Voxta docs

Wan 2.2

Local video with Wan 2.2: the 14B image-to-video pair, with first and end frames and loops, or the 5B text/image-to-video model.

Wan 2.2 animates a picture into a short, silent clip. Two sizes:

  • 14B image-to-video: two experts. A high-noise one lays out the motion and a low-noise one adds detail. Only one is on the GPU at a time. Needs a picture. Takes a first frame, an end frame, or both, and loops.
  • 5B text/image-to-video: one smaller model that also works from a prompt alone. No end frame.

They are paired with the UMT5-XXL text encoder and the Wan VAE.

Experimental. Only an admin can add it. The 14B fp8 models peak around 16 GB of video memory at 480p. On a 12 to 16 GB card, part of the model is kept in RAM and streamed in (slower, same result).

Setup

Add the service

Add Modules → Services → Wan 2.2 → Install, then follow the install steps. It installs on its defaults. The Models step offers the models to download now.

Pick a preset

PresetFor
Wan 2.2 Lightning (4 steps)14B with lightx2v's 4-step LoRA pair, fetched into the LoRAs folder the first time it runs.
Wan 2.2 (20 steps, CFG)14B without LoRAs.
Wan 2.2 5B (30 steps, CFG)The 5B model.

Settings

  • Model Size: 14B image-to-video (high + low noise) or 5B text/image-to-video.
  • High-Noise Model, Low-Noise Model: fp8 builds (~14.3 GB each) recommended. fp16 ones are quantized to fp8 as they load. With one left at None, the other runs every step.
  • Text Encoder: UMT5-XXL fp8 (~6.7 GB, recommended) or fp16.
  • Steps, CFG Scale, Shift, Expert Boundary (where the high-noise expert hands off. Wan uses 0.9).
  • Negative Guidance (NAG): makes the negative prompt work at CFG 1, for about 10% more time per step.
  • Resolution: 480p, or 720p (about 3x slower).
  • Duration (seconds): 1 to 8, default 5, at 16 fps (5B: 24 fps). RIFE can double the frame rate afterwards.
  • High-Noise LoRAs, Low-Noise LoRAs: one list per expert, since Wan 2.2 LoRAs ship as pairs.
  • Enable Preview: a fast low-resolution preview with the same seed, which Render full reproduces.

Loops

In the image viewer, Animate into a video offers a loop with the 14B model. The picture is both the first and the end frame, and the seam is blended so it does not show.

On this page