LiTo Runtime FP32 for `mlx-spatial`

Run Apple's LiTo image-to-3D Gaussian Splat model on Apple Silicon through mlx-spatial, using MLX-ready safetensors instead of local .ckpt conversion.

This runtime-pruned bundle is for researchers who want a practical Mac-native LiTo inference path: download one repository and generate a 3D Gaussian Splat PLY from an input image. No CUDA is required.

Quick Start: Image to 3DGS on Apple Silicon

Install mlx-spatial:

pip install \
  "mlx-spatial @ git+https://github.com/appautomaton/mlx-spatial.git@afa6b6512567dc9964294a1cec8601e1b505802e"

This model card requires the bundle-local LiTo dependency layout introduced by the immutable runtime commit above.

Download this model bundle:

hf download appautomaton/lito-research-mlx \
  --local-dir weights/lito-research-mlx

Validate the local layout:

mlx-spatial-lito validate weights/lito-research-mlx
mlx-spatial-lito inspect weights/lito-research-mlx --limit 10

Generate a Gaussian-splat PLY:

mlx-spatial-lito generate inputs/lito/sample.png \
  --weights-root weights/lito-research-mlx \
  --output outputs/lito/sample.ply \
  --memory-profile balanced \
  --print-metrics

The output is a 3D Gaussian Splat PLY, not a mesh. Use a 3DGS-aware viewer such as KIRI Engine's 3DGS Render Blender add-on. Blender's native PLY importer can read the container but does not render LiTo Gaussian splat fields correctly.

What This Model Bundle Provides

This Hugging Face repository is a self-contained LiTo runtime bundle for mlx-spatial:

tokenizer/lito_new.safetensors
image_to_3d/lito_dit_rgba.safetensors
dependencies/trellis/LICENSE
dependencies/trellis/SOURCE.json
dependencies/trellis/ckpts/ss_dec_conv3d_16l8_fp16.json
dependencies/trellis/ckpts/ss_dec_conv3d_16l8_fp16.safetensors

The embedded decoder is the exact 147,591,972-byte checkpoint from microsoft/TRELLIS-image-large revision 25e0d31ffbebe4b5a97464dd851910efc3002d96, with SHA-256 1c76d4a40519aa2d711cc263a8404105231ac26db31d946bed48b84fee79009a. It converts LiTo voxel latents into Gaussian initialization coordinates. mlx-spatial-lito validate requires this bundle-local decoder and does not search a separate TRELLIS checkout.

File	Logical tensors	Bytes
`image_to_3d/lito_dit_rgba.safetensors`	1,016	3,713,563,945
`tokenizer/lito_new.safetensors`	467	521,201,761
`dependencies/trellis/ckpts/ss_dec_conv3d_16l8_fp16.safetensors`	74	147,591,972
Runtime weights	1,557	4,382,357,678

Best For

Apple Silicon MLX inference experiments.
Image-to-3D Gaussian Splat generation with mlx-spatial.
Research workflows that need LiTo weights in safetensors format.
Local 3DGS inspection in KIRI, Gaussian-splat-aware Blender add-ons, or compatible 3DGS viewers.

Current Limitations

Research-only, non-commercial license boundary from Apple.
This is an unofficial converted derivative bundle, not an Apple-hosted official MLX package.
Current mlx-spatial LiTo support targets image-to-3D Gaussian Splat inference; it does not provide LiTo training, fine-tuning, mesh extraction, multi-image conditioning, or video conditioning.
Visual quality depends strongly on input matting and alpha quality. Inputs with broad or noisy alpha masks can produce weaker holes, handles, and fine structures.
CUDA is not required and is not used by mlx-spatial LiTo inference.

Conversion Details

The LiTo files were converted from Apple's original .ckpt checkpoints to safetensors and pruned to the modules read by the main inference path. The bundle retains the EMA velocity estimator, DINO/RGBA image conditioner, Gaussian decoder, and voxel decoder. It removes the non-EMA training copy, duplicate tokenizer, mesh/fpoint/LPIPS modules, and training-only decoders. Retained tensor values, shapes, and dtypes are unchanged; no retained LiTo tensor is quantized. The embedded TRELLIS decoder is redistributed unchanged from the immutable Microsoft source revision recorded above.

Verification

Both checkpoints and the embedded decoder pass mlx-spatial-lito validate.
Architecture inspection recovers 28 DiT blocks, 6 Gaussian Perceiver blocks, and 4 voxel decoder blocks.
The bundle exposes 1,483 LiTo logical inference tensors.
A bundle-local 20-step runtime validation produced 8,192 finite Gaussians after explicitly capping occupied cells for packaging verification. This is not an uncapped quality benchmark.

Project Links

Runtime package: mlx-spatial
mlx-spatial PyPI package: https://pypi.org/project/mlx-spatial/
mlx-spatial source: https://github.com/appautomaton/mlx-spatial
This model repo: https://huggingface.co/appautomaton/lito-research-mlx

Apple LiTo Source and License

This bundle is based on Apple's LiTo research release:

Apple LiTo project: https://apple.github.io/ml-lito/
Apple LiTo source code: https://github.com/apple/ml-lito
Apple model license: https://github.com/apple/ml-lito/blob/main/LICENSE_MODEL

Apple's LiTo model weights are released under the Apple Machine Learning Research Model License Agreement. Use is limited to non-commercial scientific research and academic development activities. Commercial product use is not permitted.

License and source access last checked: 2026-08-06.

This repository is not an Apple release and is not endorsed by Apple. Redistribution of this converted bundle must keep Apple's license terms, attribution notice, and modification disclosure.

Required attribution notice:

Apple Machine Learning Research Model is licensed under the Apple Machine Learning Research Model License Agreement.

The embedded TRELLIS dependency is licensed under the MIT License. Its license copy and immutable source metadata are included under dependencies/trellis/.

Downloads last month: -; Downloads are not tracked for this model. How to track

MLX

Hardware compatibility

Quantized

Inference Providers NEW

Image-to-3D

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for appautomaton/lito-research-mlx

Finetunes

1 model

LiTo Runtime FP32 for mlx-spatial