WilD3DGen · Complete 3D Benchmark Suite

TL;DR

The reported suite is the complete version; fast evaluation is only a debugging convenience and will not appear in the paper table.
Toys4K is the common image-to-3D anchor. Hi3DEval/Hi3DBench is a separate hierarchical quality track, not a replacement for exact geometry evaluation.
The final report keeps image quality, text alignment, geometry, material/part quality, and production health as separate scorecards.

1 · Suite topology

2 · Complete suite

Track	Full split	Purpose	Models
Toys4K image	3,218	Common single-view image-to-3D comparison	All seven
Hi3DBench image	510 image prompts	Object, part, material, and spatial quality	All seven
Hard-OOD geometry	100 fixed GT assets	Chamfer, F-score, normals, and novel-view reconstruction	All seven
T3Bench text	300 prompts	Single object, surroundings, and multiple objects	Native text models
Hi3DBench text	510 text prompts	Hierarchical text-to-3D quality	Native text models
Production health	Every artifact	Export validity, latency, VRAM, and failure rate	Every model/track

Track

Full split

Purpose

Models

Toys4K image

3,218

Common single-view image-to-3D comparison

All seven

Hi3DBench image

510 image prompts

Object, part, material, and spatial quality

All seven

Hard-OOD geometry

100 fixed GT assets

Chamfer, F-score, normals, and novel-view reconstruction

All seven

T3Bench text

300 prompts

Single object, surroundings, and multiple objects

Native text models

Hi3DBench text

510 text prompts

Hierarchical text-to-3D quality

Native text models

Production health

Every artifact

Export validity, latency, VRAM, and failure rate

Every model/track

3 · Model × modality matrix

Model	Image-to-3D	Native text-to-3D	Primary evaluation
EVA01	YES	YES	Image + text tracks
UniLat3D	YES	N/A	Image tracks
Pixal3D	YES	N/A	Single-view common track
TRELLIS-1	YES	YES	Image + text tracks
TRELLIS.2	YES	N/A	Image tracks
Hunyuan3D-2.1	YES	N/A	Image tracks
WilD3DGen	YES	According to released checkpoint	Target model

Model

Image-to-3D

Native text-to-3D

Primary evaluation

EVA01

YES

Image + text tracks

UniLat3D

YES

N/A

Image tracks

Pixal3D

YES

N/A

Single-view common track

TRELLIS-1

YES

Image + text tracks

TRELLIS.2

YES

N/A

Image tracks

Hunyuan3D-2.1

YES

N/A

Image tracks

WilD3DGen

YES

According to released checkpoint

Target model

N/A is an explicit result, not a zero.
Text is never converted to an image to manufacture a text-to-3D score.

4 · Metric coverage

Track	Implemented foundation	Adapter required
Toys4K image	PSNR/MAE, SSIM optional, mesh health, JSONL aggregation	CLIP, DINOv2 FD/KD, ULIP-2, Uni3D, LPIPS
Hi3DBench image/text	Artifact protocol and aggregation	Official Hi3DEval object/part/material evaluator
Hard-OOD geometry	Chamfer-L2, artifact loading	F-score, normal consistency, fixed GT sampling
T3Bench text	Artifact protocol and aggregation	Official T3Bench quality/alignment evaluator
Production health	Mesh existence, finite vertices, watertightness, components	Texture coverage, mesh size, peak VRAM, failure taxonomy

Track

Implemented foundation

Adapter required

Toys4K image

PSNR/MAE, SSIM optional, mesh health, JSONL aggregation

CLIP, DINOv2 FD/KD, ULIP-2, Uni3D, LPIPS

Hi3DBench image/text

Artifact protocol and aggregation

Official Hi3DEval object/part/material evaluator

Hard-OOD geometry

Chamfer-L2, artifact loading

F-score, normal consistency, fixed GT sampling

T3Bench text

Artifact protocol and aggregation

Official T3Bench quality/alignment evaluator

Production health

Mesh existence, finite vertices, watertightness, components

Texture coverage, mesh size, peak VRAM, failure taxonomy

5 · Full implementation gates

Phase	Deliverable	Exit criterion
1	Freeze manifests, prompt IDs, cameras, seeds, and GLB schema	Every model consumes identical manifests
2	Finish deterministic geometry/render/health metrics	Clean-environment run succeeds
3	Integrate CLIP, DINOv2, LPIPS, ULIP-2, and Uni3D	Checkpoint and preprocessing recorded
4	Integrate official Hi3DEval and T3Bench	Native image/text tracks score successfully
5	Run all seven models on complete manifests	JSONL, aggregates, failures, and render grids exist
6	Publish scorecards and qualitative galleries	No silent missing-data averaging

Phase

Deliverable

Exit criterion

Freeze manifests, prompt IDs, cameras, seeds, and GLB schema

Every model consumes identical manifests

Finish deterministic geometry/render/health metrics

Clean-environment run succeeds

Integrate CLIP, DINOv2, LPIPS, ULIP-2, and Uni3D

Checkpoint and preprocessing recorded

Integrate official Hi3DEval and T3Bench

Native image/text tracks score successfully

Run all seven models on complete manifests

JSONL, aggregates, failures, and render grids exist

Publish scorecards and qualitative galleries

No silent missing-data averaging

The current codebase is at the registry/protocol stage, not the complete-metric stage.
The full run should start only after the official adapters pass on a fixed calibration subset.

6 · Sources and decisions

Plan status: registry updated to complete_version_only; learned-metric and official-evaluator adapters remain implementation gates.