codepawl/asset-index

Asset Index

37,668 3D assets from 6 libraries in one table. An AI agent, a model or an animator can ask for "a wooden crate, 0.5 m, game tier" and get the right file. Every row has the same columns. They give the real size in metres, the kind, style, surface, biome and PBR maps, and triangles where known. They also say whether the asset tiles or is rigged, what it suits, which optimizer tier to use and how to get it.

Library	Rows	Licence
Quixel Megascans	22,696	Fab Standard Licence: claim it on Fab, then optimize locally (no files here)
Kenney	4,434	CC0
Smithsonian Open Access	3,451	CC0, files in codepawl/smithsonian-3d-glb
ambientCG	2,917	CC0
Poly Haven	2,387	CC0 (credit Poly Haven when you use their API)
Quaternius	1,783	CC0 for 82 packs, Quaternius Asset Licence (no redistribution) for one

CC0 rows point straight at a file you can download and ship. Megascans rows point at the Fab page. Fab licenses Megascans to each user and forbids passing the files on. This repository holds no Megascans files, only Fab's public listing details. The tool/ folder turns your own download into web and game GLB on your machine.

Ask it like an agent

from datasets import load_dataset
rows = load_dataset("codepawl/asset-index", split="train")
crates = rows.filter(lambda row: "crate" in row["search_text"] and row["scale_m"] and 0.25 <= row["scale_m"] <= 1)
for row in crates.select(range(5)):
    print(row["title"], row["source"], row["scale_m"], row["get"])

Or in plain words:

python tool/query.py metadata.parquet "wooden crate, 0.5 m, game tier"
python tool/query.py metadata.parquet "rock cliff, wet, 2 m, film" --top 5
python tool/query.py metadata.parquet "tileable forest floor material, web" --free
python tool/query.py metadata.parquet "lowpoly character" --rigged --json

A size such as 2 m keeps assets from half to double that size. game, web, mobile or film picks the use whose tier is shown. material, model, decal, hdri or character narrows the kind, lowpoly or photoreal the style, tileable keeps repeating surfaces. --free keeps files you can ship, --trainable drops rows Fab marks NoAI.

Columns

Column	Meaning
asset_id, source, title, source_url, collection, authors	where it comes from (collection is the Kenney kit or Quaternius pack)
licence, redistributable, credit	SPDX id, whether the files may be shipped, the credit line to show
kind, style, category_path, tags	model, material, decal, atlas, hdri, brush, terrain; photoreal, stylized, lowpoly
surface_type, biomes	rock, wood, leather…; forest, desert, urban… (one rule set for every library)
size_m, scan_area_m, scale_m, size_class, size_from	real size in metres and where it came from: the listing, the library's API or our own measurement
triangles, lods, texel_density_px_m, pbr_maps, max_texture_px	technical detail where the library gives it or we measured it
tileable, atlas, rigged, animated	repeating surface, sheet of cut-outs, has a skeleton, ships clips
use_for, animation_role, recommended_tier	where it fits and which optimizer tier to start from
get, files	how to get it now; files lists our optimized GLB tiers once a library is packed
no_ai, ai_generated	Fab's NoAI and CreatedWithAI flags (Megascans only), with read_on
search_text	one lowercase line to match a request against

Known sizes: ambientCG 572, Quixel Megascans 22,223, Poly Haven 1,390.

For animators and set dressers

animation_role sorts every library by how it is used on a shot. The roles are:

• character
• terrain or set surface
• ground cover or foliage
• hand prop or close-up detail
• set dressing
• background or layout piece
• lighting and background (HDRIs)

scale_m is the real size, so a prop drops into a scene at the right size for the camera. 192 rows are rigged.

Optimize what you download

The tool is in tool/. It needs Node 20+, Blender for FBX, KTX-Software for KTX2 and uv for Python; tool/README.md lists the setup.

cd tool && npm install
node megascans/optimize.mjs "D:/Downloads/assets" --out "D:/Downloads/assets-optimized" --no-check

Each folder or zip can hold glTF, FBX with loose maps, or a texture set. The tool writes objects/ (meshopt + WebP), objects_ktx2/, lods/ (50/20/6 %), mobile/, engine/ and collision/. Without --no-check it then loads every file in three.js WebGL, three.js WebGPU, 67js, Blender and Godot. Texture sets become a plane of the scanned size. It refuses to upload Megascans output. So far it is tested on CC0 assets laid out like Megascans exports (glTF folder, FBX with loose maps, zipped surface set).

For web games and WebGPU

Use the web tier for three.js and WebGPU scenes and ktx2 when GPU memory matters.

Use it in 67js

npm install 67js@npm:@codepawlhq/67js meshoptimizer

NoAI and training

Of the 22,656 Megascans listings read in detail by 2026-11-07, Fab marked 147 NoAI. On 2026-11-05 Quixel announced that you can now train AI models on Megascans. Quixel also said Fab listings update until 2026-11-12. The index reads the flag again after that date. CC0 rows carry no such restriction.

Licence

The index (rows, derived fields, code) is CodePawl's. Each asset keeps its own licence, given per row. Names and listing details belong to their libraries. Not affiliated with or endorsed by Epic Games, Quixel, Poly Haven, ambientCG, Kenney, Quaternius or the Smithsonian Institution.