appautomaton/lito-research-mlx-8bit

LiTo Runtime 8-bit Affine for mlx-spatial

A compact LiTo inference bundle with selective 8-bit affine weights, built for direct use by mlx-spatial on Apple Silicon. Accuracy-sensitive boundaries remain in FP32.

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	lito-research-mlx
8-bit runtime	Selective affine 8-bit	2.173 GB	This model

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

Compatibility

Quantized loading requires mlx-spatial commit 2a4fac6 or a later release containing it:

pip install \
  "mlx-spatial @ git+https://github.com/appautomaton/mlx-spatial.git@2a4fac6"

The runtime targets Apple Silicon and MLX 0.32.x. It does not use Torch, CUDA, or a dequantized FP32 checkpoint.

Use

hf download appautomaton/lito-research-mlx-8bit \
  --local-dir weights/lito-research-mlx-8bit

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-8bit

mlx-spatial-lito generate inputs/lito/object-rgba.png \
  --weights-root weights/lito-research-mlx-8bit \
  --output outputs/lito/object-8bit.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	Quantized matrices	Bytes
image_to_3d/lito_dit_rgba.safetensors	1,016	224	2,004,347,727
tokenizer/lito_new.safetensors	467	124	168,537,103
Total	1,483	348	2,172,884,830

The quantization scheme is affine 8-bit with group size 64. Packed weights are stored as uint32 with FP32 scales and biases, then executed directly through mx.quantized_matmul.

Internal attention and MLP matrices in the EMA DiT, Gaussian decoder, and voxel decoder are quantized. The image conditioner, convolutions, embeddings, normalization parameters, timestep and condition projections, boundary projections, and final output heads remain FP32.

The bundle is also runtime-pruned: non-EMA, duplicate, mesh, loss, and training-only modules are absent. Format details are embedded under mlx_spatial.lito.runtime and mlx_spatial.lito.quantization.

Verification

• Both checkpoints pass mlx-spatial-lito validate.
• Architecture inspection recovers 28 DiT blocks, 6 Gaussian Perceiver blocks, and 4 voxel decoder blocks.
• Real-weight Linear probes measured 0.51%–0.61% relative RMSE and cosine similarity above 0.99998 against FP32.
• An uncapped 20-step run produced 557,568 finite Gaussians in 2 minutes 39.76 seconds, with 11.60 GiB peak active MLX memory.
• The implementation passed 1143 repository tests.

The runtime figures are one local Apple Silicon observation, not a general benchmark or an official quality evaluation.

Limitations and License

• Inference only; training and mesh-specific modules are intentionally absent.
• Quantization can change generation details relative to FP32.
• 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
• FP32 runtime variant
• Apple LiTo project
• Apple LiTo source
• MLX