Add GOP-aware smart rendering to video concatenation

docs/src/content/docs/Users Guide/05.Video Generation/video-workflows.mdx

@@ -128,6 +128,8 @@ This variant takes **two inputs** — the source video and a destination image:
 Because the endpoint is pinned to your image, it's ideal for bridging a clip to a known next shot or landing on a specific final frame.
 :::
 
+The **Concatenate Videos** node smart-renders compatible H.264 MP4 clips by preserving compressed video outside transition regions. Clips must share dimensions and frame rate and use compatible encoding settings. Other inputs, or a failed smart-render validation, automatically use the regular full-encoding path. Set the node's **Smart Render** input to `false` to always use full encoding.
+
 ---
 
 ## Bundled MiniMax H3 workflows

invokeai/app/services/workflow_records/default_workflows/Extend Video - LTX-2.json

@@ -458,7 +458,7 @@
       "type": "invocation",
       "data": {
         "id": "video_output",
-        "version": "1.2.0",
+        "version": "1.3.0",
         "nodePack": "invokeai",
         "label": "",
         "notes": "Joins source + extension. `transition_frames` arrives over an edge from context_frames -- the crossfade must consume exactly the held frames or they play twice.",
@@ -493,6 +493,12 @@
             "label": "",
             "description": "",
             "value": "match_first"
+          },
+          "smart_render": {
+            "name": "smart_render",
+            "label": "",
+            "description": "",
+            "value": true
           }
         },
         "isOpen": true,

invokeai/app/services/workflow_records/default_workflows/Extend Video - MiniMax H3.json

@@ -628,7 +628,7 @@
       "type": "invocation",
       "data": {
         "id": "4b32c22b-679d-4d5e-a167-c3309265fcd9",
-        "version": "1.2.0",
+        "version": "1.3.0",
         "nodePack": "invokeai",
         "label": "",
         "notes": "Joins the trimmed source and the generated continuation, cross-fading over the hand-off frame so the seam is not visible.",
@@ -657,6 +657,12 @@
             "label": "",
             "description": "",
             "value": "match_first"
+          },
+          "smart_render": {
+            "name": "smart_render",
+            "label": "",
+            "description": "",
+            "value": true
           }
         },
         "isOpen": true,

invokeai/app/services/workflow_records/default_workflows/Extend Video - Wan 2.2 Lightning w_ Concept LoRAs.json

@@ -570,7 +570,7 @@
       "type": "invocation",
       "data": {
         "id": "4b32c22b-679d-4d5e-a167-c3309265fcd9",
-        "version": "1.2.0",
+        "version": "1.3.0",
         "nodePack": "invokeai",
         "label": "",
         "notes": "Joins the source clip and the generated continuation. Size Mismatch is 'match_first', so the generated clip is resampled onto the source clip's canvas -- the model renders onto its own fixed canvas family and will rarely match your footage, and without this the join fails outright. Your source keeps its own resolution; the generated segment is the part that gets rescaled.",
@@ -615,6 +615,12 @@
             "label": "",
             "description": "",
             "value": "match_first"
+          },
+          "smart_render": {
+            "name": "smart_render",
+            "label": "",
+            "description": "",
+            "value": true
           }
         },
         "dynamicInputTemplates": {},

invokeai/app/services/workflow_records/default_workflows/Extend Video - Wan 2.2 Lightning.json

@@ -564,7 +564,7 @@
       "type": "invocation",
       "data": {
         "id": "4b32c22b-679d-4d5e-a167-c3309265fcd9",
-        "version": "1.2.0",
+        "version": "1.3.0",
         "nodePack": "invokeai",
         "label": "",
         "notes": "Joins the source clip and the generated continuation. Size Mismatch is 'match_first', so the generated clip is resampled onto the source clip's canvas -- the model renders onto its own fixed canvas family and will rarely match your footage, and without this the join fails outright. Your source keeps its own resolution; the generated segment is the part that gets rescaled.",
@@ -609,6 +609,12 @@
             "label": "",
             "description": "",
             "value": "match_first"
+          },
+          "smart_render": {
+            "name": "smart_render",
+            "label": "",
+            "description": "",
+            "value": true
           }
         },
         "isOpen": true,

invokeai/app/services/workflow_records/default_workflows/Extend Video to Image - LTX-2.json

@@ -462,7 +462,7 @@
       "type": "invocation",
       "data": {
         "id": "video_output",
-        "version": "1.2.0",
+        "version": "1.3.0",
         "nodePack": "invokeai",
         "label": "",
         "notes": "Joins source + extension. `transition_frames` arrives over an edge from context_frames -- the crossfade must consume exactly the held frames or they play twice.",
@@ -497,6 +497,12 @@
             "label": "",
             "description": "",
             "value": "match_first"
+          },
+          "smart_render": {
+            "name": "smart_render",
+            "label": "",
+            "description": "",
+            "value": true
           }
         },
         "isOpen": true,

invokeai/app/services/workflow_records/default_workflows/Extend Video to Image - Wan 2.2 Lightning.json

@@ -1062,7 +1062,7 @@
       "type": "invocation",
       "data": {
         "id": "4b32c22b-679d-4d5e-a167-c3309265fcd9",
-        "version": "1.2.0",
+        "version": "1.3.0",
         "nodePack": "invokeai",
         "label": "",
         "notes": "Joins the source clip and the generated continuation. Size Mismatch is 'match_first', so the generated clip is resampled onto the source clip's canvas -- the model renders onto its own fixed canvas family and will rarely match your footage, and without this the join fails outright. Your source keeps its own resolution; the generated segment is the part that gets rescaled.",
@@ -1107,6 +1107,12 @@
             "label": "",
             "description": "",
             "value": "match_first"
+          },
+          "smart_render": {
+            "name": "smart_render",
+            "label": "",
+            "description": "",
+            "value": true
           }
         },
         "isOpen": true,

invokeai/app/util/h264_splice.py

@@ -77,6 +77,15 @@
 _IDENTITY_MATRIX = (0x10000, 0, 0, 0, 0x10000, 0, 0, 0, 0x40000000)
 
 
+class H264AccessPoint(NamedTuple):
+    """An IDR sample and its exact position on both the packet and presentation timelines."""
+
+    sample_index: int
+    frame_index: int
+    dts: int
+    pts: int
+
+
 class H264Layout(NamedTuple):
     """What a splice needs to know about an MP4's first video track."""
 
@@ -93,6 +102,10 @@ class H264Layout(NamedTuple):
     Two tracks can share a joined track only if these bytes match."""
     frame_duration: Optional[int]
     """Every sample's duration in media ticks, or None if they are not all equal."""
+    time_scale: int
+    """The media clock rate in ticks per second."""
+    access_points: tuple[H264AccessPoint, ...]
+    """IDR points with sample and presentation indices, DTS and PTS; empty unless timing is verified CFR."""
 
 
 class SpliceError(RuntimeError):
@@ -148,7 +161,7 @@ def read_h264_layout(path: Path) -> Optional[H264Layout]:
                 fh, find_box(fh, *edts, b"elst", max_boxes=_MAX_CHILD_BOXES), stbl, movie_timescale, media_timescale
             ):
                 return None
-            return _read_video_stbl(fh, *stbl)
+            return _read_video_stbl(fh, *stbl, media_timescale)
     return None
 
 
@@ -328,7 +341,7 @@ def _is_idr(fh: BinaryIO, offset: int, size: int, length_size: int) -> bool:
     return False
 
 
-def _read_video_stbl(fh: BinaryIO, stbl_start: int, stbl_end: int) -> Optional[H264Layout]:
+def _read_video_stbl(fh: BinaryIO, stbl_start: int, stbl_end: int, time_scale: int) -> Optional[H264Layout]:
     stbl = (stbl_start, stbl_end)
     stsd = find_box(fh, *stbl, b"stsd", max_boxes=_MAX_CHILD_BOXES)
     stsz = find_box(fh, *stbl, b"stsz", max_boxes=_MAX_CHILD_BOXES)
@@ -407,16 +420,61 @@ def _read_video_stbl(fh: BinaryIO, stbl_start: int, stbl_end: int) -> Optional[H
     )
 
     runs = _read_stts(fh, stbl)
-    if runs is None:
+    if runs is None or sum(count for count, _ in runs) != sample_count or any(delta <= 0 for _, delta in runs):
         return None
     deltas = {delta for _, delta in runs}
     frame_duration = deltas.pop() if len(deltas) == 1 else None
+    access_points: tuple[H264AccessPoint, ...] = ()
+    if frame_duration is not None:
+        offsets = _read_composition_offsets(fh, stbl, sample_count)
+        if offsets is not None:
+            dts = np.arange(sample_count, dtype=np.int64) * frame_duration
+            pts = dts + offsets
+            sorted_pts = np.sort(pts)
+            if sample_count == 1 or np.all(np.diff(sorted_pts) == frame_duration):
+                access_points = tuple(
+                    H264AccessPoint(
+                        sample_index=k,
+                        frame_index=int(np.searchsorted(sorted_pts, pts[k])),
+                        dts=int(dts[k]),
+                        pts=int(pts[k]),
+                    )
+                    for k in idr_samples
+                )
     return H264Layout(
         sample_count=sample_count,
         sync_samples=idr_samples,
         decoder_config=decoder_config,
         sample_entry=sample_entry,
         frame_duration=frame_duration,
+        time_scale=time_scale,
+        access_points=access_points,
+    )
+
+
+def _read_composition_offsets(fh: BinaryIO, stbl: tuple[int, int], sample_count: int) -> Optional[np.ndarray]:
+    """Expand the bounded ``ctts`` runs, or return zero offsets when the table is absent."""
+    ctts = find_box(fh, *stbl, b"ctts", max_boxes=_MAX_CHILD_BOXES)
+    if ctts is None:
+        return np.zeros(sample_count, dtype=np.int64)
+    if ctts[0] + 8 > ctts[1]:
+        return None
+    fh.seek(ctts[0])
+    version = fh.read(1)[0]
+    if version not in (0, 1):
+        return None
+    fh.seek(ctts[0] + 4)
+    (run_count,) = struct.unpack(">I", fh.read(4))
+    if not 0 < run_count <= _MAX_SAMPLES or ctts[0] + 8 + 8 * run_count > ctts[1]:
+        return None
+    fh.seek(ctts[0] + 8)
+    fmt = ">II" if version == 0 else ">Ii"
+    runs = [struct.unpack(fmt, fh.read(8)) for _ in range(run_count)]
+    if any(count == 0 for count, _ in runs) or sum(count for count, _ in runs) != sample_count:
+        return None
+    return np.repeat(
+        np.fromiter((offset for _, offset in runs), dtype=np.int64, count=run_count),
+        np.fromiter((count for count, _ in runs), dtype=np.int64, count=run_count),
     )