USYD Community

ViTPose+ Large

About

ViTPose+ Large is a 434M open-weight model from USYD Community. 2D human pose estimation (ViT-L, 434M). It accepts text, images, and video. Released January 12, 2025.

Parameters
434M
Published parameter count.
License
Apache-2.0
License on the weights.

Pricing

Token rates are US dollars per 1M tokens. Image rates are per 1K images. Audio rates are per 1M audio seconds.

Rate	Price
Input / 1M	$3.90
Cached input / 1M	$3.90
Output / 1M	$0

Compare

ViTPose+ Large is the smallest pose model here, at 434M, and the one that accepts video.

Model	What it does	Price	Parameters	Context
ViTPose+ Large	2D human pose estimation (ViT-L, 434M).	$3.90 / 1M input	434M	—
HaMeR	Transformer-based 3D hand mesh recovery with ViTPose-H detection.	$3.90 / 1M input	—	—

Benchmarks

Published scores for ViTPose+ Large, from the ViTPose repository.

Evals

COCO AP78.6
COCO AR84.1

Reading	Example
Qualitative	Clear structure, grounded in the input
Quantitative	COCO AP: 78.6.
Cost and performance	Lower listed rate, mid-pack latency

Performance

Latency is the end-to-end round trip. Throughput is completion tokens divided by that time. It is a request rate, not decode speed.

Latency
11,823 ms
Throughput
—
Requests
272

Examples

2D pose estimation
ViTPose-Plus emits 17 keypoints per person, COCO schema.
{"poses":[{"keypoints":[[x,y,c], 17 of...]}], ...}

Supported inputs

textIgnored
A text part beside the media is accepted and ignored. A request with no media part is a 400.
image_urlYes
One per request. A second image_url part is a 400.
video_urlYes
One clip. Pose is tracked on every sampled frame; the top-level video_fps sets the detector cadence (default 10).
document_urlNo
Not accepted.

pose is the only method.
Each detected person returns a person label, a normalized bounding box, and 17 normalized COCO keypoints (kpts_xy).

Methods

Methods this model serves. Payload shapes are in the docs.

Method	Returns
pose (default)	One region per detected person, with `bbox_xywh`, `label`, and `kpts_xy`

Estimate cost

Estimate. An hour of video is 15 frames a minute at 256 tokens a frame, plus the audio in that request, and 500 output tokens a minute.

Hours of video

$1,347.84

Quick start

Call this model on the OpenAI-compatible gateway. The model id is already filled in.

from openai import OpenAI

client = OpenAI(
    base_url="https://gateway.vlm.run/v1/openai",
    api_key="<VLMRUN_API_KEY>",
)

response = client.chat.completions.create(
    model="usyd-community/vitpose-plus-large",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "image_url",
                    "image_url": {
                        "url": "http://images.cocodataset.org/val2017/000000000785.jpg"
                    },
                },
            ],
        }
    ],
    extra_body={"method": "pose"},
)

print(response.choices[0].message.content)

Response

A single image returns the json block alone, with no wrapper:

{
  "object": "img.pose.kpts",
  "items": [
    {
      "bbox_xywh": [0.4387, 0.1035, 0.3395, 0.8176],
      "label": "person",
      "kpts_xy": [[0.573, 0.1872], [0.584, 0.1706], [0.5639, 0.1738], "..."],
      "kpts_score": [0.9638, 0.9702, 0.9541, "..."]
    }
  ],
  "kpts_labels": ["nose", "left_eye", "right_eye", "..."]
}

kpts_xy lists the 17 COCO keypoints in order, and kpts_labels on the container names them once for every item: nose, the eyes, the ears, the shoulders, elbows, wrists, hips, knees and ankles. Each [x, y] pair is normalized against the frame, and a joint may fall outside 0-1 when the model places it past the edge. kpts_score is the model's raw per-joint confidence, in the same order.

On a video_url, video_fps at the top level of the request sets the detector cadence (default 10). frames lists every sampled frame, including frames with nobody in them, so a gap in items is distinguishable from a frame that was never read.