Upload folder using huggingface_hub

predict.py

@@ -174,7 +174,7 @@ def predict_next_frame(ens, context_frames: np.ndarray) -> np.ndarray:
         ens.direct_cache = []
         for i in range(PRED_FRAMES):
             frame = np.transpose(predicted_np[i], (1, 2, 0))
-            frame = (frame * 255).clip(0, 255).astype(np.uint8)
+            frame = np.round(frame * 255).clip(0, 255).astype(np.uint8)
             ens.direct_cache.append(frame)
 
         result = ens.direct_cache[ens.cache_step]
@@ -191,20 +191,6 @@ def predict_next_frame(ens, context_frames: np.ndarray) -> np.ndarray:
             return result
 
         ens.reset_cache()
-
-        # Medium static detection: all frames vs first frame, max diff < 2/255
-        max_diff = 0
-        first_frame = frames[0]
-        for i in range(1, len(frames)):
-            diff = np.abs(frames[i].astype(np.float32) - first_frame.astype(np.float32)).mean()
-            max_diff = max(max_diff, diff)
-        if max_diff < 2.0 / 255.0:
-            last_ctx_uint8 = (last_frame * 255).clip(0, 255).astype(np.uint8)
-            ens.direct_cache = [last_ctx_uint8.copy() for _ in range(PRED_FRAMES)]
-            result = ens.direct_cache[ens.cache_step]
-            ens.cache_step += 1
-            return result
-
         with torch.no_grad():
             context_tensor = torch.from_numpy(context).to(DEVICE)
             last_tensor = torch.from_numpy(last_frame_t).to(DEVICE)
@@ -254,7 +240,7 @@ def predict_next_frame(ens, context_frames: np.ndarray) -> np.ndarray:
         ens.direct_cache = []
         for i in range(PRED_FRAMES):
             frame = np.transpose(predicted_np[i], (1, 2, 0))
-            frame = (frame * 255).clip(0, 255).astype(np.uint8)
+            frame = np.round(frame * 255).clip(0, 255).astype(np.uint8)
             ens.direct_cache.append(frame)
 
         result = ens.direct_cache[ens.cache_step]
@@ -286,7 +272,7 @@ def predict_next_frame(ens, context_frames: np.ndarray) -> np.ndarray:
         ens.direct_cache = []
         for i in range(PRED_FRAMES):
             frame = np.transpose(predicted_np[i], (1, 2, 0))
-            frame = (frame * 255).clip(0, 255).astype(np.uint8)
+            frame = np.round(frame * 255).clip(0, 255).astype(np.uint8)
             ens.direct_cache.append(frame)
 
         result = ens.direct_cache[ens.cache_step]