geopavlakos

HaMeR

About

HaMeR is an open-weight model from geopavlakos. Transformer-based 3D hand mesh recovery with ViTPose-H detection. It accepts text and images.

License
—
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

HaMeR is one of the catalog's pose models. The table is each model's published price, size, and context.

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

Benchmarks

Published scores for HaMeR, from the HaMeR paper.

Evals

FreiHAND F@5mm76.5
FreiHAND PA-MPJPE mm13

A lower score is better.

Reading	Example
Qualitative	Clear structure, grounded in the input
Quantitative	FreiHAND F@5mm: 76.5.
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
275 ms
Throughput
—
Requests
52

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_urlNo
Not accepted.
document_urlNo
Not accepted.

pose is the only method.
Each detected hand returns a left_hand or right_hand label, a detection score, a normalized bounding box, and 21 hand keypoints in OpenPose order, twice over: kpts_xy in the image plane and kpts_xyz in 3D. That is why the payload is tagged world.pose.kpts rather than img.pose.kpts, which is what the 2D-only ViTPose returns.

Method parameters

Parameter	Default	Description
`body_conf`	`0.5`	Minimum confidence for the body detection that locates the hands
`hand_conf`	`0.5`	Minimum confidence for a hand to be reported
`rescale_factor`	`2.0`	How far the detected hand box is expanded before the mesh is fit
`include_mesh`	`false`	Return the full MANO mesh alongside the keypoints

Methods

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

Method	Returns
pose (default)	One entry per detected hand, with `bbox_xywh`, `label`, `score`, `kpts_xy`, and `kpts_xyz`

Estimate cost

Input tokensOutput tokens

$3.90

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="geopavlakos/hamer",
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "image_url",
                    "image_url": {
                        "url": "http://images.cocodataset.org/val2017/000000000785.jpg"
                    },
                },
            ],
        }
    ],
    extra_body={"method": "pose", "method_params": {"hand_conf": 0.5}},
)

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

Response

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

{
  "object": "world.pose.kpts",
  "items": [
    {
      "bbox_xywh": [0.6735, 0.3556, 0.0623, 0.0938],
      "label": "left_hand",
      "score": 1.0,
      "kpts_xy": [[0.695, 0.3822], [0.6881, 0.3928], "..."],
      "kpts_xyz": [[-0.0957, 0.0064, 0.0062], [-0.1146, 0.0257, -0.0211], "..."]
    }
  ],
  "kpts_labels": ["wrist", "thumb_mcp", "thumb_pip", "thumb_dip", "thumb_tip", "..."]
}

label is the hand's laterality, left_hand or right_hand, and score is the detection confidence.

kpts_xy is normalized against the source image, so multiply by image_width and image_height for pixels. kpts_xyz is metric 3D in the model's own hand-centred frame and is not tied to image coordinates; it is the reason the payload is tagged world.pose.kpts rather than the img.pose.kpts that ViTPose returns. Both lists hold the same 21 joints in OpenPose hand order, named once on the container by kpts_labels: the wrist, then the thumb, index, middle, ring and pinky fingers, four joints each from base to tip (_mcp, _pip, _dip, _tip).

method_params.include_mesh adds the full MANO mesh beside the keypoints.