appautomaton/lito-research-mlx-8bit

LiTo Runtime 8-bit Affine for mlx-spatial

A self-contained LiTo inference bundle with selective affine INT8 weights for Apple Silicon. Accuracy-sensitive boundaries remain in FP32. This is an unofficial derivative for non-commercial research use, not an Apple release.

Use

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

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

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

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

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
dependencies/trellis/ckpts/ss_dec_conv3d_16l8_fp16.safetensors	74	0	147,591,972
Runtime weights	1,557	348	2,320,476,802

The bundle also includes the decoder config, Microsoft MIT license, and an immutable source manifest under dependencies/trellis/.

The LiTo quantization scheme is affine 8-bit with group size 64. Packed weights are stored as uint32 with FP32 scales and biases and execute 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, normalizations, boundary projections, and output heads remain FP32.

Embedded TRELLIS Decoder

LiTo uses the TRELLIS sparse-structure decoder to convert voxel latents into Gaussian initialization coordinates. This bundle embeds the exact checkpoint from microsoft/TRELLIS-image-large revision 25e0d31ffbebe4b5a97464dd851910efc3002d96:

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 decoder safetensors SHA-256 is 1c76d4a40519aa2d711cc263a8404105231ac26db31d946bed48b84fee79009a. The runtime does not search a separate TRELLIS checkout.

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.
• A separate bundle-local 20-step validation produced 8,192 finite Gaussians after explicitly capping occupied cells for packaging verification.

These figures are one local Apple Silicon observation, not a general benchmark or an official quality-equivalence claim.

Limitations and Licenses

• 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.
• Commercial use is not permitted by Apple's model license.

LiTo weights are covered by the bundled LICENSE_MODEL. The embedded TRELLIS decoder is covered by the MIT License under dependencies/trellis/LICENSE.

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

Links

• FP32 runtime variant
• appautomaton/mlx-spatial
• Apple LiTo project
• Apple LiTo source