Make LiTo runtime bundles self-contained

README.md

@@ -103,10 +103,6 @@ uv run hf download appautomaton/sam-3d-objects-mlx --local-dir weights/sam-3d-ob
 uv run mlx-spatial-sam3d validate weights/sam-3d-objects-mlx
 
 uv run hf download appautomaton/lito-research-mlx --local-dir weights/lito-research-mlx
-uv run 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
 uv run mlx-spatial-lito validate weights/lito-research-mlx
 ```
 

docs/lito.md

@@ -21,10 +21,6 @@ Recommended runtime bundle:
 ```bash
 uv run hf download appautomaton/lito-research-mlx \
   --local-dir weights/lito-research-mlx
-uv run 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
 uv run mlx-spatial-lito validate weights/lito-research-mlx
 uv run mlx-spatial-lito inspect weights/lito-research-mlx --limit 10
 ```
@@ -34,16 +30,19 @@ Expected converted layout:
 ```text
 weights/lito-research-mlx/tokenizer/lito_new.safetensors
 weights/lito-research-mlx/image_to_3d/lito_dit_rgba.safetensors
-weights/trellis2/microsoft/TRELLIS-image-large/ckpts/ss_dec_conv3d_16l8_fp16.json
-weights/trellis2/microsoft/TRELLIS-image-large/ckpts/ss_dec_conv3d_16l8_fp16.safetensors
+weights/lito-research-mlx/dependencies/trellis/LICENSE
+weights/lito-research-mlx/dependencies/trellis/SOURCE.json
+weights/lito-research-mlx/dependencies/trellis/ckpts/ss_dec_conv3d_16l8_fp16.json
+weights/lito-research-mlx/dependencies/trellis/ckpts/ss_dec_conv3d_16l8_fp16.safetensors
 ```
 
-The first two files are the LiTo bundle. The final two are the sparse-structure
-decoder used to convert LiTo voxel latents into Gaussian initialization
-coordinates. They come from `microsoft/TRELLIS-image-large`, not from the
-`microsoft/TRELLIS.2-4B` bundle used by the separate TRELLIS.2 pipeline.
-`mlx-spatial-lito validate` checks all four runtime files; `inspect` reads only
-the two LiTo safetensors.
+The embedded sparse-structure decoder converts LiTo voxel latents into Gaussian
+initialization coordinates. It is the exact checkpoint from
+`microsoft/TRELLIS-image-large` revision
+`25e0d31ffbebe4b5a97464dd851910efc3002d96`, not a dependency on the separate
+TRELLIS.2 pipeline. `mlx-spatial-lito validate` requires the bundle-local
+decoder and never searches an external TRELLIS root. `inspect` reads only the
+two LiTo safetensors.
 
 Maintainers can print Apple CDN download commands and convert local `.ckpt` files:
 
@@ -85,6 +84,8 @@ checkpoint records the exact policy, logical tensor shapes, bit width, group
 size, and affine mode in safetensors metadata. The normal LiTo loader detects
 that metadata and executes packed matrices directly with MLX quantized matrix
 multiplication; no Torch or intermediate dequantized checkpoint is involved.
+The root-level quantizer copies the embedded TRELLIS decoder and its provenance
+files into the output bundle unchanged.
 
 Pass the new root to inference exactly as you would the full-precision root:
 

docs/model-publishing.md

@@ -42,9 +42,14 @@ The tracked model-card source lives under:
 
