EditHero
A Benchmark for Long-Horizon
Part-Level 3D Editing and Vibe Modeling

Ruihan Yu1,2*Yu-Ju Tsai3,4*Muyao Niu1,2Runyi Li1,2Lian Fu1,2Hanqing Liu2Zheng-Hui Huang1Yonghao Yu1Sho Kuno1,2Ming-Hsuan Yang4†Kaipeng Zhang1†Zhixiang Wang1†

1 Alaya Lab2 The University of Tokyo3 Institute of Science Tokyo4 University of California, Merced

* Equal contribution† Corresponding authors

Paper Code Hugging Face Video BibTeX

Who edits 3D better? EditHero puts conventional methods and LLMs that write code on the same long edit chains, with the right answer after every edit.

Top downConventional methods regenerate the whole object.
Bottom upCode agents change only the named part.
A figurine edited over five turns: sunglasses added, cap removed, hair recoloured, outfit replaced and a witch hat addedt0t1t2t3t4t5AddRemoveReplaceRetexture
Leaderboard on the same 55 chains: instruction following and how much of the original object is kept, for six LLMs writing code (Claude Opus 5.5, OpenAI Astra, Claude Fable 5.1, OpenAI Sol 6, DeepSeek V4.1 Flash, Z.ai GLM 5.3 Flash) and four 3D editing methods (VoxHammer, Nano3D, 3DEditFormer, PartFlow)
TRAILER

Editing is a sequence

TRAILER · 120 SECONDS

Edit chains with an exact target after every turn, and what happens when editors keep editing their own output.

Benchmark

Every turn has an exact target

Library parts in, a JSON log out. The log rebuilds every state exactly. Every turn checked by a person.

457edit chains
2,755edits
252host objects
30turns at most

Four kinds of edit

add
before→after: Add half a boiled egg on top of the uppermost patty.
before→after: Add a large orange pumpkin with a green stem on the mossy ground beside the shrine's arched entrance.
before→after: Add a small hourglass with a brown wooden frame and a purple tint to the open area of the shelf top.
remove
before→after: Remove the tree farthest from the front door among the trees beside the cabin.
before→after: Remove the yellow omelet nigiri wrapped in a strip of dark seaweed from the serving tray.
before→after: Remove the knight's sword.
replace
before→after: Replace the cream crackle-glaze soup bowl and its ladle with a plain white footed bowl, standing upright in the same place beside the sushi tray.
before→after: Replace the large smoky glass jar topped by a copper funnel with a tall white ceramic jar.
before→after: Replace the knight's sword with a steel longsword with a gilded crossguard and a leather-wrapped grip.
retexture
before→after: Change only the rectangular sushi serving tray to ivory ceramic with flowing dark-blue wave patterns. Preserve every sushi piece, its original food colors and texture details, and all geometry.
before→after: Give the feed roll brushed brass with copper scratches and dark circumferential bands, preserving the original support.
before→after: Change the material of the arched roof of the little yellow cottage to terracotta red clay roof tiles.

Six kinds of object

Characters48 objects
Animals & creatures30 objects
Robots41 objects
Vehicles39 objects
Objects61 objects
Scenes & buildings33 objects

How a chain is made

The data engine: a host with named slots and a part library, one turn of the assembly engine, and a chain rebuilt from its logged operations
Chains

Follow a chain, turn by turn

Each method keeps editing its own output.

operationobjectattribute

Same camera and light for every tile.

Leaderboard

Conventional methods vs LLMs writing code

Conventional methodsLLMs writing code

Follows the instruction

IF, share of the asked change made

  1. 1Opus 5.50.62
  2. 2Astra0.55
  3. 3Fable 5.10.54
  4. 4Sol 60.43
  5. 5DeepSeek V4.1 Flash0.42
  6. 6VoxHammer0.37
  7. 7Nano3D0.34
  8. 8GLM 5.3 Flash0.26
  9. 93DEditFormer0.21
  10. 10PartFlow0.13

Keeps the original

CC0, untouched part of the first object kept

  1. 1Astra0.95
  2. 2Opus 5.50.95
  3. 3Fable 5.10.94
  4. 4Sol 60.93
  5. 5DeepSeek V4.1 Flash0.90
  6. 6GLM 5.3 Flash0.89
  7. 7Nano3D0.73
  8. 8PartFlow0.66
  9. 93DEditFormer0.54
  10. 10VoxHammer0.48

55 chains, 416 edits. Every method keeps editing its own output.

BibTeX

@misc{yu2026editherobenchmarklonghorizonpartlevel,
      title={EditHero: A Benchmark for Long-Horizon Part-Level 3D Editing and Vibe Modeling},
      author={Ruihan Yu and Yu-Ju Tsai and Muyao Niu and Runyi Li and Lian Fu and Hanqing Liu and Zheng-Hui Huang and Yonghao Yu and Sho Kuno and Ming-Hsuan Yang and Kaipeng Zhang and Zhixiang Wang},
      year={2026},
      eprint={2610.02298},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.02298},
}