Hugging Face's logo

AlayaLab
/
WorldSculpt

English
Model card Files
xet
 Community
1
MyNiuuu commited on
Commit
924a1dc
·
verified ·
1 Parent(s): 9ddc6b0

Upload folder using huggingface_hub

Browse files
Files changed (6) hide show
  1. shape_ft1024_mv_lora_ibr_texverse_fixedmem05/ckpts/denoiser_step0015000.pt +3 -0
  2. shape_ft1024_mv_lora_ibr_texverse_fixedmem05/ckpts/mv_aggregator_step0015000.pt +3 -0
  3. shape_ft1024_mv_lora_ibr_texverse_fixedmem05/config.json +261 -0
  4. ss_ft64_mv_lora_ibr_texverse/ckpts/denoiser_step0015000.pt +3 -0
  5. ss_ft64_mv_lora_ibr_texverse/ckpts/mv_aggregator_step0015000.pt +3 -0
  6. ss_ft64_mv_lora_ibr_texverse/config.json +188 -0
shape_ft1024_mv_lora_ibr_texverse_fixedmem05/ckpts/denoiser_step0015000.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3bba4f203f4d4e3a6f2aea3f2b4cf0d6c53b14b22c8ee3b4ee3e3f71887fadfe
3
+ size 88605893
shape_ft1024_mv_lora_ibr_texverse_fixedmem05/ckpts/mv_aggregator_step0015000.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b55786caa4bd940ad12de3be791c73aba4593d41f8d8836b66f17e6032f0a241
3
+ size 134263077
shape_ft1024_mv_lora_ibr_texverse_fixedmem05/config.json ADDED
@@ -0,0 +1,261 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_comment": "Stage 2 SLat Shape Flow (1024), multi-view LoRA + ControlNet-style zero-init aggregator. Same pattern as Stage 1 MV LoRA+zeroagg, but: (a) uses MultiViewImageConditionedSLatShapeView dataset (sparse coords/feats + MV padding), (b) MultiViewLoraZeroAggImageConditionedProjSparseFlowMatchingCFGTrainer, (c) mv_aggregator channels=2048 (NAF-concat of LR+HR DINOv3 features), (d) batch_size_per_gpu=2 default since 1024-res images + sparse latents are heavy.",
3
+ "models": {
4
+ "denoiser": {
5
+ "name": "ElasticSLatFlowModel",
6
+ "args": {
7
+ "resolution": 64,
8
+ "in_channels": 32,
9
+ "out_channels": 32,
10
+ "model_channels": 1536,
11
+ "cond_channels": 1024,
12
+ "num_blocks": 30,
13
+ "num_heads": 12,
14
+ "mlp_ratio": 5.3334,
15
+ "pe_mode": "rope",
16
+ "share_mod": true,
17
+ "initialization": "scaled",
18
+ "qk_rms_norm": true,
19
+ "qk_rms_norm_cross": true,
20
+ "image_attn_mode": "proj",
21
+ "proj_in_channels": 2048
22
+ }
23
+ }
24
+ },
25
+ "dataset": {
26
+ "name": "MultiViewRandomImageConditionedSLatShapeView",
27
+ "args": {
28
+ "resolution": 1024,
29
+ "image_size": 1024,
30
+ "min_aesthetic_score": 4.5,
31
+ "max_tokens": 32768,
32
+ "num_views": 2,
33
+ "num_input_views_range": [
34
+ 1,
35
+ 20
36
+ ],
37
+ "num_views_total": 8,
38
+ "load_camera_info": true,
39
+ "skip_aesthetic_score_datasets": [
40
+ "texverse"
41
+ ],
42
+ "normalization": {
43
+ "mean": [
44
+ 0.781296,
45
+ 0.018091,
46
+ -0.495192,
47
+ -0.558457,
48
+ 1.06053,
49
+ 0.093252,
50
+ 1.518149,
51
+ -0.933218,
52
+ -0.732996,
53
+ 2.604095,
54
+ -0.118341,
55
+ -2.143904,
56
+ 0.495076,
57
+ -2.179512,
58
+ -2.130751,
59
+ -0.996944,
60
