appautomaton/lito-research-mlx

LiTo Runtime FP32 for mlx-spatial

A compact FP32 inference bundle for running Apple's LiTo image-to-3D Gaussian Splat model with mlx-spatial on Apple Silicon. It contains only the modules used by the main inference path.

This is an unofficial derivative for non-commercial research use. It is not an Apple release.

Variants

Variant	Precision	Size	Model
FP32 runtime	Source FP32	4.235 GB	This model
8-bit runtime	Selective affine 8-bit	2.173 GB	lito-research-mlx-8bit

Both variants expose the same 1,483 logical inference tensors.

Use

pip install mlx-spatial

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

LiTo also requires two sparse-structure decoder files:

hf download microsoft/TRELLIS-image-large \
  ckpts/ss_dec_conv3d_16l8_fp16.json \
  ckpts/ss_dec_conv3d_16l8_fp16.safetensors \
  --local-dir weights/trellis2/microsoft/TRELLIS-image-large

Validate and run:

mlx-spatial-lito validate weights/lito-research-mlx

mlx-spatial-lito generate inputs/lito/object-rgba.png \
  --weights-root weights/lito-research-mlx \
  --output outputs/lito/object.ply \
  --format ply \
  --num-steps 20 \
  --cfg-scale 3.0 \
  --print-metrics

A clean RGBA foreground matte is strongly recommended. The output is a 3D Gaussian Splat PLY, not a triangle mesh.

Bundle

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
Total	1,483	4,234,765,706

The bundle retains:

• the EMA velocity estimator;
• the DINO/RGBA image conditioner;
• the Gaussian decoder;
• the voxel decoder.

It removes the non-EMA training copy, duplicate embedded tokenizer, mesh/fpoint/LPIPS modules, and training-only decoders: 2,418 tensors and 4,287,076,368 bytes in total. Retained tensor values, shapes, and dtypes are unchanged. No retained tensor is quantized.

The pruning record is embedded under mlx_spatial.lito.runtime in each safetensors header.

Verification

• Both checkpoints pass mlx-spatial-lito validate.
• Architecture inspection recovers 28 DiT blocks, 6 Gaussian Perceiver blocks, and 4 voxel decoder blocks.
• The implementation at 2a4fac6 passed 1143 repository tests.

These checks cover the bundle format and MLX runtime. They do not claim bit-exact parity with the upstream Torch/CUDA implementation.

Limitations and License

• Inference only; training and mesh-specific modules are intentionally absent.
• Single-view reconstruction cannot determine unseen geometry with certainty.
• Fine detail depends on the input view, alpha matte, occlusion, reflections, and thin structures.
• Commercial use is not permitted by Apple's model license.

Read the bundled LICENSE_MODEL and Apple's model license before use.

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

Links

• appautomaton/mlx-spatial — MLX-native 3D and spatial inference for Apple Silicon: object reconstruction, image-to-mesh, scene geometry, Gaussian splats, and multi-view bundles.
• mlx-spatial documentation
• mlx-spatial on PyPI
• LiTo guide
• 8-bit runtime variant
• Apple LiTo project
• Apple LiTo source
• MLX