tamarher commited on
Commit
0a5844c
·
verified ·
1 Parent(s): ae325e0

Upload SAM3D MLX 8-bit affine runtime bundle

Browse files
.hfignore ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ .DS_Store
2
+ .cache/
3
+ __pycache__/
4
+ *.pyc
5
+ *.pyo
6
+ .pytest_cache/
7
+ .ruff_cache/
LICENSE ADDED
@@ -0,0 +1,52 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ SAM License
2
+ Last Updated: November 19, 2025
3
+
4
+ “Agreement” means the terms and conditions for use, reproduction, distribution and modification of the SAM Materials set forth herein.
5
+
6
+ “SAM Materials” means, collectively, Documentation and the models, software and algorithms, including machine-learning model code, trained model weights, inference-enabling code, training-enabling code, fine-tuning enabling code, and other elements of the foregoing distributed by Meta and made available under this Agreement.
7
+
8
+ “Documentation” means the specifications, manuals and documentation accompanying
9
+ SAM Materials distributed by Meta.
10
+
11
+ “Licensee” or “you” means you, or your employer or any other person or entity (if you are entering into this Agreement on such person or entity’s behalf), of the age required under applicable laws, rules or regulations to provide legal consent and that has legal authority to bind your employer or such other person or entity if you are entering in this Agreement on their behalf.
12
+
13
+ “Meta” or “we” means Meta Platforms Ireland Limited (if you are located in or, if you are an entity, your principal place of business is in the EEA or Switzerland) or Meta Platforms, Inc. (if you are located outside of the EEA or Switzerland).
14
+
15
+ “Sanctions” means any economic or trade sanctions or restrictions administered or enforced by the United States (including the Office of Foreign Assets Control of the U.S. Department of the Treasury (“OFAC”), the U.S. Department of State and the U.S. Department of Commerce), the United Nations, the European Union, or the United Kingdom.
16
+
17
+ “Trade Controls” means any of the following: Sanctions and applicable export and import controls.
18
+
19
+ By using or distributing any portion or element of the SAM Materials, you agree to be bound by this Agreement.
20
+
21
+ 1. License Rights and Redistribution.
22
+
23
+ a. Grant of Rights. You are granted a non-exclusive, worldwide, non-transferable and royalty-free limited license under Meta’s intellectual property or other rights owned by Meta embodied in the SAM Materials to use, reproduce, distribute, copy, create derivative works of, and make modifications to the SAM Materials.
24
+
25
+ i. Grant of Patent License. Subject to the terms and conditions of this License, you are granted a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this section) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Work, where such license applies only to those patent claims licensable by Meta that are necessarily infringed alone or by combination of their contribution(s) with the SAM 3 Materials. If you institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the SAM 3 Materials incorporated within the work constitutes direct or contributory patent infringement, then any patent licenses granted to you under this License for that work shall terminate as of the date such litigation is filed.
26
+
27
+ b. Redistribution and Use.
28
+
29
+ i. Distribution of SAM Materials, and any derivative works thereof, are subject to the terms of this Agreement. If you distribute or make the SAM Materials, or any derivative works thereof, available to a third party, you may only do so under the terms of this Agreement and you shall provide a copy of this Agreement with any such SAM Materials.
30
+
31
+ ii. If you submit for publication the results of research you perform on, using, or otherwise in connection with SAM Materials, you must acknowledge the use of SAM Materials in your publication.
32
+
33
+ iii. Your use of the SAM Materials must comply with applicable laws and regulations, including Trade Control Laws and applicable privacy and data protection laws.
34
+ iv. Your use of the SAM Materials will not involve or encourage others to reverse engineer, decompile or discover the underlying components of the SAM Materials.
35
+ v. You are not the target of Trade Controls and your use of SAM Materials must comply with Trade Controls. You agree not to use, or permit others to use, SAM Materials for any activities subject to the International Traffic in Arms Regulations (ITAR) or end uses prohibited by Trade Controls, including those related to military or warfare purposes, nuclear industries or applications, espionage, or the development or use of guns or illegal weapons.
36
+ 2. User Support. Your use of the SAM Materials is done at your own discretion; Meta does not process any information nor provide any service in relation to such use. Meta is under no obligation to provide any support services for the SAM Materials. Any support provided is “as is”, “with all faults”, and without warranty of any kind.
37
+
38
+ 3. Disclaimer of Warranty. UNLESS REQUIRED BY APPLICABLE LAW, THE SAM MATERIALS AND ANY OUTPUT AND RESULTS THEREFROM ARE PROVIDED ON AN “AS IS” BASIS, WITHOUT WARRANTIES OF ANY KIND, AND META DISCLAIMS ALL WARRANTIES OF ANY KIND, BOTH EXPRESS AND IMPLIED, INCLUDING, WITHOUT LIMITATION, ANY WARRANTIES OF TITLE, NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. YOU ARE SOLELY RESPONSIBLE FOR DETERMINING THE APPROPRIATENESS OF USING OR REDISTRIBUTING THE SAM MATERIALS AND ASSUME ANY RISKS ASSOCIATED WITH YOUR USE OF THE SAM MATERIALS AND ANY OUTPUT AND RESULTS.
39
+
40
+ 4. Limitation of Liability. IN NO EVENT WILL META OR ITS AFFILIATES BE LIABLE UNDER ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, TORT, NEGLIGENCE, PRODUCTS LIABILITY, OR OTHERWISE, ARISING OUT OF THIS AGREEMENT, FOR ANY LOST PROFITS OR ANY DIRECT OR INDIRECT, SPECIAL, CONSEQUENTIAL, INCIDENTAL, EXEMPLARY OR PUNITIVE DAMAGES, EVEN IF META OR ITS AFFILIATES HAVE BEEN ADVISED OF THE POSSIBILITY OF ANY OF THE FOREGOING.
41
+
42
+ 5. Intellectual Property.
43
+
44
+ a. Subject to Meta’s ownership of SAM Materials and derivatives made by or for Meta, with respect to any derivative works and modifications of the SAM Materials that are made by you, as between you and Meta, you are and will be the owner of such derivative works and modifications.
45
+
46
+ b. If you institute litigation or other proceedings against Meta or any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the SAM Materials, outputs or results, or any portion of any of the foregoing, constitutes infringement of intellectual property or other rights owned or licensable by you, then any licenses granted to you under this Agreement shall terminate as of the date such litigation or claim is filed or instituted. You will indemnify and hold harmless Meta from and against any claim by any third party arising out of or related to your use or distribution of the SAM Materials.
47
+
48
+ 6. Term and Termination. The term of this Agreement will commence upon your acceptance of this Agreement or access to the SAM Materials and will continue in full force and effect until terminated in accordance with the terms and conditions herein. Meta may terminate this Agreement if you are in breach of any term or condition of this Agreement. Upon termination of this Agreement, you shall delete and cease use of the SAM Materials. Sections 3, 4 and 7 shall survive the termination of this Agreement.