+ 0.261421,
61
+ -2.217463,
62
+ 1.260067,
63
+ -0.150213,
64
+ 3.790713,
65
+ 1.481266,
66
+ -1.046058,
67
+ -1.523667,
68
+ -0.059621,
69
+ 2.22078,
70
+ 1.621212,
71
+ 0.87723,
72
+ 0.567247,
73
+ -3.175944,
74
+ -3.186688,
75
+ 1.578665
76
+ ],
77
+ "std": [
78
+ 5.972266,
79
+ 4.706852,
80
+ 5.44501,
81
+ 5.209927,
82
+ 5.32022,
83
+ 4.547237,
84
+ 5.020802,
85
+ 5.444004,
86
+ 5.226681,
87
+ 5.683095,
88
+ 4.831436,
89
+ 5.286469,
90
+ 5.652043,
91
+ 5.367606,
92
+ 5.525084,
93
+ 4.730578,
94
+ 4.805265,
95
+ 5.124013,
96
+ 5.530808,
97
+ 5.619001,
98
+ 5.10393,
99
+ 5.41767,
100
+ 5.269677,
101
+ 5.547194,
102
+ 5.634698,
103
+ 5.235274,
104
+ 6.110351,
105
+ 5.511298,
106
+ 6.237273,
107
+ 4.879207,
108
+ 5.347008,
109
+ 5.405691
110
+ ]
111
+ },
112
+ "pretrained_slat_dec": "microsoft/TRELLIS.2-4B/ckpts/shape_dec_next_dc_f16c32_fp16",
113
+ "occlusion_cfg": {
114
+ "p_apply": 0.5,
115
+ "anneal_iters": 3000,
116
+ "max_area": [
117
+ 0,
118
+ 0.75
119
+ ],
120
+ "p_occluder3d": 0.8,
121
+ "occluder_pool": "/gs/bs/tga-koike-shanda/niumuyao/Pixal3D/datasets/TexVerse_Refine/occluder_pool",
122
+ "occluder3d": {
123
+ "scale": [
124
+ 0.5,
125
+ 0.95
126
+ ],
127
+ "rho": [
128
+ 0.4,
129
+ 0.85
130
+ ],
131
+ "jitter": 0.35,
132
+ "front_margin": 0.5,
133
+ "min_cover": 0.12,
134
+ "max_cover": [
135
+ 0.12,
136
+ 0.75
137
+ ],
138
+ "max_tries": 20,
139
+ "max_objects": 6,
140
+ "splat": 4,
141
+ "close_kernel": 5
142
+ }
143
+ },
144
+ "camera_noise_cfg": {
145
+ "p_apply": 0.5,
146
+ "anneal_iters": 3000,
147
+ "rot_deg": [
148
+ 0,
149
+ 5
150
+ ],
151
+ "trans": [
152
+ 0,
153
+ 0.05
154
+ ],
155
+ "fov": [
156
+ 0,
157
+ 0.05
158
+ ]
159
+ },
160
+ "mask_noise_cfg": {
161
+ "p_apply": 0.5,
162
+ "anneal_iters": 3000,
163
+ "ragged": [
164
+ 0,
165
+ 0.25
166
+ ],
167
+ "rim_frac": 0.012,
168
+ "p_erode": 0.0
169
+ },
170
+ "downscale_cfg": {
171
+ "p_apply": 0.5,
172
+ "anneal_iters": 3000,
173
+ "downscale": [
174
+ 0,
175
+ 0.75
176
+ ]
177
+ },
178
+ "num_test": 200
179
+ }
180
+ },
181
+ "trainer": {
182
+ "name": "MultiViewLoraZeroAggImageConditionedProjSparseFlowMatchingCFGTrainer",
183
+ "args": {
184
+ "max_steps": 20000,
185
+ "batch_size_per_gpu": 2,
186
+ "batch_split": 2,
187
+ "snapshot_batch_size": 1,
188
+ "snapshot_num_samples": 1,
189
+ "num_workers": 14,
190
+ "optimizer": {
191
+ "name": "AdamW",
192
+ "args": {
193
+ "lr": 0.0001,
194
+ "weight_decay": 0.0,
195
+ "betas": [
196
+ 0.9,
197
+ 0.95
198
+ ],
199
+ "eps": 1e-08
200
+ }
201
+ },
202
+ "ema_rate": [
203
+ 0.9999
204
+ ],
205
+ "mix_precision_mode": "amp",
206
+ "mix_precision_dtype": "bfloat16",
207
+ "elastic": {
208
+ "name": "FixedMemoryController",
209
+ "args": {
210
+ "mem_ratio": 0.5
211
+ }
212
+ },
213
+ "grad_clip": {
214
+ "name": "AdaptiveGradClipper",
215
+ "args": {
216
+ "max_norm": 1.0,
217
+ "clip_percentile": 95
218
+ }
219
+ },
220
+ "lora_config": {
