Rapidata/image-to-3d-single-object

Rapidata Image-to-3D Benchmark

Built by Rapidata.

We gave 14 image-to-3D models the same single image and asked people which reconstruction is better. Every model turned each image into a textured 3D asset, we rendered all assets the same way, and annotators compared two of them at a time next to the input image. This dataset holds 57,777 of those comparisons and all 307,965 votes behind them, collected with the Rapidata Python SDK. There are no automated metrics in it.

Each comparison was judged on one of two questions:

• Geometry: is the shape right? Assets are shown in plain grey, with their textures removed.
• Appearance: does it look like the object in the image? Assets are shown as the model delivered them.

The live leaderboard, with filters by model and image category, is on benchmark.ai/image-to-3d.

If this dataset is useful to you, a like helps us decide what to release next ❤️

To evaluate your own models and create a leaderboard, check out our MRI.

At a glance

Comparisons (rows)	57,777
Human votes	307,965, at least 5 per comparison
Models	14
Input images	357, each rendered twice (grey and textured), so 714 prompts
Turntable videos	10 s, 240 frames at 24 fps, 720×720
Annotators	a dedicated audience that passed a 3D screening test (see below)

What a comparison looks like

Annotators see the input image and two turntables side by side, without model names and in random order. The two examples below are real comparisons from this dataset.

Geometry: "Which video reflects the shape of the object in the photo?"

Input image	Hi3D v3.0 (chosen 5 of 5)	Tripo H3.1 (0 of 5)

Appearance: "Which video looks more like the object in the photo?"

Input image	Hi3D v3.0 (chosen 3 of 5)	Tripo H3.1 (2 of 5)

The word "photo" in both questions means the input image, which is either a photograph or a generated picture.

Rankings

Scores come from a Bradley–Terry fit on all votes, put on the Elo scale. The rank is in brackets. These numbers update with every sync of this dataset; the live version is on benchmark.ai.

Model	Overall	Geometry	Appearance
Hi3D v3.0	1375 (#1)	1325 (#1)	1437 (#1)
Meshy v7.1	1304 (#2)	1269 (#2)	1349 (#2)
Meshy v7	1292 (#3)	1257 (#3)	1337 (#3)
Tripo H3.1	1271 (#4)	1246 (#4)	1304 (#4)
Hunyuan3D 3.1 Pro	1155 (#5)	1126 (#5)	1190 (#5)
Seed3D 2.0	1105 (#6)	1047 (#8)	1170 (#6)
TRELLIS.2	1012 (#7)	1086 (#7)	936 (#8)
Rodin Gen-2.5 High	953 (#8)	1119 (#6)	776 (#10)
Tripo P2	908 (#9)	750 (#11)	1066 (#7)
Rodin Gen-2.5 Low	790 (#10)	869 (#9)	702 (#12)
Pixal3D	767 (#11)	731 (#12)	800 (#9)
Hunyuan3D 2.1	763 (#12)	769 (#10)	752 (#11)
TRELLIS	667 (#13)	705 (#13)	620 (#13)
SAM 3D Objects	638 (#14)	702 (#14)	560 (#14)

Some models move a lot between the two columns. Rodin Gen-2.5 High is 6th on Geometry and 10th on Appearance: good shapes, weaker textures. Tripo P2 goes the other way.

The input images

357 images from four sources, picked so that no evaluated model is likely to have trained on them:

Source	Images	License
Amazon Berkeley Objects catalogue photos, only products that ship no 3D model	103	CC BY 4.0
3DGen-Bench generated images	90	MIT
FLUX-Reason-6M, generated with FLUX.1 [dev]	26	Apache 2.0
Generated for this benchmark with GPT Image, with made-up brands wherever text or a logo appears	138	released with this dataset

261 images show a single object and 96 show several. Some were chosen because they are hard to rebuild from one view: thin structures, holes and cavities, one object hiding another, glass or shiny surfaces, fur, hair or leaves, and printed text.

Every image was flattened onto white, padded to a square and scaled to at most 1024 px.

How the assets were made and rendered

• Each model got the same single image and ran with fixed settings, including whatever preprocessing its provider applies, such as background removal.
• SAM 3D Objects needs to be told what to segment, so it also got a short noun naming the object. It sometimes segments the wrong thing, which hurts its scores.
• One asset per model per image. We never picked the best of several tries.
• Every mesh was centred and scaled to the same size, then rendered with the same camera path, lens, lighting and background. The turntable starts from the input image's viewpoint and then turns to show the sides the image could not.
• The Geometry video is the same file rendered with all textures removed, not a second generation.

Who voted

The comparisons were judged by a dedicated audience, not Rapidata's general pool. To get in, an annotator answered gold-standard 3D comparisons (44 pairs with a known right answer, on both grey and textured renders) and needed at least 75% correct over ten or more of them. Each comparison was then shown to at least five different annotators.

Columns

One row is one comparison: two models, one input image, one question.

Column	Type	Meaning
input_image	string	URL of the image both models received
video1, video2	string	URLs of the two turntable videos
model1, model2	string	the model behind video1 and video2
weighted_results_video1_geometry, weighted_results_video2_geometry	float32	votes for each side on Geometry, each vote weighted by the annotator's reliability score
detailed_results_geometry	string	every Geometry vote as a JSON array
weighted_results_video1_appearance, weighted_results_video2_appearance	float32	the same for Appearance
detailed_results_appearance	string	every Appearance vote as a JSON array

A row belongs to one question only. On a Geometry row the Appearance columns are empty, and the other way round.

The two weighted_results_* values of a row are weighted vote sums, not probabilities, so they don't add up to 1. Divide one by their sum to get a win share.

Each entry in detailed_results_* has the annotator's country, language, gender, ageBucket and occupation, their reliability userScore, the side they picked (votedFor: A is video 1, B is video 2) and the time of the vote (votedAt).

import json
from datasets import load_dataset

ds = load_dataset("Rapidata/image-to-3d-single-object", split="train")

geometry = ds.filter(lambda row: row["detailed_results_geometry"] is not None)
row = geometry[0]
votes = json.loads(row["detailed_results_geometry"])
print(row["model1"], "vs", row["model2"], [v["votedFor"] for v in votes])

About Rapidata

Rapidata collects human feedback at scale. Learn more at rapidata.ai.