diff --git a/docs/how-to/write-a-converter.md b/docs/how-to/write-a-converter.md
index 51ddaa1b..87f67a77 100644
--- a/docs/how-to/write-a-converter.md
+++ b/docs/how-to/write-a-converter.md
@@ -10,7 +10,7 @@ Non-camera channels have no HFlow-specific payload contract. Their topic, schema
 
 Camera handling is selected by exact schema name. A custom schema name that happens to contain images is treated as an ordinary pass-through channel, so it will not appear in `Episode.cameras` and camera checks will not inspect it. The accepted camera inputs are:
 
-- `sensor_msgs/msg/CompressedImage` or `foxglove.CompressedImage`: the message encoding and schema must be decodable by the installed ROS 2 or protobuf MCAP decoder. Every message on one channel must use the same image format, and that format string must contain `jpeg`, `jpg`, or `png`. The payloads must contain valid images of that format. For a channel with at least two messages, the median difference between consecutive log times must be positive so HFlow can infer a frame rate. HFlow accepts a one-message channel using a 1 fps fallback and drops an empty camera channel with a warning. See [`_decode_compressed_images`](../../src/hflow/transform.py), [`_input_codec_for_image_format`](../../src/hflow/transform.py), and [`estimate_fps_from_log_times`](../../src/hflow/video.py).
+- `sensor_msgs/msg/CompressedImage` or `foxglove.CompressedImage`: the message encoding and schema must be decodable by the installed ROS 2 or protobuf MCAP decoder. Every message on one channel must use the same image format, and that format string must contain `jpeg`, `jpg`, or `png`. The payloads must contain valid images of that format. For a channel with at least two messages, the median difference between consecutive log times must be positive so HFlow can infer a frame rate. HFlow accepts a one-message channel using a 1 fps fallback and preserves an empty declared camera channel as an empty canonical video channel. See [`_decode_compressed_images`](../../src/hflow/transform.py), [`_input_codec_for_image_format`](../../src/hflow/transform.py), and [`estimate_fps_from_log_times`](../../src/hflow/video.py).
 - `foxglove.CompressedVideo` or `foxglove_msgs/msg/CompressedVideo`: the message encoding and schema must be decodable. Each message must declare `format="h264"` and contain one Annex B access unit. Every keyframe must include SPS and PPS, and the first message on a channel must be a keyframe. HFlow inserts a missing access-unit delimiter without re-encoding when the message can also be serialized by its protobuf or ROS 2 CDR encoder; it rejects other stream repairs. The current validator does not inspect whether an otherwise accepted stream uses B-frames, so a converter passing through H.264 must still disable them to satisfy the canonical output convention. See [`_resolve_decoder`, `_resolve_encoder`, and `_validate_passthrough_video_payload`](../../src/hflow/transform.py).
 
 Raw `sensor_msgs/msg/Image` and `foxglove.RawImage` channels are not supported. Record compressed JPEG or PNG images instead. Any derived topic configured by the pipeline must also be absent from the source because HFlow refuses to create a second channel with that topic; this is normally a pipeline-author concern rather than converter logic. These refusals are in [`write_canonical_episode`](../../src/hflow/transform.py).
diff --git a/src/hflow/transform.py b/src/hflow/transform.py
index 06e09ac7..a77f8f03 100644
--- a/src/hflow/transform.py
+++ b/src/hflow/transform.py
@@ -756,7 +756,9 @@ def group_for(info: TopicInfo) -> str:
             payloads = camera_payloads[channel_id]
             topic = infos[channel_id].topic
             if not payloads:
-                logger.warning("camera topic %r has no messages; dropping it", topic)
+                # Preserve the declared channel below even when there are no
+                # messages. Channel declarations are source information and
+                # participate in the canonical content identity (#528).
                 continue
             images, frame_ids, input_codec = _decode_compressed_images(
                 topic, infos[channel_id], payloads
@@ -895,8 +897,6 @@ def output_topic(message: _OutgoingMessage) -> str:
             for source_channel_id in sorted(infos, key=lambda cid: (infos[cid].topic, cid)):
                 info = infos[source_channel_id]
                 if source_channel_id in camera_payloads:
-                    if not camera_payloads[source_channel_id]:
-                        continue
                     if video_schema_id is None:
                         video_schema_id = writer.register_schema(
                             name=CANONICAL_VIDEO_SCHEMA_NAME,
diff --git a/tests/test_transform.py b/tests/test_transform.py
index 492f6eee..0a93dc5c 100644
--- a/tests/test_transform.py
+++ b/tests/test_transform.py
@@ -133,6 +133,63 @@ def test_camera_channels_became_compressed_video(
     assert joint_channel.message_encoding == "cdr"
 
 
+def _write_state_source(path: Path, *, include_empty_camera: bool) -> Path:
+    from mcap.writer import Writer as StockWriter
+
+    with path.open("wb") as stream:
+        writer = StockWriter(stream)
+        writer.start(profile="", library="test")
+        if include_empty_camera:
+            camera_schema_id = writer.register_schema(
+                name="sensor_msgs/msg/CompressedImage",
+                encoding="ros2msg",
+                data=b"",
+            )
+            writer.register_channel(
+                topic="/camera/empty",
+                message_encoding="cdr",
+                schema_id=camera_schema_id,
+            )
+        state_schema_id = writer.register_schema(name="state", encoding="jsonschema", data=b"{}")
+        state_channel_id = writer.register_channel(
+            topic="/state",
+            message_encoding="json",
+            schema_id=state_schema_id,
+        )
+        writer.add_message(
+            state_channel_id,
+            log_time=1_000_000_000,
+            publish_time=1_000_000_000,
+            data=b'{"value": 1}',
+        )
+        writer.finish()
+    return path
+
+
+def test_empty_camera_declaration_survives_transform_and_changes_identity(tmp_path: Path) -> None:
+    with_camera = _write_state_source(tmp_path / "with-camera.mcap", include_empty_camera=True)
+    without_camera = _write_state_source(tmp_path / "without-camera.mcap", include_empty_camera=False)
+    canonical_with = tmp_path / "with-camera.canonical.mcap"
+    canonical_without = tmp_path / "without-camera.canonical.mcap"
+
+    write_canonical_episode(with_camera, canonical_with)
+    write_canonical_episode(without_camera, canonical_without)
+
+    with canonical_with.open("rb") as stream:
+        summary = make_reader(stream).get_summary()
+        assert summary is not None
+        by_topic = {
+            channel.topic: (channel, summary.schemas[channel.schema_id])
+            for channel in summary.channels.values()
+        }
+
+    camera_channel, camera_schema = by_topic["/camera/empty"]
+    assert camera_schema.name == CANONICAL_VIDEO_SCHEMA_NAME
+    assert camera_schema.encoding == "protobuf"
+    assert camera_channel.message_encoding == "protobuf"
+    assert canonical_with.read_bytes() != canonical_without.read_bytes()
+
+
 def test_state_messages_pass_through_byte_for_byte(
     source_episode: Path, canonical_episode: Path
 ) -> None: