hunyuan3d-swift

swift and python mlx ports of the hunyuan3d shape and paint pipelines.

the repo has two parts:

• swift package at the root
• python-mlx ports under python/

the swift code is checked against python fixtures. the python ports are checked against the original pytorch code.

this is exciting

that's hunyuan3D-shape running in FIVE POINT SIX gigabytes of RAM (thanks to MLX). with a Q8 or Q4, we are easily in mobile territory. now, what would you do with a image to 3d model running on your phone (iPhone 15 onwards)? not sure, but that's really really cool, the first time it has been possible (AFAIK). here's some demos (checkout Modelr:

https://github.com/user-attachments/assets/495de7e8-6c76-4b3f-af37-ffa23e0b0a64

https://github.com/user-attachments/assets/7c1f5008-8a68-4e08-b3f6-80864d3c1a00

package

main products:

• Hy3DMLX for shape generation
• HunyuanPaintMLX for texture generation
• hy3d for the command line

main commands:

swift build -c release
swift run -c release hy3d --help

weights

download model weights into local folders:

hf download zimengxiong/hunyuan3d-mlx-shape-small --local-dir weights/shape-small
hf download zimengxiong/hunyuan3d-mlx-paint-large --local-dir weights/paint-large

four supported slots:

• shape small: hunyuan3d-dit-v2-mini
• shape large: hunyuan3d-dit-v2-0-turbo
• paint small: hunyuan3d-paint-v2-0
• paint large: hunyuan3d-paintpbr-v2-1

run

shape plus paint:

swift run -c release hy3d generate photo.png -o out.glb \
  --shape-weights weights/shape-small \
  --paint-weights weights/paint-large

shape only:

swift run -c release hy3d shape photo.png -o mesh.glb \
  --weights weights/shape-small

paint an existing mesh:

swift run -c release hy3d paint mesh.glb photo.png -o textured.glb \
  --weights weights/paint-large --model pbr

license

swift source is mit. dependencies, model weights, and algorithm ports keep their own licenses. see THIRD_PARTY_LICENSES.md.