singam96/ShadeNet-5-35M: upload hf_draft

README.md

@@ -264,6 +264,8 @@ python infer.py photos/ \
 
 Each input produces a panel image with the albedo, depth and normal maps. A GPU is recommended: the model has 35M parameters and runs at 448px.
 
+`config.json` in this repository lists the architecture, the input and output convention (7 maps in [-1, 1]: albedo, depth, normal), the supported sizes and the recommended inference settings.
+
 ## Limitations
 
 - Input sizes: multiples of 448 only (448 is what it was trained at). Other sizes must be resized first.

config.json

@@ -0,0 +1,39 @@
+{
+  "model_name": "ShadeNet-5-35M",
+  "task": "intrinsic-decomposition",
+  "architecture": "ParallelUNet + frozen DINOv2 ViT-S/14",
+  "backbone": "facebook/dinov2-small",
+  "parameters": 34652431,
+  "trainable_parameters": 12595855,
+  "image_size": 448,
+  "supported_image_sizes": [448, 896],
+  "input": {
+    "name": "input_rgb",
+    "channels": 3,
+    "range": [-1.0, 1.0],
+    "note": "RGB photo, resized and padded to a supported size"
+  },
+  "output": {
+    "name": "maps",
+    "channels": 7,
+    "range": [-1.0, 1.0],
+    "layout": {"albedo": [0, 3], "depth": [3, 4], "normal": [4, 7]},
+    "note": "depth is relative; normals are x right, y up, z toward the viewer"
+  },
+  "checkpoint": "checkpoints/best-l1-pruned10-ep19.ckpt",
+  "onnx": {
+    "onnx/shadenet5_mix_int4_int8.onnx": "ViT fc1/fc2 int8, everything else int4",
+    "onnx/shadenet5_mix2_int4_int8.onnx": "ViT fc2 int8, everything else int4"
+  },
+  "recommended_inference": {
+    "flip_avg": true,
+    "rot_avg_deg": 4.0,
+    "unit_normals": true,
+    "forward_passes": 6,
+    "command": "python infer.py <photos> --checkpoint best-l1-pruned10-ep19.ckpt --image-size 448 --enhanced"
+  },
+  "validation_mean_l1": {
+    "single_pass": 0.1763,
+    "enhanced": 0.1628
+  }
+}