49
+
50
+ 7. Governing Law and Jurisdiction. This Agreement will be governed and construed under the laws of the State of California without regard to choice of law principles, and the UN Convention on Contracts for the International Sale of Goods does not apply to this Agreement. The courts of California shall have exclusive jurisdiction of any dispute arising out of this Agreement.
51
+
52
+ 8. Modifications and Amendments. Meta may modify this Agreement from time to time; provided that they are similar in spirit to the current version of the Agreement, but may differ in detail to address new problems or concerns. All such changes will be effective immediately. Your continued use of the SAM Materials after any modification to this Agreement constitutes your agreement to such modification. Except as provided in this Agreement, no modification or addition to any provision of this Agreement will be binding unless it is in writing and signed by an authorized representative of both you and Meta.
README.md ADDED
@@ -0,0 +1,239 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ license_name: sam-license
4
+ license_link: https://github.com/facebookresearch/sam-3d-objects/blob/main/LICENSE
5
+ library_name: mlx
6
+ pipeline_tag: image-to-3d
7
+ base_model:
8
+ - appautomaton/sam-3d-objects-mlx
9
+ base_model_relation: quantized
10
+ tags:
11
+ - mlx
12
+ - apple-silicon
13
+ - safetensors
14
+ - sam-3d
15
+ - sam-3d-objects
16
+ - image-to-3d
17
+ - 3d-reconstruction
18
+ - gaussian-splatting
19
+ - mesh
20
+ - glb
21
+ - 8-bit
22
+ - affine-quantization
23
+ - mixed-precision
24
+ ---
25
+
26
+ # SAM 3D Objects MLX 8-bit Affine for `mlx-spatial`
27
+
28
+ <p align="center">
29
+ <a href="https://appautomaton.github.io"><img alt="App Automaton project" src="https://img.shields.io/badge/App_Automaton-Project-5B5BD6?style=for-the-badge"></a>
30
+ <a href="https://github.com/appautomaton/mlx-spatial"><img alt="GitHub — appautomaton/mlx-spatial" src="https://img.shields.io/badge/GitHub-mlx--spatial-181717?style=for-the-badge&amp;logo=github&amp;logoColor=white"></a>
31
+ <a href="https://appautomaton.renocrypt.com/mlx-spatial/"><img alt="mlx-spatial documentation" src="https://img.shields.io/badge/Documentation-mlx--spatial-0A7BBB?style=for-the-badge&amp;logo=readthedocs&amp;logoColor=white"></a>
32
+ <a href="https://pypi.org/project/mlx-spatial/"><img alt="mlx-spatial on PyPI" src="https://img.shields.io/pypi/v/mlx-spatial?style=for-the-badge&amp;logo=pypi&amp;logoColor=white&amp;label=PyPI"></a>
33
+ </p>
34
+
35
+ <p align="center">
36
+ <a href="https://huggingface.co/appautomaton/sam-3d-objects-mlx"><img alt="SAM 3D Objects MLX full-precision variant" src="https://img.shields.io/badge/Hugging_Face-MLX_FP-FFD21E?style=for-the-badge&amp;logo=huggingface&amp;logoColor=000"></a>
37
+ <a href="https://huggingface.co/appautomaton/sam-3d-objects-mlx-8bit"><img alt="SAM 3D Objects MLX 8-bit variant" src="https://img.shields.io/badge/Hugging_Face-MLX_8--bit-FF9D00?style=for-the-badge&amp;logo=huggingface&amp;logoColor=000"></a>
38
+ </p>
39
+
40
+ A self-contained SAM 3D Objects inference bundle with selective 8-bit affine
41
+ weights, built for direct execution by
42
+ [`mlx-spatial`](https://github.com/appautomaton/mlx-spatial) on Apple Silicon.
43
+ Large transformer matrices run through MLX's packed quantized matrix
44
+ multiplication, while convolutional and accuracy-sensitive boundary tensors
45
+ remain in their original precision.
46
+
47
+ > This is an unofficial quantized derivative. It is not a Meta or MoGe
48
+ > release.
49
+
50
+ ## Variants
51
+
52
+ | Variant | Precision | Checkpoint size | Model |
53
+ | --- | --- | ---: | --- |
54
+ | MLX converted | Source FP16/FP32 | 13.705 GB | [`sam-3d-objects-mlx`](https://huggingface.co/appautomaton/sam-3d-objects-mlx) |
55
+ | MLX 8-bit | Selective affine INT8 with retained FP16/FP32 | 4.625 GB | **This model** |
56
+
57
+ The 8-bit checkpoint payload is 66.3% smaller. Both variants expose the same
58
+ 3,783 logical tensors with identical names, shapes, and declared source
59
+ dtypes. The full-precision repository is not required to run this variant.
60
+
61
+ ## Compatibility
62
+
63
+ This format requires an `mlx-spatial` build that includes SAM3D affine
64
+ checkpoint support. Until that support is available in a tagged PyPI release,
65
+ install the current project revision:
66
+
67
+ ```bash
68
+ pip install \
69
+ "mlx-spatial @ git+https://github.com/appautomaton/mlx-spatial.git@main"
70
+ ```
71
+
72
+ The runtime targets Apple Silicon and MLX `0.32.x`. It does not use Torch,
73
+ CUDA, or a dequantized full-precision checkpoint.
74
+
75
+ ## Use
76
+
77
+ Download the complete bundle:
78
+
79
+ ```bash
80
+ hf download appautomaton/sam-3d-objects-mlx-8bit \
81
+ --local-dir weights/sam-3d-objects-mlx-8bit
82
+ ```
83
+
84
+ Validate and inspect it without loading all weights:
85
+
86
+ ```bash
87
+ mlx-spatial-sam3d validate weights/sam-3d-objects-mlx-8bit
88
+ mlx-spatial-sam3d inspect weights/sam-3d-objects-mlx-8bit
89
+ ```
90
+
91
+ Generate a Gaussian Splat PLY:
92
+
93
+ ```bash
94
+ mlx-spatial-sam3d reconstruct \
95
+ weights/sam-3d-objects-mlx-8bit \
96
+ inputs/sam3d/object-rmbg.png \
97
+ --mask inputs/sam3d/object-mask.png \
98
+ --moge-root weights/sam-3d-objects-mlx-8bit/moge \
99
+ --output outputs/sam3d/object-8bit/gaussians.ply \
100
+ --trace-output outputs/sam3d/object-8bit/trace.json \
101
+ --memory-profile balanced
102
+ ```
103
+
104
+ Add `--glb-output outputs/sam3d/object-8bit/object.glb` to decode and export a
105
+ mesh-backed GLB in the same run.