221
+ "r": 32,
222
+ "lora_alpha": 32,
223
+ "target_modules": [
224
+ "to_qkv",
225
+ "to_q",
226
+ "to_kv",
227
+ "to_out",
228
+ "proj_linear"
229
+ ],
230
+ "lora_dropout": 0.0,
231
+ "bias": "none"
232
+ },
233
+ "mv_aggregator": {
234
+ "mode": "ibr",
235
+ "channels": 2048,
236
+ "chunk_size": 16384
237
+ },
238
+ "finetune_ckpt": {
239
+ "denoiser": "/gs/fs/tga-koike-shanda4/niumuyao/Pixal3D/pretrained/Pixal3D/ckpts/slat_flow_img2shape_dit_1_3B_1024_bf16.safetensors"
240
+ },
241
+ "i_print": 20,
242
+ "i_log": 5,
243
+ "i_sample": -1,
244
+ "i_save": 500,
245
+ "p_uncond": 0.1,
246
+ "sigma_min": 1e-05,
247
+ "run_projection_test": false,
248
+ "image_cond_model": {
249
+ "name": "DinoV3ProjFeatureExtractor",
250
+ "args": {
251
+ "model_name": "camenduru/dinov3-vitl16-pretrain-lvd1689m",
252
+ "image_size": 1024,
253
+ "grid_resolution": 64,
254
+ "use_naf_upsample": true,
255
+ "naf_target_size": 512
256
+ },
257
+ "image_attn_mode": "proj"
258
+ }
259
+ }
260
+ }
261
+ }
ss_ft64_mv_lora_ibr_texverse/ckpts/denoiser_step0015000.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f7ca40e5460a1e2798a295275be894b5e01461eb3529349e615a6cffd3ddffdd
3
+ size 84673669
ss_ft64_mv_lora_ibr_texverse/ckpts/mv_aggregator_step0015000.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e601159960561f553c041b9d1335378502162690cd25425505cd2e8bf1e7bb14
3
+ size 33579301
ss_ft64_mv_lora_ibr_texverse/config.json ADDED
@@ -0,0 +1,188 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_comment": "Stage 1 SS Flow (64), multi-view LoRA + ControlNet-style ZERO-INIT aggregator. The denoiser is wrapped with peft.LoraConfig (rank=32). The non-anchor views' proj features get added through a zero-init Linear (mv_aggregator) on top of the anchor view's features. At t=0 this is identical to single-view inference. The aggregator is registered as a sibling model so basic.py auto-handles DDP, optimizer, and save/load.",
3
+ "models": {
4
+ "denoiser": {
5
+ "name": "SparseStructureFlowModel",
6
+ "args": {
7
+ "resolution": 16,
8
+ "in_channels": 8,
9
+ "out_channels": 8,
10
+ "model_channels": 1536,
11
+ "cond_channels": 1024,
12
+ "num_blocks": 30,
13
+ "num_heads": 12,
14
+ "mlp_ratio": 5.3334,
15
+ "pe_mode": "rope",
16
+ "share_mod": true,
17
+ "initialization": "scaled",
18
+ "qk_rms_norm": true,
19
+ "qk_rms_norm_cross": true,
20
+ "image_attn_mode": "proj"
21
+ }
22
+ }
23
+ },
24
+ "dataset": {
25
+ "name": "MultiViewRandomImageConditionedSparseStructureLatentView",
26
+ "args": {
27
+ "min_aesthetic_score": 4.5,
28
+ "image_size": 512,
29
+ "num_views": 2,
30
+ "num_input_views_range": [
31
+ 1,
32
+ 20
33
+ ],
34
+ "num_views_total": 8,
35
+ "load_camera_info": true,
36
+ "skip_aesthetic_score_datasets": [
37
+ "texverse"
38
+ ],
39
+ "pretrained_ss_dec": "microsoft/TRELLIS-image-large/ckpts/ss_dec_conv3d_16l8_fp16",
40
+ "occlusion_cfg": {
41
+ "p_apply": 0.5,
