Add weights

README.md

@@ -0,0 +1,32 @@
+---
+license: mit
+pipeline_tag: image-to-3d
+tags:
+  - 3d-generation
+  - part-aware
+  - trellis
+---
+
+# KaiNinja
+
+Weights of KaiNinja, a part-aware image-to-3D generator built on TRELLIS.2.
+
+| File | Stage | Model | Weights |
+|---|---|---|---|
+| `stage1_layout_flow.safetensors` | Stage 1, dual-volume sparse-structure flow | `TriSparseStructureFlowModel` | step 75k |
+| `stage2_slat_flow.safetensors` | Stage 2, dual-stream structured-latent flow | `TriSLatSeparateFlowModel` | EMA (0.9999), step 100k |
+
+The matching `*.json` files hold the model constructor arguments. The shape decoder is the original
+TRELLIS.2 one and is not included here.
+
+## Inference settings
+
+- Stage 1: Euler flow sampler, 50 steps, classifier-free guidance 3.0; the two decoded volumes go
+  through overlap deduplication before Stage 2.
+- Stage 2: Euler flow sampler, 50 steps, classifier-free guidance 3.0.
+- Each volume is decoded with the TRELLIS.2 shape decoder; parts are the connected components within
+  each volume.
+
+## License
+
+MIT, as the base model.

stage1_layout_flow.json

@@ -0,0 +1,28 @@
+{
+  "name": "TriSparseStructureFlowModel",
+  "args": {
+    "resolution": 16,
+    "channels_per_stream": 8,
+    "num_streams": 2,
+    "model_channels": 1536,
+    "cond_channels": 1024,
+    "num_blocks": 30,
+    "num_heads": 12,
+    "mlp_ratio": 5.3334,
+    "pe_mode": "rope",
+    "share_mod": true,
+    "qk_rms_norm": true,
+    "qk_rms_norm_cross": true,
+    "dtype": "bfloat16",
+    "weight_share": "separate",
+    "cross_volume_at": [
+      6,
+      12,
+      18,
+      24,
+      29
+    ],
+    "per_stream_io": true,
+    "rope_stream_offset": 16.0
+  }
+}

stage1_layout_flow.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e1d430ce4677d709be4a3035ef401c8413a3224a1cea00455273b91aa95737a1
+size 5733917816

stage2_slat_flow.json

@@ -0,0 +1,29 @@
+{
+  "name": "TriSLatSeparateFlowModel",
+  "args": {
+    "resolution": 32,
+    "in_channels": 32,
+    "out_channels": 32,
+    "model_channels": 1536,
+    "cond_channels": 1024,
+    "num_blocks": 30,
+    "num_heads": 12,
+    "mlp_ratio": 5.3334,
+    "share_mod": true,
+    "qk_rms_norm": true,
+    "qk_rms_norm_cross": true,
+    "global_attn_at": [
+      2,
+      5,
+      8,
+      11,
+      14,
+      17,
+      20,
+      23,
+      26,
+      29
+    ],
+    "num_streams": 2
+  }
+}

stage2_slat_flow.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cb252ad8997ecacaef7f20e667f9bab38ef19dd014d1199c0e9ea272991c6946
+size 2619732720