Video Render Batch (A-S7.5 in-process render)

OTR_VideoRenderBatch is where the OldTimeRadio pipeline stops planning and starts burning GPU time. Everything upstream - the script, the voices, the frozen audio master, the per-role video policy from OTR_VideoDirector, the shot plan from OTR_ShotLock, the episode stills - has already happened. This node walks the engine registry and renders one real per-beat clip per shot, then hands the finished clips downstream to OTR_SilentComposite for assembly. It's the workhorse, and it's why this pack needs a GPU at all if you dial past the procedural floor.

The one widget that matters is mode. episode is the production path - render a real episode from patched_ledger_json, the ShotLock-planned, image-genned ledger. soak and single are diagnostic harness modes for the author's own QA rig; most users will never touch them, and the widgets that look like they matter there (beats, oom_index, frame_count, engine) are only read in those modes. In episode mode they're inert - the beat count and frame budget come from the planned ledger, not your widgets. So when you're looking at this node in a real workflow, set mode to episode and leave the rest alone.

Wiring, for when you're building the graph by hand rather than loading the canonical workflow:

• patched_ledger_json - the ShotLock-planned ledger. Non-negotiable in episode mode.
• master_audio_path - path to the frozen master mix from OTR_EpisodeAssembler.output_path. Beats that lack their own per-line audio get it sliced from this file, read-only, by ffmpeg. The master is never mutated. Leave it unset and the node degrades loudly on missing audio.
• image_done - an ordering gate from OTR_ImageGenDispatcher.image_done. Consuming that opaque token forces the image phase (every episode still on disk) to finish before video rendering starts. It's never parsed - just a handshake.

It returns render_report_json (a structured report of what each engine delivered) and clip_manifest_json (the beat-ordered manifest that feeds OTR_SilentComposite). It's an output node, so it can sit at the end of a render-gate workflow with nothing wired downstream and still run.

The mechanism is the interesting part. This node must run inside ComfyUI's executor thread (a normal /prompt queue) rather than a background route - only there does ComfyUI's model_management evict the umt5/whisper encoders between encode and sample, which is what keeps the heavy in-process forward under the VRAM ceiling. It holds a single resident engine at a time, and any per-beat detach/reclaim happens inside the selected engine wrapper. The frozen audio section stays read-only the whole time. That "one engine resident, explicit reclaim" discipline is the same philosophy as the rest of the pack: no surprises, and when an engine OOMs or a model is missing, it stops with a named error rather than quietly swapping in something else.

Two troubleshooting notes worth keeping. First, install via ComfyUI Manager ("ComfyUI-OldTimeRadio") or clone from GitHub into custom_nodes/, check out v2.0-alpha, restart - then place the Stable Audio / voice / image weights yourself, because nothing auto-downloads and a missing weight fails this node loudly. Second, if you're experimenting: mode=soak with oom_index set will inject a simulated OOM at a chosen beat index to test the fallback trail - it's a genuinely nice diagnostic, but it's for understanding failure behavior, not for making episodes. And know the pack's honest known limitation going in: occasionally an episode hands a line to the wrong character, and nothing fixes that. The render still completes; it just isn't perfect.