106
+
107
+ SAM3D requires a binary object mask aligned with the input image. A clean
108
+ foreground extraction is strongly recommended: transparent background pixels
109
+ should have alpha zero, and the supplied mask should cover the intended object
110
+ without including the surrounding scene.
111
+
112
+ ## Bundle
113
+
114
+ The repository preserves the directory and checkpoint names used by the
115
+ full-precision MLX bundle:
116
+
117
+ | Checkpoint | Logical tensors | INT8 matrices | Bytes |
118
+ | --- | ---: | ---: | ---: |
119
+ | `checkpoints/ss_generator.safetensors` | 1,741 | 556 | 2,197,959,476 |
120
+ | `checkpoints/ss_decoder.safetensors` | 74 | 0 | 147,592,136 |
121
+ | `checkpoints/slat_generator.safetensors` | 1,225 | 384 | 1,578,129,504 |
122
+ | `checkpoints/slat_decoder_gs.safetensors` | 101 | 48 | 97,148,339 |
123
+ | `checkpoints/slat_decoder_gs_4.safetensors` | 101 | 48 | 95,942,210 |
124
+ | `checkpoints/slat_decoder_mesh.safetensors` | 120 | 48 | 119,546,584 |
125
+ | `moge/model.safetensors` | 421 | 96 | 388,551,606 |
126
+ | **Total** | **3,783** | **1,180** | **4,624,869,855** |
127
+
128
+ The remaining YAML configuration, conversion metadata, audit, and license
129
+ files are small and are included alongside the checkpoints.
130
+
131
+ ## Quantization
132
+
133
+ The quantization scheme is affine 8-bit with group size 64. Packed weights are
134
+ stored as `uint32` with FP32 scales and biases and are executed directly with
135
+ `mx.quantized_matmul`.
136
+
137
+ The following block-internal two-dimensional weights are quantized:
138
+
139
+ - DINO condition-encoder attention and MLP matrices;
140
+ - the point-condition transformer block;
141
+ - sparse-structure and structured-latent generator transformer blocks;
142
+ - Gaussian and mesh decoder transformer torso blocks;
143
+ - the bundled MoGe ViT backbone attention and MLP matrices.
144
+
145
+ The following tensors remain in their original precision:
146
+
147
+ - all dense, sparse, and transposed convolutions;
148
+ - the complete sparse-structure decoder;
149
+ - patch embeddings, input mappings, condition projections, and latent
150
+ mappings;
151
+ - normalization parameters, biases, learned tokens, and positional tensors;
152
+ - final Gaussian, mesh, geometry, pose, and MoGe output heads;
153
+ - matrices whose input dimension is incompatible with group size 64.
154
+
155
+ Physical packed arrays use internal qweight, scale, and bias suffixes. The
156
+ runtime reconstructs the original logical tensor names from metadata, so model
157
+ code and configuration continue to use the same checkpoint contract. Format
158
+ details are embedded under `mlx_spatial.sam3d.quantization` in each quantized
159
+ safetensors file.
160
+
161
+ ## Reproducing the Bundle
162
+
163
+ Starting from the full-precision MLX conversion:
164
+
165
+ ```bash
166
+ mlx-spatial-sam3d-quantize \
167
+ weights/sam-3d-objects-mlx \
168
+ weights/sam-3d-objects-mlx-8bit \
169
+ --bits 8 \
170
+ --group-size 64
171
+ ```
172
+
173
+ The quantized output is a complete runtime root. Do not place copies of the
174
+ full-precision checkpoints in the 8-bit repository.
175
+
176
+ ## Verification
177
+
178
+ - Logical checkpoint inspection matched all 3,783 source tensor names, shapes,
179
+ and dtypes exactly.
180
+ - The quantization inventory contains 1,180 packed matrices across six
181
+ quantized checkpoints; the convolutional sparse-structure decoder remains
182
+ unchanged.
183
+ - A sampled static weight audit measured median 8-bit reconstruction SQNR of
184
+ approximately 45 dB across checkpoint groups, with a worst observed matrix
185
+ at 41 dB.
186
+ - The SAM3D test suite passed 163 tests, with one test deselected.
187
+ - A background-removed 900 × 900 input with an aligned mask completed all nine
188
+ inference stages with no blocker and produced 495,008 finite Gaussians.
189
+ - The generated geometry passed the runtime's nominal axis-range check.
190
+
191
+ The end-to-end run establishes runtime compatibility and artifact health. It
192
+ is not a formal claim of visual equivalence to the full-precision model, and
193
+ the observed run is not presented as a general performance benchmark.
194
+
195
+ ## Limitations
196
+
197
+ - Quantization can change sparse occupancy, geometry, appearance, and other
198
+ generation details relative to the full-precision variant.
199
+ - A useful object-aligned mask is required. Background leakage, broad masks,
200
+ soft edges, reflections, and thin structures can reduce reconstruction
201
+ quality.
202
+ - Single-view reconstruction cannot determine unseen geometry with certainty.
203
+ - The packed checkpoint format requires `mlx-spatial`; generic safetensors
204
+ loaders will see the physical packed arrays rather than the logical weights.
205
+ - This bundle supports inference, not training or fine-tuning.
206
+ - Standard PLY viewers may not render Gaussian Splat fields correctly. Use a
207
+ 3DGS-aware viewer for the Gaussian artifact, or request GLB export.
208
+
209
+ ## License and Attribution
210
+
211
+ This bundle is derived from Meta's SAM 3D Objects release and is distributed
212
+ under the SAM License. Read the bundled `LICENSE` and the
213
+ [upstream SAM License](https://github.com/facebookresearch/sam-3d-objects/blob/main/LICENSE)
214
+ before use. Redistribution of SAM Materials and derivative works remains
215
+ subject to that agreement.
216
+
217
+ The bundle also contains a converted and quantized
218
+ [`Ruicheng/moge-vitl`](https://huggingface.co/Ruicheng/moge-vitl) checkpoint
219
+ used for pointmap estimation. That checkpoint is published under Apache 2.0.
220
+ Users are responsible for complying with both sets of terms.
221
+
222
+ This repository is not affiliated with or endorsed by Meta or the MoGe
223
+ authors. Publications using these weights should acknowledge the original SAM
224
+ 3D Objects and MoGe work.
225
+
226
+ ## Links
227
+
228
+ - [App Automaton](https://appautomaton.github.io)
229
+ - [AppAutomaton models on Hugging Face](https://huggingface.co/appautomaton)
230
+ - [`appautomaton/mlx-spatial`](https://github.com/appautomaton/mlx-spatial) — MLX-native 3D and spatial inference for Apple Silicon.
