| { |
| "format": 1, |
| "base_model": { |
| "repo": "TencentARC/Pixal3D", |
| "revision": "0b31f9160aa400719af409098bff7936a932f726" |
| }, |
| "key_convention": "Each key is 'backbone.model.<path>.lora_A.weight' or 'backbone.model.<path>.lora_B.weight', where <path> names an nn.Linear in the state dict of the base flow model given by base_submodel (minus its '.weight' suffix); lora_A has shape [rank, in_features], lora_B has shape [out_features, rank], and the layer's weight update is (alpha / rank) * lora_B @ lora_A; example: 'backbone.model.blocks.0.self_attn.to_qkv.lora_A.weight'.", |
| "fold_rule": { |
| "none": "The adapters stay separate in float32: y = base(x) + (lora_B(lora_A(x.float())) * (alpha / rank)).to(base(x).dtype).", |
| "bf16": "At load time each adapted weight W (bfloat16, the dtype of the base model's transformer blocks) is replaced by W + ((lora_B.float() @ lora_A.float()) * (alpha / rank)).to(bfloat16), the sum being taken in bfloat16, and the adapters are removed." |
| }, |
| "students": { |
| "structure": { |
| "file": "structure_student.safetensors", |
| "sha256": "a938130271c96eaaf67643bb428aa2aa8e53033b09608ea3eedc5531d9b97396", |
| "size_bytes": 1274085008, |
| "base_submodel": "ckpts/ss_flow_img_dit_1_3B_64_bf16", |
| "pipeline_key": "sparse_structure_flow_model", |
| "lora": { |
| "rank": 256, |
| "alpha": 256, |
| "target_modules": [ |
| "self_attn.to_qkv", |
| "self_attn.to_out", |
| "cross_attn.cross_attn_block.to_q", |
| "cross_attn.cross_attn_block.to_kv", |
| "cross_attn.cross_attn_block.to_out", |
| "cross_attn.proj_linear", |
| "mlp.mlp.0", |
| "mlp.mlp.2" |
| ], |
| "num_blocks": 30, |
| "num_adapters": 240, |
| "num_tensors": 480, |
| "num_elements": 318504960 |
| }, |
| "steps": { |
| "num_steps": 4, |
| "t": [ |
| 1.0, |
| 0.9, |
| 0.75, |
| 0.6, |
| 0.0 |
| ], |
| "grid_indices": [ |
| 0, |
| 3, |
| 6, |
| 8, |
| 12 |
| ], |
| "teacher_grid": "t = linspace(1, 0, 13) warped by t' = 3t / (1 + 2t); grid_indices index this grid", |
| "update": "euler", |
| "guidance_strength": 1.0 |
| }, |
| "fold": "none", |
| "dtype": "float32", |
| "source_checkpoint_sha256": "be04198a3751c999343510611e627ff92a696b02bdeb83ec8ea93c966663c1f8", |
| "training_updates": 500 |
| }, |
| "coarse_shape": { |
| "file": "coarse_shape_student.safetensors", |
| "sha256": "59be306522816b4d1defa525cfecf3f9666a161902b245fee387cd7a34fb9a66", |
| "size_bytes": 1305542328, |
| "base_submodel": "ckpts/slat_flow_img2shape_dit_1_3B_512_bf16", |
| "pipeline_key": "shape_slat_flow_model_512", |
| "lora": { |
| "rank": 256, |
| "alpha": 256, |
| "target_modules": [ |
| "self_attn.to_qkv", |
| "self_attn.to_out", |
| "cross_attn.cross_attn_block.to_q", |
| "cross_attn.cross_attn_block.to_kv", |
| "cross_attn.cross_attn_block.to_out", |
| "cross_attn.proj_linear", |
| "mlp.mlp.0", |
| "mlp.mlp.2" |
| ], |
