Arxiv Preprint 2026

Alchemy3D

Scalable 3D foundation model for versatile Assets Editing

Abstract

Alchemy3D editing types: addition, removal, replacement, local appearance, global appearance, and animation.

Dataset

Alchemy3D-1M

A million-scale editing corpus with 1.25M unique assets and 1.38M pairs spanning seven edit types, using PBR-ready assets. Type-specific pipelines cover addition, removal, replacement, local and global appearance, and animation. A segmentation split is held out of training (825K pairs after balancing).

  • 7 edit types
  • 1.25M unique assets
  • 1.38M editing pairs
  • PBR ready assets

Method

Alchemy3D model

The Alchemy3D model is a strong baseline for versatile 3D asset editing. Trained on the constructed Alchemy3D-1M dataset, it concatenates encoded source tokens directly with noisy tokens, enabling in-depth interaction within the transformer block.

Alchemy3D architecture: source and target encoders, optional image or text instruction, cross-attention transformer, and decoder.

Benchmark

GEdit3D-Bench

GEdit3D-Bench is a large-scale, multi-dimensional benchmark built from assets independent of our training data. It combines newly synthesized 3D content with recently released Sketchfab assets; all samples are curated through automated and human filtering, and evaluation jointly measures view quality, reference alignment, and MLLM-based scores rather than similarity to a single generated target.

GEdit3D-Bench construction: asset collection, captioning, instruction and image generation, human-in-the-loop curation, and evaluation metrics.

Results

Alchemy3D

Side-by-side Source / Edited wipes from our base model, Alchemy3D. Click Download to view on a card to fetch that pair (with progress). Then drag to orbit and wipe Source / Edited. Use Normals for real-time normal shading.

Alchemy3D-Segment

Comparison

Quantitative results

Add / remove / replace on GEdit3D-Bench. Alchemy3D reaches 89.6% success and 70.7% human preference; prior methods sit near 5–7% preference.

Table 2. GEdit3D-Bench, add / remove / replace. Higher is better. Best in bold, second-best underlined.
Method View quality Reference alignment MLLM Human
Pref.
Aes. MUSIQ ACLIP I ACLIP T ADINO I SR VQ IF IP
Nano3D 4.40 72.87 81.59 65.10 69.15 51.14% 70.16 53.84 72.09 5.1%
3DEditFormer 4.34 70.87 81.83 63.19 68.07 58.90% 66.31 59.79 63.39 5.3%
PartFlow 4.31 70.96 81.25 62.24 67.78 55.50% 65.27 56.20 66.60 6.5%
Alchemy3D-Turbo 4.58 72.94 82.36 68.31 72.88 88.18% 69.91 70.71 73.22 —
Alchemy3D 4.43 72.36 84.91 67.47 73.18 89.55% 72.74 71.38 76.19 70.7%

Citation

BibTeX