231
+ - [`mlx-spatial` documentation](https://appautomaton.renocrypt.com/mlx-spatial/)
232
+ - [`mlx-spatial` on PyPI](https://pypi.org/project/mlx-spatial/)
233
+ - [SAM3D guide](https://github.com/appautomaton/mlx-spatial/blob/main/docs/sam3d.md)
234
+ - [Full-precision MLX variant](https://huggingface.co/appautomaton/sam-3d-objects-mlx)
235
+ - [Upstream SAM 3D Objects model](https://huggingface.co/facebook/sam-3d-objects)
236
+ - [Upstream SAM 3D Objects source](https://github.com/facebookresearch/sam-3d-objects)
237
+ - [MoGe ViT-L checkpoint](https://huggingface.co/Ruicheng/moge-vitl)
238
+ - [MoGe source](https://github.com/microsoft/moge)
239
+ - [MLX](https://github.com/ml-explore/mlx)
checkpoints/conversion_metadata/slat_decoder_gs.yaml ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ source_sha256: f8077c36a06eaf890dd93cda1937411f793dea1eb80b3dd9329f2038ba84a111
2
+ converter: mlx-spatial-restricted-torch-zip
3
+ tensor_count: 101
checkpoints/conversion_metadata/slat_decoder_gs_4.yaml ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ source_sha256: 731a0eceaa47945b52aa27f650d695b2aea9cc70945751e5609e5cb5b49f0186
2
+ converter: mlx-spatial-restricted-torch-zip
3
+ tensor_count: 101
checkpoints/conversion_metadata/slat_decoder_mesh.yaml ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ source_sha256: 85907b37b67d8ce5b099a96629bdcfbd873eb407dee6b3aa9a75deb15038db33
2
+ converter: mlx-spatial-restricted-torch-zip
3
+ tensor_count: 120
checkpoints/conversion_metadata/slat_generator.yaml ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ source_sha256: 91529bde8e7daa12d09618a66c319e3a5a6398db6b23b958cedcb1c3f28faabb
2
+ converter: mlx-spatial-restricted-torch-zip
3
+ tensor_count: 1225
checkpoints/conversion_metadata/ss_decoder.yaml ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ source_sha256: 6dac1cd7b7fda5a38e0614fadae441f1794f80e39ea2981f1ac8aff0a7e99340
2
+ converter: mlx-spatial-restricted-torch-zip
3
+ tensor_count: 74
checkpoints/conversion_metadata/ss_generator.yaml ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ source_sha256: 225f40479e4cff4f39d6fa14c55be3abad1475bf55b61af3bec1e19ed2f6c146
2
+ converter: mlx-spatial-restricted-torch-zip
3
+ tensor_count: 1741
checkpoints/pipeline.yaml ADDED
@@ -0,0 +1,106 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ _target_: sam3d_objects.pipeline.inference_pipeline_pointmap.InferencePipelinePointMap
2
+ ss_generator_config_path: ss_generator.yaml
3
+ ss_generator_ckpt_path: ss_generator.safetensors
4
+ slat_generator_config_path: slat_generator.yaml
5
+ slat_generator_ckpt_path: slat_generator.safetensors
6
+ ss_decoder_config_path: ss_decoder.yaml
7
+ ss_decoder_ckpt_path: ss_decoder.safetensors
8
+ slat_decoder_gs_config_path: slat_decoder_gs.yaml
9
+ slat_decoder_gs_ckpt_path: slat_decoder_gs.safetensors
10
+ slat_decoder_gs_4_config_path: slat_decoder_gs_4.yaml
11
+ slat_decoder_gs_4_ckpt_path: slat_decoder_gs_4.safetensors
12
+ slat_decoder_mesh_config_path: slat_decoder_mesh.yaml
13
+ slat_decoder_mesh_ckpt_path: slat_decoder_mesh.safetensors
14
+ pad_size: 1.0
15
+ dtype: float16
16
+ version: 3dfy_v9
17
+ slat_cfg_strength: 1
18
+ slat_rescale_t: 1
19
+ downsample_ss_dist: 1
20
+ compile_model: true
21
+ ss_condition_input_mapping: []
22
+ ss_preprocessor:
23
+ _target_: sam3d_objects.data.dataset.tdfy.preprocessor.PreProcessor
24
+ img_mask_joint_transform: []
25
+ img_mask_pointmap_joint_transform:
26
+ - _partial_: true
27
+ _target_: sam3d_objects.data.dataset.tdfy.img_and_mask_transforms.resize_all_to_same_size
28
+ - _partial_: true
29
+ _target_: sam3d_objects.data.dataset.tdfy.img_and_mask_transforms.crop_around_mask_with_padding
30
+ box_size_factor: 1.2
31
+ padding_factor: 0.0
32
+ img_transform:
33
+ _target_: torchvision.transforms.Compose
34
+ transforms:
35
+ - _partial_: true
36
+ _target_: sam3d_objects.data.dataset.tdfy.img_processing.pad_to_square_centered
37
+ - _target_: torchvision.transforms.Resize
38
+ size: 518
39
+ mask_transform:
40
+ _target_: torchvision.transforms.Compose
41
+ transforms:
42
+ - _partial_: true
43
+ _target_: sam3d_objects.data.dataset.tdfy.img_processing.pad_to_square_centered
44
+ - _target_: torchvision.transforms.Resize
45
+ interpolation: 0
46
+ size: 518
47
+ normalize_pointmap: true
48
+ pointmap_normalizer:
49
+ _target_: sam3d_objects.data.dataset.tdfy.img_and_mask_transforms.ObjectCentricSSI
50
+ allow_scale_and_shift_override: true
51
+ use_scene_scale: true
52
+ pointmap_transform:
53
+ _target_: torchvision.transforms.Compose
54
+ transforms:
55
+ - _partial_: true
56
+ _target_: sam3d_objects.data.dataset.tdfy.img_processing.pad_to_square_centered
57
+ - _target_: torchvision.transforms.Resize
58
+ interpolation: 0
59
+ size: 518
60
+ pose_decoder_name: ScaleShiftInvariant
61
+ depth_model:
62
+ _target_: sam3d_objects.pipeline.depth_models.moge.MoGe
63
+ model:
64
+ _target_: moge.model.v1.MoGeModel.from_pretrained
65
+ pretrained_model_name_or_path: Ruicheng/moge-vitl
66
+ slat_condition_input_mapping: []
67
+ slat_preprocessor:
68
