, collect these three values from the running server and pass them to the pipeline spec:

base_url — e.g. http://localhost:8000/v1 (no /chat/completions suffix). If vLLM runs on another host, use http://<host_or_ip>:8000/v1 and make sure the port is reachable.
model_name — must match --served-model-name exactly.
api_key — vLLM ignores it but the OpenAI SDK requires a non-null string; use "EMPTY" if no auth is configured.

YAML snippet:

vlm:
  backend: "openai"
  openai:
    base_url: "http://localhost:8000/v1"
    model_name: "Qwen3-VL-8B-Instruct"
    api_key: "EMPTY"
    temperature: 0.3
    max_tokens: 4096
    timeout: 300

Common issues

Symptom	Fix
`CUDA out of memory` on startup	Lower `--max-model-len`, drop `--gpu-memory-utilization` (e.g. `0.85`), pick a smaller model, or raise `--tensor-parallel-size` to spread across more GPUs
`Model architectures ['…'] are not supported`	Upgrade vLLM (`pip install -U vllm`) or use a newer Docker tag — VLM support changes per release
401 / 403 during HuggingFace download	Set `HF_TOKEN` in the launch env and accept the model's license on the HuggingFace web UI
First request hangs for minutes	The model is still warming up — wait for the `Uvicorn running` log line and a successful `GET /v1/models`
`image is too large` / token overflow	Pre-resize images before sending, or raise `--max-model-len`
Empty / truncated responses	Raise `vlm.openai.max_tokens` in the pipeline spec; lower `temperature` for more deterministic output

Requirements and compatibility

Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta

“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.

Needs: Declares allowed tools, Needs an API key

The author says: Requires docker + nvidia-container-toolkit + at least one VLM endpoint (Gemini API key or OpenAI-compatible).

Security

What our scanner noticed· Medium risk

Large base64-encoded blobs
Instructions that use the network

A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.

Attribution and licence

Created by

NVIDIA Corporation

Publisher

NVIDIA (Official)

Source

NVIDIA/skills · skills/tao-generate-referring-expressions · commit fce3cca

Licence

Apache License 2.0 (redistribution allowed)

LICENSE (added by AutoGineer from LICENSE-APACHE)
ATTRIBUTION.md (written by AutoGineer)

Synced

2026-11-09

AutoGineer is not affiliated with or endorsed by NVIDIA Corporation or NVIDIA.

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download tao-generate-referring-expressions.zip

From the author

AutoGineer summary

What it does

When to use it

Advantages

Usage examples

Install

Documentation by NVIDIA Corporation

Image Referring Expression Pipeline

Purpose

Pipeline Architecture

Instructions

Initial setup

Running the pipeline

Recommended pilot workflow

Configuration

Inputs

Outputs

Prerequisites

Files

references/vllm_server.md

vLLM Server Setup for Vision-Language Models

Contents

1. Prerequisites

2. Install vLLM

Option A — Docker (default, recommended)

Option B — pip (host install, advanced)

3. Pick a VLM checkpoint

4. Launch the server

Option A — Docker (default, recommended)

Option B — pip install (host launch, advanced)

5. Verify the endpoint

6. Wire the server into the pipeline

Common issues

Requirements and compatibility

Security

Attribution and licence

Related skills

Claude API

Transformers

Open Notebook

Vision Sft

LLM Evaluation

Lora Qlora Recipes

Download ZIP

Option B — `pip` (host install, advanced)

Option B — `pip` install (host launch, advanced)