ComfyUI Node
vlo Time-to-Move (TTM)
Patches a video model to follow the motion in a reference clip, as in Time-to-Move (https://github.com/time-to-move/TTM). The reference latents seed the sampler's starting latent, which is what carries the intended motion, and the masked region is then held to the reference for the opening steps via Comfy's own inpaint path, so it works with any sampler. Patch only the model of the sampler that starts the schedule, and leave that sampler's start_at_step at 0. Drives the denoise mask, so any noise_mask already on the sampled latent is replaced.
vlo Time-to-Move (TTM)
- model
- reference_latents
- mask
- MODEL
◄start_step1►
◄end_step2►
Categoryadvanced/model
Inputs (5)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| reference_latents | LATENT | Encoded reference video, e.g. the cut-and-drag clip. Must match the resolution and frame count being sampled. This replaces whatever latent is wired into the sampler. | |
| mask | MASK | White marks the region held to the reference; black is left free for the model to generate. For a moving subject that means white background and a black hole over the subject. Pixel resolution and frame count are matched to the latent grid automatically. | |
| start_step | INT | 10–1000 | Step whose noise level the reference is seeded at. The sampler skips the steps before it. Higher values leave less noise on the reference, binding the result more tightly to it at the cost of a step and of the model's freedom to clean up paste artifacts. 0 is a no-op: sigma is 1.0 there and the reference washes out entirely. |
| end_step | INT | 20–1000 | The step at which the region stops being held to the reference and starts denoising freely. Exclusive, and counted the same way the TTM reference implementation counts it. Set at or below start_step to seed the init only and never hold the region at all. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |