diff --git a/tools/compare_fl2va_steps.py b/tools/compare_fl2va_steps.py
index 1379a4f..56dee76 100644
--- a/tools/compare_fl2va_steps.py
+++ b/tools/compare_fl2va_steps.py
@@ -21,8 +21,12 @@ text = H3TokenRefiner(checkpoint)(Qwen3VLPromptConditioner("/text-encoders/qwen3
 model = H3PackedDenoiser.from_checkpoint(checkpoint, attention_backend="sage2").eval()
 packer = H3PromptPacker(checkpoint)
 
-video, audio_carried = initial["initial_x"]
-video, audio_carried = video.to("cuda"), audio_carried.to("cuda")
+packed_initial = initial["initial_x"].to("cuda").reshape(-1)
+video_shape = (1, 24, 7, 12, 20)
+audio_shape = (1, 32, 2, 37)
+video_count = torch.tensor(video_shape).prod().item()
+video = packed_initial[:video_count].reshape(video_shape)
+audio_carried = packed_initial[video_count:].reshape(audio_shape)
 old_video = old_audio = old_sigma = None
 for index, reference in enumerate(steps):
     sigma, sigma_down = sigmas[index], sigmas[index + 1]
@@ -34,11 +38,13 @@ for index, reference in enumerate(steps):
     carry = _audio_sigma(sigma) / sigma
     velocity_audio = (1.0 - 4.0) * (audio_carried * carry) + (1.0 + 3.0 * _audio_sigma(sigma)) * velocity_audio
     denoised = (video - sigma * velocity_video, audio_carried - sigma * velocity_audio)
-    reference_denoised = reference["denoised"]
-    denoised_delta = (denoised[0].float() - reference_denoised[0].float()).abs()
+    reference_denoised = reference["denoised"].to("cuda").reshape(-1)
+    reference_denoised_video = reference_denoised[:video_count].reshape(video_shape)
+    denoised_delta = (denoised[0].float() - reference_denoised_video.float()).abs()
     previous_sigma = sigmas[index - 1] if index else None
     video = res_multistep_update(video, denoised[0], sigma, sigma_down, old_video, old_sigma, previous_sigma)
     audio_carried = res_multistep_update(audio_carried, denoised[1], sigma, sigma_down, old_audio, old_sigma, previous_sigma)
-    latent_delta = (video.float() - reference["x"][0].float()).abs()
+    reference_latent = reference["x"].to("cuda").reshape(-1)[:video_count].reshape(video_shape)
+    latent_delta = (video.float() - reference_latent.float()).abs()
     print(f"step={index:02d} x0_video_mean={denoised_delta.mean().item():.6g} x0_video_max={denoised_delta.max().item():.6g} latent_video_mean={latent_delta.mean().item():.6g} latent_video_max={latent_delta.max().item():.6g}")
     old_video, old_audio, old_sigma = denoised[0], denoised[1], sigma_down
diff --git a/tools/inspect_fl2va_initial.py b/tools/inspect_fl2va_initial.py
new file mode 100644
index 0000000..675d4f5
--- /dev/null
+++ b/tools/inspect_fl2va_initial.py
@@ -0,0 +1,11 @@
+import torch
+
+
+state = torch.load("/artifacts/fl2va-sampler-reference/initial.pt", map_location="cpu", weights_only=False)
+for name, value in state.items():
+    print(name, type(value), getattr(value, "shape", None))
+    if hasattr(value, "unbind"):
+        try:
+            print("unbind", [(type(item), getattr(item, "shape", None)) for item in value.unbind()])
+        except Exception as error:
+            print("unbind error", error)