ComfyUI Node
MonarchRT Generate (Self-Forcing T2V)
480x832, 81 frames at 16 fps. Runs the upstream Self-Forcing / MonarchRT pipeline in the runtime, returns real VIDEO outputs and a JSON report.
MonarchRT Generate (Self-Forcing T2V)
- runtime
- video
- report
◄prompt►
◄seed0►
◄attentionmonarch_h2►
◄videos1►
CategoryMonarchRT
Inputs (5)
| Name | Type | Default | Description |
|---|---|---|---|
| runtime | MONARCHRT_RUNTIME | — | |
| prompt | STRING | Used as written (no prompt rewriting). | |
| seed | INT | 00–2147483647 | — |
| attention | COMBO | monarch_h2 | monarch_h2: MonarchRT, training-free: Monarch attention on the public dense Self-Forcing weights, h_reduce=2 (about 90% effective attention sparsity; the training-free setting discussed in upstream issue #2). | monarch_h1: MonarchRT, training-free, h_reduce=1 (the upstream config default, about 95% effective sparsity; the paper pairs this with trained weights, so expect lower quality without them). | dense: Baseline: the same Self-Forcing weights with dense attention (flash-attn). Use it to compare against MonarchRT. |
| videos | INT | 11–4 | Videos in this job (seeds seed, seed+1, ...), generated one after another by the same runtime process. |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| video | VIDEO | — |
| report | STRING | — |