refactor(video): carry output resolution on the video result

Both runners round the requested resolution and the two-stage LTX path derives
its own, so the request never recorded what came out. Read it off the decoded
frames instead.

modules/video_models/video_run.py

@@ -29,6 +29,8 @@ class VideoResult:
     has_audio: bool
     still: bool
     processed: processing.Processed
+    width: int = 0 # what was generated, which is not what was requested whenever a runner rounds or a model picks
+    height: int = 0
 
 
 def resolve_model(engine: str | None, model: str | None) -> tuple[models_def.Model, bool]:
@@ -302,7 +304,8 @@ def run(selected: models_def.Model, *,
 
     if getattr(p, 'video_still', False):
         stills = processed.images[:1] # already trimmed in process_decode; defensive
-        return VideoResult(images=stills, video_path=None, thumb_path=None, num_frames=len(stills), fps=0.0, has_audio=False, still=True, processed=processed)
+        still_w, still_h = video_utils.pixel_size(stills, fallback=(p.width, p.height))
+        return VideoResult(images=stills, video_path=None, thumb_path=None, num_frames=len(stills), fps=0.0, has_audio=False, still=True, processed=processed, width=still_w, height=still_h)
 
     if hasattr(processed, 'images') and processed.images is not None:
         pixels = video_save.images_to_tensor(processed.images)
@@ -341,8 +344,9 @@ def run(selected: models_def.Model, *,
         mp4_interpolate=mp4_interpolate,
         metadata={},
     )
+    out_w, out_h = video_utils.pixel_size(processed.images, fallback=(p.width, p.height))
     del pixels
-    return VideoResult(images=processed.images, video_path=video_file, thumb_path=thumb_file, num_frames=num_frames, fps=float(save_fps), has_audio=waveform is not None, still=False, processed=processed)
+    return VideoResult(images=processed.images, video_path=video_file, thumb_path=thumb_file, num_frames=num_frames, fps=float(save_fps), has_audio=waveform is not None, still=False, processed=processed, width=out_w, height=out_h)
 
 
 def generate(task_id, ui_state,

modules/video_models/video_utils.py

@@ -73,6 +73,26 @@ def has_torchaudio():
         return False
 
 
+def pixel_size(pixels, fallback: tuple[int, int] = (0, 0)) -> tuple[int, int]:
+    """Width and height of decoded frames, whether they arrive as PIL images or as a tensor.
+
+    The runners request a resolution and the model answers with another one often enough that
+    the request is not a usable substitute, so read it off the pixels and keep the fallback for
+    the case where nothing was decoded at all.
+    """
+    if isinstance(pixels, list):
+        return pixels[0].size if len(pixels) > 0 and hasattr(pixels[0], 'size') else fallback
+    ndim = getattr(pixels, 'ndim', None)
+    if ndim == 5: # NCTHW
+        return int(pixels.shape[-1]), int(pixels.shape[-2])
+    if ndim == 4: # NHWC
+        return int(pixels.shape[2]), int(pixels.shape[1])
+    shape = getattr(pixels, 'shape', None)
+    if shape is not None and len(shape) >= 2:
+        return int(shape[-1]), int(shape[-2])
+    return fallback
+
+
 def classify_extension(fn: str):
     """Media kind of a filename, None when the extension is not one sdnext reads."""
     lower = str(fn).lower()