42
+ "anneal_iters": 3000,
43
+ "max_area": [
44
+ 0,
45
+ 0.75
46
+ ],
47
+ "p_occluder3d": 0.8,
48
+ "occluder_pool": "/gs/bs/tga-koike-shanda/niumuyao/Pixal3D/datasets/TexVerse_Refine/occluder_pool",
49
+ "occluder3d": {
50
+ "scale": [
51
+ 0.5,
52
+ 0.95
53
+ ],
54
+ "rho": [
55
+ 0.4,
56
+ 0.85
57
+ ],
58
+ "jitter": 0.35,
59
+ "front_margin": 0.5,
60
+ "min_cover": 0.12,
61
+ "max_cover": [
62
+ 0.12,
63
+ 0.75
64
+ ],
65
+ "max_tries": 20,
66
+ "max_objects": 6,
67
+ "splat": 4,
68
+ "close_kernel": 5
69
+ }
70
+ },
71
+ "camera_noise_cfg": {
72
+ "p_apply": 0.5,
73
+ "anneal_iters": 3000,
74
+ "rot_deg": [
75
+ 0,
76
+ 5
77
+ ],
78
+ "trans": [
79
+ 0,
80
+ 0.05
81
+ ],
82
+ "fov": [
83
+ 0,
84
+ 0.05
85
+ ]
86
+ },
87
+ "mask_noise_cfg": {
88
+ "p_apply": 0.5,
89
+ "anneal_iters": 3000,
90
+ "ragged": [
91
+ 0,
92
+ 0.25
93
+ ],
94
+ "rim_frac": 0.012,
95
+ "p_erode": 0.0
96
+ },
97
+ "downscale_cfg": {
98
+ "p_apply": 0.5,
99
+ "anneal_iters": 3000,
100
+ "downscale": [
101
+ 0,
102
+ 0.75
103
+ ]
104
+ },
105
+ "num_test": 200
106
+ }
107
+ },
108
+ "trainer": {
109
+ "name": "MultiViewLoraZeroAggImageConditionedProjFlowMatchingCFGTrainer",
110
+ "args": {
111
+ "debug": false,
112
+ "max_steps": 20000,
113
+ "batch_size_per_gpu": 8,
114
+ "batch_split": 2,
115
+ "snapshot_batch_size": 1,
116
+ "snapshot_num_samples": 1,
117
+ "num_workers": 14,
118
+ "optimizer": {
119
+ "name": "AdamW",
120
+ "args": {
121
+ "lr": 0.0001,
122
+ "weight_decay": 0.0,
123
+ "betas": [
124
+ 0.9,
125
+ 0.95
126
+ ],
127
+ "eps": 1e-08
128
+ }
129
+ },
130
+ "ema_rate": [
131
+ 0.9999
132
+ ],
133
+ "mix_precision_mode": "amp",
134
+ "mix_precision_dtype": "bfloat16",
135
+ "grad_clip": {
136
+ "name": "AdaptiveGradClipper",
137
+ "args": {
138
+ "max_norm": 1.0,
139
+ "clip_percentile": 95
140
+ }
141
+ },
142
+ "lora_config": {
143
+ "r": 32,
144
+ "lora_alpha": 32,
145
+ "target_modules": [
146
+ "to_qkv",
147
+ "to_q",
148
+ "to_kv",
149
+ "to_out",
150
+ "proj_linear"
151
+ ],
152
+ "lora_dropout": 0.0,
153
+ "bias": "none"
154
+ },
155
+ "mv_aggregator": {
156
+ "mode": "ibr",
157
+ "channels": 1024,
158
+ "chunk_size": 16384
159
+ },
160
+ "finetune_ckpt": {
161
+ "denoiser": "/gs/fs/tga-koike-shanda4/niumuyao/Pixal3D/pretrained/Pixal3D/ckpts/ss_flow_img_dit_1_3B_64_bf16.safetensors"
162
+ },
163
+ "i_print": 20,
164
+ "i_log": 5,
165
+ "i_sample": 500,
166
+ "i_save": 500,
167
+ "p_uncond": 0.1,
168
+ "t_schedule": {
169
+ "name": "logitNormal",
170
+ "args": {
171
+ "mean": 1.0,
172
+ "std": 1.0
173
+ }
174
+ },
175
+ "sigma_min": 1e-05,
176
+ "run_projection_test": false,
177
+ "image_cond_model": {
178
+ "name": "DinoV3ProjFeatureExtractor",
179
+ "args": {
180
+ "model_name": "camenduru/dinov3-vitl16-pretrain-lvd1689m",
181
+ "image_size": 512,
182
+ "grid_resolution": 16
183
+ },
184
+ "image_attn_mode": "proj"
185
+ }
186
+ }
187
+ }
188
+ }