ComfyUI Node
Looking Ahead Image to Video
Text-and-image-to-video with Looking Ahead 1.3B: the start image is resized to cover the frame and center-cropped like the official inference, then the video is generated autoregressively in 3-latent-frame chunks (KV cache, 16 fps). The released models were trained on 5-second clips.
Looking Ahead Image to Video
- model
- positive
- negative
- vae
- start_image
- video
- images
- latent
- report
◄seconds5►
◄seed42►
◄steps50►
◄cfg3.0►
◄shift5.0►
◄width768►
◄height432►
◄keep_model_loadedfalse►
Categoryvideo/looking-ahead
Inputs (13)
| Name | Type | Default | Description |
|---|---|---|---|
| model | LOOKING_AHEAD_MODEL | — | |
| positive | CONDITIONING | CLIPTextEncode with a umT5-XXL (Wan) text encoder | |
| negative | CONDITIONING | The official default negative prompt is in the example workflow | |
| vae | VAE | The Wan 2.1 VAE (wan_2.1_vae.safetensors) | |
| start_image | IMAGE | — | |
| seconds | INT | 51–30 | Output length including the start frame, at 16 fps. Longer than 5 s uses the rolling KV cache with a 3-frame attention sink; quality beyond 5 s is not guaranteed upstream. |
| seed | INT | 420–18446744073709550000 | — |
| steps | INT | 501–200 | Denoising steps per chunk (official: 50) |
| cfg | FLOAT | 3.00–20 | Official: 3.0 |
| shift | FLOAT | 5.00.1–20 | Timestep shift (official: 5.0) |
| width | INT | 768256–1280 | Trained at 768x432 |
| height | INT | 432256–1280 | Trained at 768x432 |
| keep_model_loaded | BOOLEAN | false | Keep the model on the GPU after sampling (faster reruns, about 2.8 GB of VRAM) |
Outputs (4)
| Name | Type | Description |
|---|---|---|
| video | VIDEO | — |
| images | IMAGE | — |
| latent | LATENT | — |
| report | STRING | — |