 ```text
 model-cards/lito-research-mlx/
+model-cards/lito-research-mlx-8bit/
 ```
 
 LiTo is research-only and non-commercial under Apple's model license. The model repository must include `LICENSE_MODEL`, identify the safetensors files as an unofficial converted derivative, and avoid language that implies Apple endorsement.
+Both LiTo variants must embed the required Microsoft TRELLIS sparse-structure
+decoder under `dependencies/trellis/`, together with its MIT license and an
+immutable source manifest. Published LiTo bundles must not depend on a sibling
+TRELLIS checkout.
 
 ## What The Model Card Should Contain
 

model-cards/lito-research-mlx-8bit/LICENSE_MODEL

@@ -0,0 +1,88 @@
+Disclaimer: IMPORTANT: This Apple Machine Learning Research Model is
+specifically developed and released by Apple Inc. ("Apple") for the sole purpose
+of scientific research of artificial intelligence and machine-learning
+technology. “Apple Machine Learning Research Model” means the model, including
+but not limited to algorithms, formulas, trained model weights, parameters,
+configurations, checkpoints, and any related materials (including
+documentation).
+
+This Apple Machine Learning Research Model is provided to You by
+Apple in consideration of your agreement to the following terms, and your use,
+modification, creation of Model Derivatives, and or redistribution of the Apple
+Machine Learning Research Model constitutes acceptance of this Agreement. If You
+do not agree with these terms, please do not use, modify, create Model
+Derivatives of, or distribute this Apple Machine Learning Research Model or
+Model Derivatives.
+
+* License Scope: In consideration of your agreement to abide by the following
+  terms, and subject to these terms, Apple hereby grants you a personal,
+  non-exclusive, worldwide, non-transferable, royalty-free, revocable, and
+  limited license, to use, copy, modify, distribute, and create Model
+  Derivatives (defined below) of the Apple Machine Learning Research Model
+  exclusively for Research Purposes. You agree that any Model Derivatives You
+  may create or that may be created for You will be limited to Research Purposes
+  as well. “Research Purposes” means non-commercial scientific research and
+  academic development activities, such as experimentation, analysis, testing
+  conducted by You with the sole intent to advance scientific knowledge and
+  research. “Research Purposes” does not include any commercial exploitation,
+  product development or use in any commercial product or service.
+
+* Distribution of Apple Machine Learning Research Model and Model Derivatives:
+  If you choose to redistribute Apple Machine Learning Research Model or its
+  Model Derivatives, you must provide a copy of this Agreement to such third
+  party, and ensure that the following attribution notice be provided: “Apple
+  Machine Learning Research Model is licensed under the Apple Machine Learning
+  Research Model License Agreement.” Additionally, all Model Derivatives must
+  clearly be identified as such, including disclosure of modifications and
+  changes made to the Apple Machine Learning Research Model. The name,
+  trademarks, service marks or logos of Apple may not be used to endorse or
+  promote Model Derivatives or the relationship between You and Apple. “Model
+  Derivatives” means any models or any other artifacts created by modifications,
+  improvements, adaptations, alterations to the architecture, algorithm or
+  training processes of the Apple Machine Learning Research Model, or by any
+  retraining, fine-tuning of the Apple Machine Learning Research Model.
+
+* No Other License: Except as expressly stated in this notice, no other rights
+  or licenses, express or implied, are granted by Apple herein, including but
+  not limited to any patent, trademark, and similar intellectual property rights
+  worldwide that may be infringed by the Apple Machine Learning Research Model,
+  the Model Derivatives or by other works in which the Apple Machine Learning
+  Research Model may be incorporated.
+
+* Compliance with Laws: Your use of Apple Machine Learning Research Model must
+  be in compliance with all applicable laws and regulations.
+
+* Term and Termination: The term of this Agreement will begin upon your
+  acceptance of this Agreement or use of the Apple Machine Learning Research
+  Model and will continue until terminated in accordance with the following
+  terms. Apple may terminate this Agreement at any time if You are in breach of
+  any term or condition of this Agreement. Upon termination of this Agreement,
+  You must cease to use all Apple Machine Learning Research Models and Model
+  Derivatives and permanently delete any copy thereof. Sections 3, 6 and 7 will
+  survive termination.
+
+* Disclaimer and Limitation of Liability: This Apple Machine Learning Research
+  Model and any outputs generated by the Apple Machine Learning Research Model
+  are provided on an “AS IS” basis. APPLE MAKES NO WARRANTIES, EXPRESS OR
+  IMPLIED, INCLUDING WITHOUT LIMITATION THE IMPLIED WARRANTIES OF
+  NON-INFRINGEMENT, MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE,
+  REGARDING THE APPLE MACHINE LEARNING RESEARCH MODEL OR OUTPUTS GENERATED BY
+  THE APPLE MACHINE LEARNING RESEARCH MODEL. You are solely responsible for
+  determining the appropriateness of using or redistributing the Apple Machine
+  Learning Research Model and any outputs of the Apple Machine Learning Research
+  Model and assume any risks associated with Your use of the Apple Machine
+  Learning Research Model and any output and results. IN NO EVENT SHALL APPLE BE
+  LIABLE FOR ANY SPECIAL, INDIRECT, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING
+  IN ANY WAY OUT OF THE USE, REPRODUCTION, MODIFICATION AND/OR DISTRIBUTION OF
+  THE APPLE MACHINE LEARNING RESEARCH MODEL AND ANY OUTPUTS OF THE APPLE MACHINE
+  LEARNING RESEARCH MODEL, HOWEVER CAUSED AND WHETHER UNDER THEORY OF CONTRACT,
+  TORT (INCLUDING NEGLIGENCE), STRICT LIABILITY OR OTHERWISE, EVEN IF APPLE HAS
+  BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+* Governing Law: This Agreement will be governed by and construed under the laws
+  of the State of California without regard to its choice of law principles. The
+  Convention on Contracts for the International Sale of Goods shall not apply to
+  the Agreement except that the arbitration clause and any arbitration hereunder
+  shall be governed by the Federal Arbitration Act, Chapters 1 and 2.
+
+Copyright (C) 2026 Apple Inc. All Rights Reserved.

model-cards/lito-research-mlx-8bit/README.md

@@ -0,0 +1,121 @@
+---
+license: other
+license_name: apple-machine-learning-research-model-license-agreement
+license_link: https://github.com/apple/ml-lito/blob/main/LICENSE_MODEL
+library_name: mlx
+pipeline_tag: image-to-3d
+base_model:
+  - appautomaton/lito-research-mlx
+tags:
+  - mlx
+  - apple-silicon
+  - safetensors
+  - image-to-3d
+  - gaussian-splatting
+  - 3dgs
+  - 8-bit
+  - affine-quantization
+  - runtime-only
+  - research-only
+  - non-commercial
+---
+
+# 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
+
+```bash
+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`:
+
+```text
+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](https://huggingface.co/appautomaton/lito-research-mlx)
+- [`appautomaton/mlx-spatial`](https://github.com/appautomaton/mlx-spatial)
+- [Apple LiTo project](https://apple.github.io/ml-lito/)
+- [Apple LiTo source](https://github.com/apple/ml-lito)

model-cards/lito-research-mlx-8bit/dependencies/trellis/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) Microsoft Corporation.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

model-cards/lito-research-mlx-8bit/dependencies/trellis/SOURCE.json

@@ -0,0 +1,15 @@
+{
+  "files": {
+    "ckpts/ss_dec_conv3d_16l8_fp16.json": {
+      "bytes": 245,
+      "sha256": "646781293f1cda74720de85d1cef50a957fb4aebd9a4bd014e454e32f2330ac5"
+    },
+    "ckpts/ss_dec_conv3d_16l8_fp16.safetensors": {
+      "bytes": 147591972,
+      "sha256": "1c76d4a40519aa2d711cc263a8404105231ac26db31d946bed48b84fee79009a"
+    }
+  },
+  "license": "MIT",
+  "repo_id": "microsoft/TRELLIS-image-large",
+  "revision": "25e0d31ffbebe4b5a97464dd851910efc3002d96"
+}