+ _target_: sam3d_objects.data.dataset.tdfy.preprocessor.PreProcessor
69
+ img_transform:
70
+ _target_: torchvision.transforms.Compose
71
+ transforms:
72
+ - _target_: sam3d_objects.data.dataset.tdfy.img_processing.pad_to_square_centered
73
+ _partial_: true
74
+ - _target_: torchvision.transforms.Resize
75
+ size: 518
76
+ mask_transform:
77
+ _target_: torchvision.transforms.Compose
78
+ transforms:
79
+ - _target_: sam3d_objects.data.dataset.tdfy.img_processing.pad_to_square_centered
80
+ _partial_: true
81
+ - _target_: torchvision.transforms.Resize
82
+ size: 518
83
+ interpolation: 0
84
+ img_mask_joint_transform:
85
+ - _target_: sam3d_objects.data.dataset.tdfy.img_and_mask_transforms.crop_around_mask_with_padding
86
+ _partial_: true
87
+ box_size_factor: 1.2
88
+ padding_factor: 0.0
89
+ slat_mean:
90
+ - 0.12211431
91
+ - 0.37204156
92
+ - -1.26521907
93
+ - -2.05276058
94
+ - -3.10432536
95
+ - -0.11294304
96
+ - -0.85146744
97
+ - 0.45506954
98
+ slat_std:
99
+ - 2.37326008
100
+ - 2.13174402
101
+ - 2.2413953
102
+ - 2.30589401
103
+ - 2.1191894
104
+ - 1.8969511
105
+ - 2.41684989
106
+ - 2.08374642
checkpoints/slat_decoder_gs.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c3583403258ffa5c5e967e8fd224b06e5859608c4fa26c9a7122622bdf337c12
3
+ size 97148339
checkpoints/slat_decoder_gs.yaml ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ _target_: sam3d_objects.model.backbone.tdfy_dit.models.structured_latent_vae.decoder_gs.SLatGaussianDecoderTdfyWrapper
2
+ resolution: 64
3
+ model_channels: 768
4
+ latent_channels: 8
5
+ num_blocks: 12
6
+ num_heads: 12
7
+ mlp_ratio: 4
8
+ attn_mode: swin
9
+ window_size: 8
10
+ representation_config:
11
+ lr:
12
+ _xyz: 1.0
13
+ _features_dc: 1.0
14
+ _opacity: 1.0
15
+ _scaling: 1.0
16
+ _rotation: 0.1
17
+ perturb_offset: true
18
+ voxel_size: 1.5
19
+ num_gaussians: 32
20
+ 2d_filter_kernel_size: 0.1
21
+ 3d_filter_kernel_size: 0.0009
22
+ scaling_bias: 0.004
23
+ opacity_bias: 0.1
24
+ scaling_activation: softplus
25
+ use_fp16: true
checkpoints/slat_decoder_gs_4.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e1c6d84b537648b8a02796c1d11c73b17b6b2e44150c1e6b9a1b2ab26bb309ce
3
+ size 95942210
checkpoints/slat_decoder_gs_4.yaml ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ _target_: sam3d_objects.model.backbone.tdfy_dit.models.structured_latent_vae.decoder_gs.SLatGaussianDecoderTdfyWrapper
2
+ resolution: 64
3
+ model_channels: 768
4
+ latent_channels: 8
5
+ num_blocks: 12
6
+ num_heads: 12
7
+ mlp_ratio: 4
8
+ attn_mode: swin
9
+ window_size: 8
10
+ representation_config:
11
+ lr:
12
+ _xyz: 1.0
13
+ _features_dc: 1.0
14
+ _opacity: 1.0
15
+ _scaling: 1.0
16
+ _rotation: 0.1
17
+ perturb_offset: true
18
+ voxel_size: 1.5
19
+ num_gaussians: 4
20
+ 2d_filter_kernel_size: 0.1
21
+ 3d_filter_kernel_size: 0.0009
22
+ scaling_bias: 0.004
23
+ opacity_bias: 0.1
24
+ scaling_activation: softplus
25
+ use_fp16: true
checkpoints/slat_decoder_mesh.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a242278fe5a7fd4ae96d9af9e2c4860c0bd584d2fa9d52a708ed296ba0f7ae6d
3
+ size 119546584
checkpoints/slat_decoder_mesh.yaml ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ _target_: sam3d_objects.model.backbone.tdfy_dit.models.structured_latent_vae.decoder_mesh.SLatMeshDecoderTdfyWrapper
2
+ resolution: 64
3
+ model_channels: 768
4
+ latent_channels: 8
5
+ num_blocks: 12
6
+ num_heads: 12
7
+ mlp_ratio: 4
8
+ attn_mode: swin
9
+ window_size: 8
10
+ representation_config:
11
+ use_color: true
12
+ use_fp16: true
checkpoints/slat_generator.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c2225a8d84bbf154d4b56e65624111fc0dd89222bf17974885e8704cca1864a7
3
+ size 1578129504
checkpoints/slat_generator.yaml ADDED
@@ -0,0 +1,60 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ module:
2
+ condition_embedder:
3
+ backbone:
4
+ _target_: sam3d_objects.model.backbone.dit.embedder.embedder_fuser.EmbedderFuser
5
+ embedder_list:
6
+ - - _target_: sam3d_objects.model.backbone.dit.embedder.dino.Dino
7
+ dino_model: dinov2_vitl14_reg
8
+ input_size: 518
9
+ normalize_images: true
10
+ prenorm_features: true
11
+ - - - image
12
+ - cropped
13
+ - - rgb_image
14
+ - full
15
+ - - _target_: sam3d_objects.model.backbone.dit.embedder.dino.Dino
16
+ dino_model: dinov2_vitl14_reg
17
+ input_size: 518
18
+ normalize_images: true
19
+ prenorm_features: true
20
+ - - - mask
21
+ - cropped
22
+ - - rgb_image_mask
23
+ - full
24
+ projection_net_hidden_dim_multiplier: 4.0
25
+ use_pos_embedding: learned
26
+ generator:
27
+ backbone:
28
+ _target_: sam3d_objects.model.backbone.generator.flow_matching.model.FlowMatching
29
+ inference_steps: 12
30
+ reverse_fn:
31
+ _target_: sam3d_objects.model.backbone.generator.classifier_free_guidance.ClassifierFreeGuidance
32
+ backbone:
33
+ _target_: sam3d_objects.model.backbone.tdfy_dit.models.structured_latent_flow.SLatFlowModelTdfyWrapper
34
+ cond_channels: 1024
35
+ condition_embedder: null
36
+ force_zeros_cond: true
37
+ in_channels: 8
38
+ io_block_channels:
39
+ - 128
40
+ mlp_ratio: 4
41
+ model_channels: 1024
42
+ num_blocks: 24
43
+ num_heads: 16
44
+ num_io_res_blocks: 2
