Publish camera inference in dataset layout

Author: Junchao-cs

README.md

@@ -161,6 +161,7 @@ torchrun --standalone --nproc-per-node=1 -m solarwm infer \
   --set checkpoint.path=/path/to/SolarWM-models/SolarWM-5B-sgf-stage2-81f \
   --set data.index_root=/path/to/SolarWM-Data/releases-v1 \
   --set data.transport.root=/path/to/SolarWM-Data/releases-v1 \
+  --set inference.run_id=sekai-game-fix-release \
   --set runtime.output_dir=/path/to/output
 
 # Preencode raw-WDS for the Wan2.2-5B 153f recipe.

configs/examples/wan22_ti2v_5b/infer_stage2_sgf_camera_length.yaml

@@ -96,6 +96,8 @@ checkpoint:
 inference:
   source: validation
   length: camera
+  output_layout: dataset_triplet_v1
+  run_id: wan22-ti2v-5b-stage2-sgf-camera-length-inference
 runtime:
   output_dir: /path/to/outputs/wan22-ti2v-5b-stage2-sgf-camera-length-inference
   compile_flex: false

docs/backends/wan22-ti2v-5b.md

@@ -144,18 +144,50 @@ torchrun --standalone --nproc-per-node=1 -m solarwm infer \
   --set checkpoint.path="$SOLAR_MODEL_ROOT/SolarWM-5B-sgf-stage2-81f" \
   --set data.index_root="$SOLAR_DATA_ROOT" \
   --set data.transport.root="$SOLAR_DATA_ROOT" \
+  --set inference.run_id=sekai-game-fix-release \
   --set runtime.output_dir="$SOLAR_OUTPUT_ROOT/wan5-stage2-sgf-infer"
 ```
 
 The Stage2 command loads the checkpoint directory as one model, resolves its
 weight role from `release-manifest.json` (`ema` for the published checkpoint),
 and uses the longest camera-backed horizon available for each selected test
 sample. The horizon is rounded down to complete three-latent chunks; at the
-configured 16 fps, a 960-frame camera track produces 240 latents and 957
-decoded frames. Outputs longer than 60 latents are VAE-decoded as consecutive
-60-latent tiles with one continuous temporal cache.
-Override `data.test_index` to select a different raw-WDS test index. Outputs
-are written below `runtime.output_dir/generation`.
+configured 16 fps, a 960-frame camera track produces 240 latents and 957 model
+frames. Publication repeats the final generated frame three times so the
+generated and comparison MP4s, and their camera trajectory, all contain 960
+frames. Outputs longer than 60 latents are VAE-decoded as consecutive 60-latent
+tiles with one continuous temporal cache.
+
+Camera-length inference defaults to `inference.output_layout=dataset_triplet_v1`.
+`runtime.output_dir` is the shared publication root, while `inference.run_id`
+selects a create-only provenance transaction beneath `runs/`. Every selected
+index row must provide `clip_id`. The physical dataset is `physical_generation`
+when that field is non-empty, and otherwise `dataset`; outputs are:
+
+```text
+runtime.output_dir/
+  generate/<physical_dataset>/<clip_id>.mp4
+  compare/<physical_dataset>/<clip_id>.mp4
+  camera/<physical_dataset>/<clip_id>.npy
+  runs/<run_id>/
+    generation/...
+    publication/...
+    resolved-config.json
+    launch-manifest.json
+    run-result.json
+    COMPLETE.json
+```
+
+The camera file is the source authoritative absolute C2W trajectory selected
+at the published frame timestamps and cast directly to contiguous float64 for
+compatibility. It is never the rebased or translation-transformed model camera.
+All public artifacts are create-only, and consumers must require the run-level
+`COMPLETE.json`. A run ID cannot be reused. The dataset-triplet layout currently
+requires a single shared-filesystem node; multi-node node-local output is
+rejected. Set `inference.output_layout=transaction_v1` only when the legacy
+private `runtime.output_dir/generation` transaction is explicitly desired.
+
+Override `data.test_index` to select a different raw-WDS test index.
 
 The other released checkpoints use the same command with the config, GPU
 count, checkpoint directory, and output directory shown above. The fixed

src/solarwm/backends/wan22/generation.py

@@ -8,7 +8,10 @@
 from typing import Any
 
 from solarwm.errors import BackendContractError
-from solarwm.runtime.output_layout import validation_pass_component
+from solarwm.runtime.output_layout import (
+    camera_inference_output_layout,
+    validation_pass_component,
+)
 
 
 @dataclass(frozen=True)
@@ -83,6 +86,8 @@ def resolve_generation_plan(config: Mapping[str, Any]) -> GenerationPlan:
         raise BackendContractError(
             "camera-length inference is supported only for standalone Stage2 inference"
         )
+    if camera_length:
+        camera_inference_output_layout(config)
     if action == "infer" and str(inference.get("source", "")) != "validation":
         raise BackendContractError("inference.source must be validation")
     raw_passes = validation.get("passes", [])

src/solarwm/backends/wan22/runtime/data.py

@@ -248,6 +248,7 @@ def build_camera_tokens(
         payload,
         _camera_array_key(manifest, configured_array_key),
     )
+    matrices = np.asarray(matrices, dtype=np.float32)
     declared_shape = tuple(int(value) for value in manifest["camera"]["shape"])
     if matrices.shape != declared_shape:
         raise DataContractError(