| "num_blocks": 30, |
| "num_adapters": 240, |
| "num_tensors": 480, |
| "num_elements": 326369280 |
| }, |
| "steps": { |
| "num_steps": 1, |
| "t": [ |
| 1.0, |
| 0.0 |
| ], |
| "update": "x0_prediction", |
| "guidance_strength": 1.0, |
| "x0_formula": "x0 = (1 - s) * x_t - (s + (1 - s) * t) * v with s = 1e-5, in float64" |
| }, |
| "fold": "none", |
| "dtype": "float32", |
| "source_checkpoint_sha256": "46bc526388cc55f226d5747e64d7cd8a32fbe9153fdd0ff67eac7695ed148e1f", |
| "training_updates": 3000 |
| }, |
| "fine_shape": { |
| "file": "fine_shape_student.safetensors", |
| "sha256": "6beb37c82f5f618f84d9bc988564d41e3971ff85c42ef06217e580186e7dd583", |
| "size_bytes": 1305542328, |
| "base_submodel": "ckpts/slat_flow_img2shape_dit_1_3B_1024_bf16", |
| "pipeline_key": "shape_slat_flow_model_1024", |
| "lora": { |
| "rank": 256, |
| "alpha": 256, |
| "target_modules": [ |
| "self_attn.to_qkv", |
| "self_attn.to_out", |
| "cross_attn.cross_attn_block.to_q", |
| "cross_attn.cross_attn_block.to_kv", |
| "cross_attn.cross_attn_block.to_out", |
| "cross_attn.proj_linear", |
| "mlp.mlp.0", |
| "mlp.mlp.2" |
| ], |
| "num_blocks": 30, |
| "num_adapters": 240, |
| "num_tensors": 480, |
| "num_elements": 326369280 |
| }, |
| "steps": { |
| "num_steps": 1, |
| "t": [ |
| 1.0, |
| 0.0 |
| ], |
| "update": "x0_prediction", |
| "guidance_strength": 1.0, |
| "x0_formula": "x0 = (1 - s) * x_t - (s + (1 - s) * t) * v with s = 1e-5, in float64" |
| }, |
| "fold": "bf16", |
| "dtype": "float32", |
| "source_checkpoint_sha256": "9fa67dd330e2dcfaf751e281d19ed8cc3e9c949c97a1a0d0d1a5940bafda070c", |
| "training_updates": 2000 |
| }, |
| "texture": { |
| "file": "texture_student.safetensors", |
| "sha256": "ecc4ae096358120cde9ef2a6bbdff8bfee5c04e5f3bc49db84183d0f7d37887c", |
| "size_bytes": 1305542328, |
| "base_submodel": "ckpts/slat_flow_imgshape2tex_dit_1_3B_1024_bf16", |
| "pipeline_key": "tex_slat_flow_model_1024", |
| "lora": { |
| "rank": 256, |
| "alpha": 256, |
| "target_modules": [ |
| "self_attn.to_qkv", |
| "self_attn.to_out", |
| "cross_attn.cross_attn_block.to_q", |
| "cross_attn.cross_attn_block.to_kv", |
| "cross_attn.cross_attn_block.to_out", |
| "cross_attn.proj_linear", |
| "mlp.mlp.0", |
| "mlp.mlp.2" |
| ], |
| "num_blocks": 30, |
| "num_adapters": 240, |
| "num_tensors": 480, |
| "num_elements": 326369280 |
| }, |
| "steps": { |
| "num_steps": 2, |
| "k": 2, |
| "m": 6, |
| "t": [ |
| 1.0, |
| 0.75, |
| 0.0 |
| ], |
| "grid_indices": [ |
| 0, |
| 6, |
| 12 |
| ], |
| "teacher_grid": "t = linspace(1, 0, 13) warped by t' = 3t / (1 + 2t); grid_indices index this grid", |
| "update": "euler", |
| "guidance_strength": 1.0 |
| }, |
| "fold": "bf16", |
| "dtype": "float32", |
| "source_checkpoint_sha256": "4d8c96e111239c08670fe7cf88391ac024e2aa70b952e9f3714149c36ccb970f", |
| "training_updates": 3000 |
| } |
| } |
| } |
|
|
