Nodes/ComfyUI-OldTimeRadio/ VideoDirector (per-role model select)
ComfyUI Node

VideoDirector (per-role model select)

A ComfyUI node in OldTimeRadio/v2/video with 16 inputs and 1 output.

By jbrick2070·Created 5 months ago·Updated about 4 hours ago· 7
VideoDirector (per-role model select)
    • video_policy_json
    announcer_video_model
    music_video_model
    character_video_model
    announcer_image_model
    music_image_model
    character_image_model
    fps25
    canvas_w832
    canvas_h480
    seed_moderequest_hash
    request_seed0
    custom_models_json{}
    gate_in
    device_policycuda
    dtype_policyfp8_ok
    max_render_frames0
    CategoryOldTimeRadio/v2/video

    Inputs (16)

    NameTypeDefaultDescription
    announcer_video_modelCOMBOVideo model for ANNOUNCER beats (role A).
    music_video_modelCOMBOVideo model for MUSIC beats (role B).
    character_video_modelCOMBOVideo model for CHARACTER beats (face + audio, role C). Uses the saved dropdown value; profiles only retarget this when explicitly applied.
    announcer_image_modelCOMBOImage source for the announcer (feeds its video).
    music_image_modelCOMBOImage source for music beats.
    character_image_modelCOMBOImage source for character beats (kept slot).
    fpsINT251–60
    canvas_wINT83216–7680
    canvas_hINT48016–4320
    seed_modeCOMBOrequest_hashrequest_hash (deterministic) | fixed.
    request_seedINT00–4294967295Base seed (NOT named 'seed' on purpose, V-7).
    custom_models_jsonoptSTRING{}When a role is set to '+ Add Custom Model', map the role key to a custom engine id here, e.g. {"character_video_model": "my_engine"}.
    gate_inoptSTRINGOptional ordering signal (opaque STRING).
    device_policyoptCOMBOcudaEXPLICIT local video render device (platform profile value). No auto-detect.
    dtype_policyoptCOMBOfp8_okDtype lanes allowed for local video engines (fp8/fp4 artifacts are OFF on ROCm/MPS tiers).
    max_render_framesoptINT00–240Absolute per-clip RENDER-length ceiling in frames for local video engines (0 = unpinned = the engine's own maximum). Beat LENGTH is never changed by it -- this caps what an engine renders, never what the episode plays. TWO LANES CONSUME IT DIFFERENTLY. ltx_8gb takes it as a COVERAGE-PLANNING cap: the beat is partitioned into real clips of at most this many frames and chained, so the ceiling decides how many clips cover the beat. The 8GB Wan tier takes it as an adapter-side NATIVE cap: it pins 17, renders 17 real frames, and ping-pong-extends them to the beat's audio length rather than requesting the whole 177-frame span and dying in the VRAM cost model. A value at or above an engine's own maximum has no effect on that engine.

    Outputs (1)

    NameTypeDescription
    video_policy_jsonSTRING