45
+ out_channels: 8
46
+ patch_size: 2
47
+ pe_mode: ape
48
+ qk_rms_norm: true
49
+ resolution: 64
50
+ use_fp16: true
51
+ p_unconditional: 0.0
52
+ strength: 0.0
53
+ unconditional_handling: add_flag
54
+ sigma_min: 0.0
55
+ time_scale: 1000.0
56
+ training_time_sampler_fn:
57
+ _partial_: true
58
+ _target_: sam3d_objects.model.backbone.generator.flow_matching.model.lognorm_sampler
59
+ mean: -1.0
60
+ std: 1.0
checkpoints/ss_decoder.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f5ac316ca86ca26a6bdc006ae1961c6c87d857e494f2a0463fd2c19825962baf
3
+ size 147592136
checkpoints/ss_decoder.yaml ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ _target_: sam3d_objects.model.backbone.tdfy_dit.models.sparse_structure_vae.SparseStructureDecoderTdfyWrapper
2
+ out_channels: 1
3
+ latent_channels: 8
4
+ num_res_blocks: 2
5
+ num_res_blocks_middle: 2
6
+ channels:
7
+ - 512
8
+ - 128
9
+ - 32
10
+ reshape_input_to_cube: false
checkpoints/ss_generator.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f1f372d8805a0edc15bc1ec8fdec2e14b899680ad02c0dbff98b166338be9ee9
3
+ size 2197959476
checkpoints/ss_generator.yaml ADDED
@@ -0,0 +1,141 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ module:
2
+ condition_embedder:
3
+ backbone:
4
+ _target_: sam3d_objects.model.backbone.dit.embedder.embedder_fuser.EmbedderFuser
5
+ drop_modalities_weight:
6
+ - - - pointmap
7
+ - rgb_pointmap
8
+ - 1.0
9
+ dropout_prob: 0.1
10
+ embedder_list:
11
+ - - _target_: sam3d_objects.model.backbone.dit.embedder.dino.Dino
12
+ dino_model: dinov2_vitl14_reg
13
+ input_size: 518
14
+ normalize_images: true
15
+ - - - image
16
+ - cropped
17
+ - - rgb_image
18
+ - full
19
+ - - _target_: sam3d_objects.model.backbone.dit.embedder.dino.Dino
20
+ dino_model: dinov2_vitl14_reg
21
+ input_size: 518
22
+ normalize_images: true
23
+ - - - mask
24
+ - cropped
25
+ - - rgb_image_mask
26
+ - full
27
+ - - _target_: sam3d_objects.model.backbone.dit.embedder.pointmap.PointPatchEmbed
28
+ embed_dim: 512
29
+ input_size: 256
30
+ patch_size: 8
31
+ remap_output: linear
32
+ - - - pointmap
33
+ - cropped
34
+ - - rgb_pointmap
35
+ - full
36
+ force_drop_modalities: null
37
+ freeze: true
38
+ projection_net_hidden_dim_multiplier: 4.0
39
+ use_pos_embedding: learned
40
+ generator:
41
+ backbone:
42
+ _target_: sam3d_objects.model.backbone.generator.shortcut.model.ShortCut
43
+ batch_mode: true
44
+ cfg_modalities:
45
+ - shape
46
+ inference_steps: 2
47
+ loss_weights:
48
+ 6drotation_normalized: 0.1
49
+ _target_: sam3d_objects.config.utils.make_dict
50
+ scale: 0.1
51
+ shape: 0
52
+ translation: 1.0
53
+ translation_scale: 0.0
54
+ ratio_cfg_samples_in_self_consistency_target: 0.25
55
+ rescale_t: 1
56
+ reverse_fn:
57
+ _target_: sam3d_objects.model.backbone.generator.classifier_free_guidance.ClassifierFreeGuidanceWithExternalUnconditionalProbability
58
+ backbone:
59
+ _target_: sam3d_objects.model.backbone.tdfy_dit.models.mot_sparse_structure_flow.SparseStructureFlowTdfyWrapper
60
+ cond_channels: 1024
61
+ condition_embedder: null
62
+ force_zeros_cond: true
63
+ freeze_d_time_embedder: true
64
+ freeze_shared_parameters: true
65
+ in_channels: 8
66
+ is_shortcut_model: true
67
+ latent_mapping:
68
+ 6drotation_normalized:
69
+ _target_: sam3d_objects.model.backbone.tdfy_dit.models.mm_latent.Latent
70
+ in_channels: 6
71
+ model_channels: 1024
72
+ pos_embedder:
73
+ _target_: sam3d_objects.model.backbone.tdfy_dit.models.mm_latent.LearntPositionEmbedder
74
+ model_channels: 1024
75
+ token_len: 1
76
+ scale:
77
+ _target_: sam3d_objects.model.backbone.tdfy_dit.models.mm_latent.Latent
78
+ in_channels: 3
79
+ model_channels: 1024
80
+ pos_embedder:
81
+ _target_: sam3d_objects.model.backbone.tdfy_dit.models.mm_latent.LearntPositionEmbedder
82
+ model_channels: 1024
83
+ token_len: 1
84
+ shape:
85
+ _target_: sam3d_objects.model.backbone.tdfy_dit.models.mm_latent.Latent
86
+ in_channels: 8
87
+ model_channels: 1024
88
+ pos_embedder:
89
+ _target_: sam3d_objects.model.backbone.tdfy_dit.models.mm_latent.ShapePositionEmbedder
90
+ model_channels: 1024
91
+ patch_size: 1
92
+ resolution: 16
93
+ translation:
94
+ _target_: sam3d_objects.model.backbone.tdfy_dit.models.mm_latent.Latent
95
+ in_channels: 3
96
+ model_channels: 1024
97
+ pos_embedder:
98
+ _target_: sam3d_objects.model.backbone.tdfy_dit.models.mm_latent.LearntPositionEmbedder
99
+ model_channels: 1024
100
+ token_len: 1
101
+ translation_scale:
102
+ _target_: sam3d_objects.model.backbone.tdfy_dit.models.mm_latent.Latent
103
+ in_channels: 1
104
+ model_channels: 1024
105
+ pos_embedder:
106
+ _target_: sam3d_objects.model.backbone.tdfy_dit.models.mm_latent.LearntPositionEmbedder
107
+ model_channels: 1024
108
+ token_len: 1
109
+ latent_share_transformer:
110
+ 6drotation_normalized:
111
+ - 6drotation_normalized
112
+ - translation
113
+ - scale
114
+ - translation_scale
115
+ mlp_ratio: 4
116
+ model_channels: 1024
117
+ num_blocks: 24
118
+ num_heads: 16
119
+ out_channels: 8
120
+ patch_size: 1
121
+ pe_mode: ape
122
+ qk_rms_norm: true
123
+ resolution: 16
124
+ use_checkpoint: false
