Expose per-beat H3 clip outputs

dumas_h3_longvideos.py

@@ -5448,7 +5448,7 @@ def _evict_all_but(keep_model):
 
 class H3LongVideos:
     CATEGORY = "Dumas/MiniMax"
-    FUNCTION = "run"
+    FUNCTION = "run"
     # fps is emitted as BOTH types on purpose: ComfyUI does not coerce between them,
     # and the nodes that want a frame rate are split -- CreateVideo / SaveWEBM /
     # VHS Video Combine take a FLOAT, while plenty of utility nodes take an INT.
@@ -5458,11 +5458,13 @@ class H3LongVideos:
     # output. Inserting mid-list would silently re-target them.
     # APPEND to these, never insert. A workflow stores an output link by SLOT INDEX,
     # so a new type in the middle silently re-points every link after it.
-    RETURN_TYPES = ("IMAGE", "AUDIO", "STRING", "STRING", "INT", "INT", "INT", "FLOAT", "FLOAT", "INT",
-                    "LATENT", "STRING")
-    RETURN_NAMES = ("images", "audio", "info", "script", "frames_per_shot", "total_frames",
-                    "shots", "video_seconds", "fps", "fps_int",
-                    "latent", "soundscape")
+    RETURN_TYPES = ("IMAGE", "AUDIO", "STRING", "STRING", "INT", "INT", "INT", "FLOAT", "FLOAT", "INT",
+                    "LATENT", "STRING", "IMAGE", "AUDIO")
+    RETURN_NAMES = ("images", "audio", "info", "script", "frames_per_shot", "total_frames",
+                    "shots", "video_seconds", "fps", "fps_int",
+                    "latent", "soundscape", "beat_images", "beat_audio")
+    OUTPUT_IS_LIST = (False, False, False, False, False, False, False, False, False, False,
+                      False, False, True, True)
 
     @classmethod
     def IS_CHANGED(cls, plan_only=False, **kwargs):
@@ -6413,9 +6415,9 @@ class H3LongVideos:
             # correctly-SHAPED empty one rather than None: a downstream LATENT input
             # would choke on None, and this keeps the preview wireable exactly like
             # a real run.
-            return (ph_img, ph_audio, plan, "\n---\n".join(gens), max(plan_lens),
-                    sum(plan_lens), shots, total, float(fps), int(fps),
-                    _empty_av_latent(w, h, 5, fps)[0], global_soundscape)
+            return (ph_img, ph_audio, plan, "\n---\n".join(gens), max(plan_lens),
+                    sum(plan_lens), shots, total, float(fps), int(fps),
+                    _empty_av_latent(w, h, 5, fps)[0], global_soundscape, [], [])
 
         spk = speech_flags(beats)          # which shots have real (quoted) dialogue
         vram_trace = []                    # free VRAM after each shot
@@ -6841,9 +6843,10 @@ class H3LongVideos:
         # reassigns it, so this is the derived bed when auto_soundscape fired and
         # your own text when it did not. Emitting it means you can read what was
         # generated, and feed it straight back into the widget-input to pin it.
-        return (all_frames, {"waveform": all_audio, "sample_rate": sr}, info, script,
-                max(shot_lens), all_frames.shape[0], len(gens), round(actual, 2),
-                float(fps), int(fps), latent_out, global_soundscape)
+        return (all_frames, {"waveform": all_audio, "sample_rate": sr}, info, script,
+                max(shot_lens), all_frames.shape[0], len(gens), round(actual, 2),
+                float(fps), int(fps), latent_out, global_soundscape,
+                video_chunks, [{"waveform": chunk, "sample_rate": sr} for chunk in audio_chunks])
 
 
 # The old FL2VA / REF2VA aliases were only alternate menu entries for the same class.