SilentComposite (render -> one always-silent video)

The name is the contract: OTR_SilentComposite takes everything the render stage produced and flattens it into one always-silent canonical video - bt709 color, yuv420p, constant frame rate, audio stripped. It's the node between the raw per-beat clips and the captions/credits/mux stages, and it exists to guarantee the rest of the pipeline always receives a single, normal, silent MP4 no matter how many engines and aspects contributed clips.

Audio is stripped here on purpose (V-1): the frozen master gets added only at the very end by OTR_MasterAudioMux. If you're hunting for a bug where the final video's audio is wrong, remember this node's output is supposed to be silent.

How it works

Two paths, one node. With just base_video_path (the OTR_SignalLostVideo radio-floor mp4) it's the single-base floor path: normalize that one video to the canvas. When clip_manifest_json from OTRVideoRenderBatch (mode=episode) is wired and carries per-beat clips, it assembles a frame-accurate CFR timeline: each beat is conformed to its audio-derived frame count, and gaps are filled from base_video_path. That's how a portrait HuMo clip, a landscape LTX clip, and the procedural floor end up as one continuous, correctly-timed video instead of a mash-up of mismatched resolutions.

Canvas defaults are canvas_w 1472 × canvas_h 832 at 25 fps. output_path defaults to <output>/otr/episodes/<stem>_silent.mp4. gate_in is the ordering signal.

The interesting optional is upscale_engine (default off, or spandrel_esrgan): a post-render upscale stage that runs inside the composite per-clip (sharpen branch only) - so a still- or low-res clip gets lifted before assembly. It uses spandrel 0.4.x (a pinned pack dependency), and upscale_device (cpu default, or cuda/cuda:N) picks where the model loads. The composite itself always runs on CPU - that's ffmpeg - so the upscale is the only part that can touch a GPU. Off is byte-identical to the classic lanczos+unsharp path, so you can A/B freely.

Outputs: silent_video_path and report.

Install and gotchas

Pack-level install (ComfyUI Manager → ComfyUI-OldTimeRadio, restart; v2.0-alpha). Needs ffmpeg on PATH. The gotcha to respect is the audio-strip invariant: if you wire in a base video hoping its audio survives, it won't - and that's correct. And if you enable spandrel_esrgan on a card with a loaded video engine, the upscale runs after, one clip at a time with VRAM reclaimed between - don't be surprised by the pause. Keep upscale_device off/cpu unless you're on a machine with headroom and want the speed.