125
+ use_fp16: false
126
+ interval:
127
+ - 0
128
+ - 500
129
+ p_unconditional: 0.1
130
+ strength: 2.0
131
+ unconditional_handling: add_flag
132
+ self_consistency_cfg_strength: 2.0
133
+ self_consistency_prob: 0.25
134
+ shortcut_loss_weight: 1.0
135
+ sigma_min: 0.0
136
+ time_scale: 1000.0
137
+ training_time_sampler_fn:
138
+ _partial_: true
139
+ _target_: sam3d_objects.model.backbone.generator.flow_matching.model.lognorm_sampler
140
+ mean: -1.0
141
+ std: 1.0
conversion_manifest.json ADDED
@@ -0,0 +1,123 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "items": [
3
+ {
4
+ "kind": "config",
5
+ "output_path": "weights/sam-3d-objects-mlx/checkpoints/ss_generator.yaml",
6
+ "role": "ss_generator",
7
+ "source_path": "weights/sam-3d-objects/checkpoints/ss_generator.yaml",
8
+ "source_sha256": null,
9
+ "status": "copied",
10
+ "tensor_count": null
11
+ },
12
+ {
13
+ "kind": "config",
14
+ "output_path": "weights/sam-3d-objects-mlx/checkpoints/slat_generator.yaml",
15
+ "role": "slat_generator",
16
+ "source_path": "weights/sam-3d-objects/checkpoints/slat_generator.yaml",
17
+ "source_sha256": null,
18
+ "status": "copied",
19
+ "tensor_count": null
20
+ },
21
+ {
22
+ "kind": "config",
23
+ "output_path": "weights/sam-3d-objects-mlx/checkpoints/ss_decoder.yaml",
24
+ "role": "ss_decoder",
25
+ "source_path": "weights/sam-3d-objects/checkpoints/ss_decoder.yaml",
26
+ "source_sha256": null,
27
+ "status": "copied",
28
+ "tensor_count": null
29
+ },
30
+ {
31
+ "kind": "config",
32
+ "output_path": "weights/sam-3d-objects-mlx/checkpoints/slat_decoder_gs.yaml",
33
+ "role": "slat_decoder_gs",
34
+ "source_path": "weights/sam-3d-objects/checkpoints/slat_decoder_gs.yaml",
35
+ "source_sha256": null,
36
+ "status": "copied",
37
+ "tensor_count": null
38
+ },
39
+ {
40
+ "kind": "config",
41
+ "output_path": "weights/sam-3d-objects-mlx/checkpoints/slat_decoder_mesh.yaml",
42
+ "role": "slat_decoder_mesh",
43
+ "source_path": "weights/sam-3d-objects/checkpoints/slat_decoder_mesh.yaml",
44
+ "source_sha256": null,
45
+ "status": "copied",
46
+ "tensor_count": null
47
+ },
48
+ {
49
+ "kind": "checkpoint",
50
+ "output_path": "weights/sam-3d-objects-mlx/checkpoints/ss_generator.safetensors",
51
+ "role": "ss_generator",
52
+ "source_path": "weights/sam-3d-objects/checkpoints/ss_generator.ckpt",
53
+ "source_sha256": null,
54
+ "status": "exists",
55
+ "tensor_count": 1741
56
+ },
57
+ {
58
+ "kind": "checkpoint",
59
+ "output_path": "weights/sam-3d-objects-mlx/checkpoints/slat_generator.safetensors",
60
+ "role": "slat_generator",
61
+ "source_path": "weights/sam-3d-objects/checkpoints/slat_generator.ckpt",
62
+ "source_sha256": null,
63
+ "status": "exists",
64
+ "tensor_count": 1225
65
+ },
66
+ {
67
+ "kind": "checkpoint",
68
+ "output_path": "weights/sam-3d-objects-mlx/checkpoints/ss_decoder.safetensors",
69
+ "role": "ss_decoder",
70
+ "source_path": "weights/sam-3d-objects/checkpoints/ss_decoder.ckpt",
71
+ "source_sha256": null,
72
+ "status": "exists",
73
+ "tensor_count": 74
74
+ },
75
+ {
76
+ "kind": "checkpoint",
77
+ "output_path": "weights/sam-3d-objects-mlx/checkpoints/slat_decoder_gs.safetensors",
78
+ "role": "slat_decoder_gs",
79
+ "source_path": "weights/sam-3d-objects/checkpoints/slat_decoder_gs.ckpt",
80
+ "source_sha256": null,
81
+ "status": "exists",
82
+ "tensor_count": 101
83
+ },
84
+ {
85
+ "kind": "checkpoint",
86
+ "output_path": "weights/sam-3d-objects-mlx/checkpoints/slat_decoder_mesh.safetensors",
87
+ "role": "slat_decoder_mesh",
88
+ "source_path": "weights/sam-3d-objects/checkpoints/slat_decoder_mesh.ckpt",
89
+ "source_sha256": null,
90
+ "status": "exists",
91
+ "tensor_count": 120
92
+ },
93
+ {
94
+ "kind": "config",
95
+ "output_path": "weights/sam-3d-objects-mlx/checkpoints/slat_decoder_gs_4.yaml",
96
+ "role": "slat_decoder_gs_4",
97
+ "source_path": "weights/sam-3d-objects/checkpoints/slat_decoder_gs_4.yaml",
98
+ "source_sha256": null,
99
+ "status": "copied",
100
+ "tensor_count": null
101
+ },
102
+ {
103
+ "kind": "checkpoint",
104
+ "output_path": "weights/sam-3d-objects-mlx/checkpoints/slat_decoder_gs_4.safetensors",
105
+ "role": "slat_decoder_gs_4",
106
+ "source_path": "weights/sam-3d-objects/checkpoints/slat_decoder_gs_4.ckpt",
107
+ "source_sha256": null,
108
+ "status": "exists",
109
+ "tensor_count": 101
110
+ },
111
+ {
112
+ "kind": "config",
113
+ "output_path": "weights/sam-3d-objects-mlx/checkpoints/pipeline.yaml",
114
+ "role": "pipeline",
115
+ "source_path": "weights/sam-3d-objects/checkpoints/pipeline.yaml",
116
+ "source_sha256": null,
117
+ "status": "rewritten",
118
+ "tensor_count": null
119
+ }
120
+ ],
121
+ "output_pipeline_path": "weights/sam-3d-objects-mlx/checkpoints/pipeline.yaml",
122
+ "source_root": "weights/sam-3d-objects"
123
+ }
moge/conversion_metadata/model.yaml ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ source_sha256: da96b09a0485a3c45a5aa455e67743c8b4efc4dd8437c1f2aa93c2b4303d957f
2
+ converter: mlx-spatial-restricted-torch-zip
3
+ tensor_count: 421
moge/model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:286ce03f9451de36b7a0f67a06464492f0abaad2734571d8e27e5622bcd1844e
