VideoDirector (per-role model select)

If you're running the OldTimeRadio canonical workflow, this is the node you'll eventually want to touch, and the one with the scariest-looking interface. OTR_VideoDirector is a per-role model selector: it decides which video engine renders which kind of beat, and which image engine mints the stills behind them. It doesn't render anything itself. It produces a single JSON policy, video_policy_json, that downstream nodes (OTR_ShotLock, OTR_VideoRenderBatch) honor exactly.

Here's the design that makes this node make sense: this pack is model-agnostic on purpose. There is no "primary" video model. There's a registry of engine adapters, and you pick one per role:

• announcer_video_model - the announcer bookends (role A)
• music_video_model - the opening/closing theme visuals (role B)
• character_video_model - character dialogue beats, face + audio (role C)

Each has a matching image-model dropdown (announcer_image_model, music_image_model, character_image_model) that supplies the stills the video lanes hold or move. That's six dropdowns, and they're all "registry IS the menu" - 32 video choices and 13 image choices at the time of writing, local and cloud mixed in.

The default is the thing people miss. All three video roles ship set to still_flat, a procedural still/CRT floor - no GPU video checkpoint at all. The image roles default to z_image_turbo, an Apache-2.0 model. That's deliberate: the canonical workflow runs anywhere, even a laptop, before you ask for the heavy stuff. The heavy lanes - HuMo, LTX, Wan, AnimateDiff, MiniMax H3 - are upgrades you dial in later, one role at a time.

And that's exactly what the README's VRAM table is for: it's the profile. Every engine's measured VRAM is listed per card class (8/12/16 GB), with ltx25_high_video measured at 14.48 GiB (explicitly 5080-only), h3_low_video at 7.28 GiB under an 8 GB clamp, wan22_high_video at 12.1 GiB, and so on. Read your card's column and pick from it. Two warnings the table makes you earn: the host RAM number is the one people miss (an 8 GB laptop with 16 GB of system RAM will choke regardless of VRAM), and two entries aren't commercial-clean - MiniMax H3 runs under a personal authorization that doesn't transfer to you, and the AnimateDiff motion module publishes no license grant. Fine for hobby, not cleared for resale.

The other widgets are mostly set-and-forget: fps (25), canvas_w/canvas_h (832×480), seed_mode (request_hash for deterministic renders vs fixed), and request_seed (deliberately not named "seed" - this pack is weird about that). Two worth knowing if you're on a small card: device_policy (cuda/cpu/mps, explicitly no auto-detect) and max_render_frames, a render-length ceiling that the low-VRAM Wan lane uses to render 17 real frames and ping-pong them out rather than OOMing on the full span. custom_models_json lets you map a role to your own engine id when a role is set to "+ Add Custom Model".

Install & gotchas. ComfyUI Manager → search "ComfyUI-OldTimeRadio", or git clone https://github.com/jbrick2070/ComfyUI-OldTimeRadio into custom_nodes/, check out v2.0-alpha, restart. If the node doesn't appear, you're on the wrong branch.

The big gotcha: engines load one at a time with VRAM reclaim between them, and a pick that doesn't fit your card fails loudly mid-render - it won't substitute a lighter engine for you. That's a feature, but it means the dropdown pick is a promise. Start on still_flat for everything, get one full episode through, then upgrade roles one at a time and watch the console. Cloud lanes (cloud_kling_avatar, google_veo_video, …) are paid, off by default, and only switch on when you pick them. Choose your video model like you'd choose a meal: by what's on your card, not by what's coolest.