Wan 2.2 TI2V 5B — iOS bundle

A pre-flighted bundle of Wan 2.2 TI2V 5B + UMT5-XXL (text encoder) + the Wan 2.2 high-compression VAE, quantized to fit the top tier of Apple hardware and run via Mirage — the on-device diffusion engine for iOS / macOS / visionOS.

This is video generation with no server. One model does both jobs: text→video and image→video (animate a still photo) from the same weights. Wan 2.2 TI2V 5B is Alibaba's dense 5B video diffusion transformer paired with a 16×16×4 high-compression VAE — the compression ratio that makes phone-side denoising and decode tractable at all.

What's inside

Total bundle size: ~8.5 GB. Peak memory at generation time depends heavily on resolution × frame count; the VAE decode of the full frame stack is the spike — enable tiled decode on memory-constrained devices.

Quick start (Mirage)

import Mirage

let docs = FileManager.default.urls(for: .documentDirectory, in: .userDomainMask)[0]

let engine = try Engine(models: ModelFiles(
    diffusionModel: docs.appendingPathComponent("Wan2.2-TI2V-5B-Q4_K_M.gguf"),
    vae:            docs.appendingPathComponent("wan2.2_vae.safetensors"),
    t5Encoder:      docs.appendingPathComponent("umt5-xxl-encoder-Q4_K_M.gguf")
))

// Text → video
let video = try await engine.generateVideo(.init(
    prompt: "a golden retriever puppy running through shallow ocean waves at sunset, slow motion, cinematic",
    width: 480, height: 832,   // Wan's native 480p — off-distribution sizes degrade hard
    frames: 33,          // must be 4n+1 (temporal compression is 4×)
    steps: 20,
    cfgScale: 6.0,
    flowShift: 3.0
))
// → ordered RGB frames; fps is an encode-time choice —
//   write them to .mp4 with AVAssetWriter (16 fps is Wan's native rate)

// Image → video: pass an init image and the same call animates it.

Frame-count rule

The VAE compresses time 4×, so frames must be 4n + 1: 13, 17, 21, 25, 29, 33, … A 33-frame clip at 16 fps is ~2 seconds of video. Passing an even count will be rounded by the engine.

Prompting guide

Wan conditions on UMT5-XXL, a multilingual instruction-grade encoder — like all modern DiTs it rewards full-sentence, cinematography-flavored prompts over tag soup:

• Order matters: subject → motion → camera → scene. The model weighs the start of the prompt most heavily, so lead with what matters most.
• Name the motion. Video models allocate capacity to change-over-time. "waves rolling in, hair blowing in the wind, camera slowly dollying right" beats a static scene description.
• Lead with the subject, follow with the camera. "A red fox stalking through snow, tracking shot, shallow depth of field."
• One scene per clip. 2-5 seconds is a shot, not a film. Prompts describing scene cuts produce mush.
• English and Chinese both work — Wan was trained bilingually, and the strongest negative prompt in the ecosystem is the Chinese default below.

The upstream default negative prompt (recommended as-is):

色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走

(It reads: garish tones, overexposure, static image, blurry details, subtitles, watermark-style artifacts, gray cast, worst quality, JPEG residue, malformed anatomy, fused fingers, still frame, cluttered background, walking backwards — the classic video-model failure modes.)

Performance (measured via stable-diffusion.cpp on an M2 Mac, 24 GB)

Phase timings for a 480×832 clip (Wan's native 480p geometry), 13 frames, 20 steps — diffusion on Metal with flash attention, text encoder + VAE on CPU with tiled decode (the memory-safe mobile configuration):

End-to-end: roughly 25 minutes per ~0.8 s clip on an M2. This is a patience feature, not an instant one — the point is that it runs at all with zero server involvement. The VAE decode dominates; a TAE-style tiny decoder (taehv) is the known path to cutting that to seconds and is the first planned improvement.

Resolution is a hard constraint, not a dial. Off-distribution sizes (e.g. 480×320) sample recognizable subjects wrapped in heavy artifacts. Stick to the trained 480p geometry: 480×832 portrait / 832×480 landscape, 13-33 frames, 15-25 steps.

Memory: untiled decode at 480×832×13 allocates ~20.7 GB — enable tiled decode everywhere. Default 32-tiles measured ~12.7 GB (seam-free output); smaller tiles cut it further for phone-class ceilings.

Gate availability on physical RAM, not device model:

ProcessInfo.processInfo.physicalMemory >= 12 * 1024 * 1024 * 1024

Sample output

A real 13-frame clip from this exact Q4 bundle on an M2 Mac — 480×832, 20 steps, tiled CPU decode:

Prompt: "a golden retriever puppy running through shallow ocean waves at sunset, splashing water, slow motion, cinematic lighting"

And single-frame generation (Wan does text-to-image too — frames: 1), 480×320, 20 steps:

Prompt: "a golden retriever puppy sitting on a beach at sunset, photorealistic"

Why this bundle exists

The official Wan 2.2 release is PyTorch + Diffusers — brilliant on an H100, unrunnable on a phone. Getting it on-device requires:

1. An engine that speaks GGUF video DiTs — stable-diffusion.cpp grew vid_gen support for the Wan family in 2025, and it runs on ggml-metal
2. The matching parts: TI2V 5B uses its own high-compression VAE (wan2.2_vae, not the Wan 2.1 VAE every other Wan model uses) and the UMT5-XXL encoder — mixing these up produces noise
3. Quants sized for a 12 GB memory ceiling rather than a 24 GB GPU

Picking those apart from three upstream repos takes effort. This bundle packages the working combination once, verified end-to-end on Apple Silicon.

Provenance

License

Apache 2.0 throughout — bundle, documentation, and all upstream weights. Safe for commercial, on-device use.

Built by

Haplo · @jc_builds · Mirage on GitHub