3
+ size 388551606
weight-audit-source-vs-mlx.json ADDED
@@ -0,0 +1,149 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "compared_tensors": 1741,
4
+ "converted": "weights/sam-3d-objects-mlx/checkpoints/ss_generator.safetensors",
5
+ "converted_tensors": 1741,
6
+ "dtype_pairs": {
7
+ "float32->float32": 1741
8
+ },
9
+ "exact_tensors_float32_compare": 1741,
10
+ "extra_count": 0,
11
+ "extra_in_converted": [],
12
+ "max_abs_diff_float32_compare": 0.0,
13
+ "max_abs_diff_key": null,
14
+ "max_rel_diff_float32_compare": 0.0,
15
+ "max_rel_diff_key": null,
16
+ "missing_count": 0,
17
+ "missing_in_converted": [],
18
+ "nonzero_diff_tensors_float32_compare": 0,
19
+ "role": "ss_generator",
20
+ "sample_nonzero_diffs": [],
21
+ "shape_mismatch_count": 0,
22
+ "shape_mismatch_sample": [],
23
+ "source": "weights/sam-3d-objects/checkpoints/ss_generator.ckpt",
24
+ "source_tensors": 1741
25
+ },
26
+ {
27
+ "compared_tensors": 1225,
28
+ "converted": "weights/sam-3d-objects-mlx/checkpoints/slat_generator.safetensors",
29
+ "converted_tensors": 1225,
30
+ "dtype_pairs": {
31
+ "float32->float32": 1225
32
+ },
33
+ "exact_tensors_float32_compare": 1225,
34
+ "extra_count": 0,
35
+ "extra_in_converted": [],
36
+ "max_abs_diff_float32_compare": 0.0,
37
+ "max_abs_diff_key": null,
38
+ "max_rel_diff_float32_compare": 0.0,
39
+ "max_rel_diff_key": null,
40
+ "missing_count": 0,
41
+ "missing_in_converted": [],
42
+ "nonzero_diff_tensors_float32_compare": 0,
43
+ "role": "slat_generator",
44
+ "sample_nonzero_diffs": [],
45
+ "shape_mismatch_count": 0,
46
+ "shape_mismatch_sample": [],
47
+ "source": "weights/sam-3d-objects/checkpoints/slat_generator.ckpt",
48
+ "source_tensors": 1225
49
+ },
50
+ {
51
+ "compared_tensors": 74,
52
+ "converted": "weights/sam-3d-objects-mlx/checkpoints/ss_decoder.safetensors",
53
+ "converted_tensors": 74,
54
+ "dtype_pairs": {
55
+ "float16->float32": 36,
56
+ "float32->float32": 38
57
+ },
58
+ "exact_tensors_float32_compare": 74,
59
+ "extra_count": 0,
60
+ "extra_in_converted": [],
61
+ "max_abs_diff_float32_compare": 0.0,
62
+ "max_abs_diff_key": null,
63
+ "max_rel_diff_float32_compare": 0.0,
64
+ "max_rel_diff_key": null,
65
+ "missing_count": 0,
66
+ "missing_in_converted": [],
67
+ "nonzero_diff_tensors_float32_compare": 0,
68
+ "role": "ss_decoder",
69
+ "sample_nonzero_diffs": [],
70
+ "shape_mismatch_count": 0,
71
+ "shape_mismatch_sample": [],
72
+ "source": "weights/sam-3d-objects/checkpoints/ss_decoder.ckpt",
73
+ "source_tensors": 74
74
+ },
75
+ {
76
+ "compared_tensors": 101,
77
+ "converted": "weights/sam-3d-objects-mlx/checkpoints/slat_decoder_gs.safetensors",
78
+ "converted_tensors": 101,
79
+ "dtype_pairs": {
80
+ "float16->float32": 96,
81
+ "float32->float32": 5
82
+ },
83
+ "exact_tensors_float32_compare": 101,
84
+ "extra_count": 0,
85
+ "extra_in_converted": [],
86
+ "max_abs_diff_float32_compare": 0.0,
87
+ "max_abs_diff_key": null,
88
+ "max_rel_diff_float32_compare": 0.0,
89
+ "max_rel_diff_key": null,
90
+ "missing_count": 0,
91
+ "missing_in_converted": [],
92
+ "nonzero_diff_tensors_float32_compare": 0,
93
+ "role": "slat_decoder_gs",
94
+ "sample_nonzero_diffs": [],
95
+ "shape_mismatch_count": 0,
96
+ "shape_mismatch_sample": [],
97
+ "source": "weights/sam-3d-objects/checkpoints/slat_decoder_gs.ckpt",
98
+ "source_tensors": 101
99
+ },
100
+ {
101
+ "compared_tensors": 120,
102
+ "converted": "weights/sam-3d-objects-mlx/checkpoints/slat_decoder_mesh.safetensors",
103
+ "converted_tensors": 120,
104
+ "dtype_pairs": {
105
+ "float32->float32": 120
106
+ },
107
+ "exact_tensors_float32_compare": 120,
108
+ "extra_count": 0,
109
+ "extra_in_converted": [],
110
+ "max_abs_diff_float32_compare": 0.0,
111
+ "max_abs_diff_key": null,
112
+ "max_rel_diff_float32_compare": 0.0,
113
+ "max_rel_diff_key": null,
114
+ "missing_count": 0,
115
+ "missing_in_converted": [],
116
+ "nonzero_diff_tensors_float32_compare": 0,
117
+ "role": "slat_decoder_mesh",
118
+ "sample_nonzero_diffs": [],
119
+ "shape_mismatch_count": 0,
120
+ "shape_mismatch_sample": [],
121
+ "source": "weights/sam-3d-objects/checkpoints/slat_decoder_mesh.ckpt",
122
+ "source_tensors": 120
123
+ },
124
+ {
125
+ "compared_tensors": 101,
126
+ "converted": "weights/sam-3d-objects-mlx/checkpoints/slat_decoder_gs_4.safetensors",
127
+ "converted_tensors": 101,
128
+ "dtype_pairs": {
129
+ "float16->float32": 96,
130
+ "float32->float32": 5
131
+ },
132
+ "exact_tensors_float32_compare": 101,
133
+ "extra_count": 0,
134
+ "extra_in_converted": [],
135
+ "max_abs_diff_float32_compare": 0.0,
136
+ "max_abs_diff_key": null,
137
+ "max_rel_diff_float32_compare": 0.0,
138
+ "max_rel_diff_key": null,
139
+ "missing_count": 0,
140
+ "missing_in_converted": [],
141
+ "nonzero_diff_tensors_float32_compare": 0,
142
+ "role": "slat_decoder_gs_4",
143
+ "sample_nonzero_diffs": [],
144
+ "shape_mismatch_count": 0,
145
+ "shape_mismatch_sample": [],
146
+ "source": "weights/sam-3d-objects/checkpoints/slat_decoder_gs_4.ckpt",
147
+ "source_tensors": 101
148
+ }
149
+ ]