ComfyUI Node
Pulse Render
Renders a Pulse Slate timeline window by window, writing each segment to disk as it completes. Requeue after a crash and only the windows that are missing render; edit one shot and only its window re-renders.
Pulse Render
- timeline
- model
- vae
- audio_vae
- model_fl2va
- video
- frames
- audio
- segment_paths
- report
◄schema_version3.0.0►
◄cache_modeauto►
◄run_dirpulseslate►
◄run_id►
◄save_segmentstrue►
◄low_memorytrue►
◄dry_runfalse►
◄prune_unusedfalse►
◄use_reference_audiofalse►
◄seam_treatmentaudio+colour►
CategoryAddisPulse/H3
Inputs (15)
| Name | Type | Default | Description |
|---|---|---|---|
| timeline | PULSE_TIMELINE | From a Pulse Slate node. Carries the compiled plan, the window seeds and the reference digests the segment cache keys on. | |
| model | MODEL | The ref2va checkpoint, with every attention and memory patch already applied upstream. This node samples with the model it is handed, so a patch applied downstream would do nothing. | |
| vae | VAE | — | |
| audio_vae | VAE | — | |
| schema_version | STRING | 3.0.0 | Which widget layout this node was saved with. Written by the node, read at load time to restore values by name. Do not edit. |
| cache_mode | COMBO | auto | 'auto' reuses any window whose content, seed and patch chain are unchanged. 'force_rerender' ignores the cache. 'reuse_only' refuses to render anything and aborts if a window is missing -- for assembling a final cut without regenerating a frame. |
| run_dir | STRING | pulseslate | Folder under ComfyUI/output that holds run folders. |
| run_id | STRING | Which run folder to resume into. Empty derives it from the timeline, so the same project reopens into the same folder across sessions -- and a seed change still reuses that folder. | |
| save_segments | BOOLEAN | true | Write each window to disk as it finishes. This is what makes a killed render resumable. Turning it off gives up the cache. |
| low_memory | BOOLEAN | true | Accumulate assembled frames as 8-bit and release VRAM between windows. The finished video is assembled by joining the segment files, so with this on a twelve-window film never exists in RAM. |
| dry_run | BOOLEAN | false | Produce the report and nothing else: no sampling, no decode, no file writes. A wrong reference binding renders successfully and gives you the wrong film -- this is how you catch it first. |
| prune_unused | BOOLEAN | false | Delete segments in this run folder that the current timeline no longer references. Off by default: yesterday's segments are what make flipping back to yesterday's edit free. |
| use_reference_audio | BOOLEAN | false | Put the lip_sync reference recordings into the finished film instead of the audio H3 generated. H3 always synthesises its own track, and on a lip-sync shot that track is a re-synthesis of your recording -- close, but not your take. Off by default because it silences everything the model scored around the voice; the generated audio is still written to each segment's .flac either way, so this is reversible without re-rendering. |
| seam_treatment | COMBO | audio+colour | What to do where two windows meet. The container seam is already gapless -- this is about the two sides being independent generations: a score that restarts, a level that steps, a grade that drifts. 'audio' matches the opening level to the previous tail and puts a 20-50ms equal-power dip across the join. Sample count is preserved exactly, so nothing drifts against the picture. 'audio+colour' also matches the entrance frame's per-channel mean and spread to the previous exit frame, decaying over the first frames so the correction does not simply move the cut to the next seam. 'off' is there so you can A/B it -- colour matching occasionally makes things worse. |
| model_fl2vaopt | MODEL | The First/Last-Frame checkpoint. Needed by the continuity modes that pin a frame. |
Outputs (5)
| Name | Type | Description |
|---|---|---|
| video | VIDEO | — |
| frames | IMAGE | — |
| audio | AUDIO | — |
| segment_paths | STRING | — |
| report | STRING | — |