Wan2.2 TI2V 5B — video from text or an image, 24 fps, for xOllama

Wan2.2 TI2V 5B (GGUF) · umt5-xxl text encoder · Wan2.2 VAE · 640×352, 33 frames at 24 fps · OpenAI /v1/videos

Runs on xOllama, not on stock ollama. Stock ollama pulls this model but cannot show or run it: it carries only media engines. xOllama serves it through an opencoti engine with video support and its codec sidecar (dev build 2610030147001/b112 or later, set with XOLLAMA_ENGINE_PATH). The engine pinned by the current xOllama release has no media engines and refuses the model at load, naming what it lacks.

Clips from a text prompt at 24 frames a second. The model also takes a starting image (input_reference); the measurements below are text-to-video only.

Use

A clip is a job, as in OpenAI’s videos API: create it, poll it, fetch it.

xollama pull mannix/wan2.2:ti2v-5b
curl http://localhost:22434/v1/videos \
  -d '{"model": "mannix/wan2.2:ti2v-5b", "prompt": "a red fox running through fresh snow at sunrise"}'
# {"id": "video_…", "object": "video", "status": "queued", …}
curl http://localhost:22434/v1/videos/video_…            # until "status": "completed"
curl http://localhost:22434/v1/videos/video_…/content -o fox.mp4

From an image, send it as input_reference in a multipart request (not yet measured on xOllama).

The template’s clip

A request with no size gets 640×352, 33 frames at 24 fps, cfg 6, euler, flow shift 3, h264 mp4. That is also the largest clip the template serves: its 13 GiB working reserve is sized for it, so a 24 GB card holds the Q4_K_M tag. A bigger request is refused with a 400 before anything loads. The same pixel count in another shape (352×640), or a smaller size with more frames, is fine.

The model can do more on a larger card: at 832×480×33 it peaks at 29.3 GiB (Q4_K_M) and 31.2 GiB (Q8_0). On a 32 GB or larger card, raise the template’s clip and reserve with xollama tweak model.

Measured

RTX PRO 6000 Blackwell, opencoti b112, the same prompt and seed:

Every clip came back as h264 mp4, 33 frames at 24 fps.

Tags

ti2v-5b is the one for a 24 GB card. ti2v-5b-q8_0 is the most quality, and leaves less than 2 GB free on a 24 GB card.

Source

QuantStack/Wan2.2-TI2V-5B-GGUF · VAE: Comfy-Org/Wan_2.2_ComfyUI_Repackaged · text encoder: city96/umt5-xxl-encoder-gguf · License: Apache-2.0. Docs: xOllama media, video jobs.