wan22-ti2v-5b-fast-int8

A Core AI bundle of FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers for the Aether SDK (iOS and macOS 27+).

• Source: FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers at revision 3e187042a324f6f5fb68fd22110a78725253de8f, licence Apache-2.0. The licence is included as LICENSE. FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers declares apache-2.0 in its model card but ships no licence file; LICENSE is the canonical text from apache.org.
• Changes from the source: converted from PyTorch to Core AI (.aimodel) by Aether forge (recipe wan22-ti2v-5b-fast-int8@1). Weights are int8-linear-perchannel (8-bit weights). The stages run one after another: the text encoder as 3 sequential segments (these weights, its token embedding float16), the first-frame encoder (the video autoencoder's encoder, float16), the denoiser as 3 sequential segments (these weights, float32 residual stream) and the video decoder (the tiny decoder: madebyollin's taeHV with lightx2v's taew2_2 weights, float16, never quantized). The model's own decoder is not included: clips decode with the tiny decoder. The tokenizer files are the source's own.
• Output: video clips continuing a given first frame, guided by a prompt (videoOutput), 3 steps without guidance; 49 frames (2.0 s) at 24 fps by default, at most 49.

Variants

Variant	Platform	Arch	Compute	Compiled	Assets	Download
ios-any-gpu	ios	any	gpu	no (specialized on first load)	wan22_ti2v_5b_fast_int8-denoiser0.aimodel 2.11 GB, wan22_ti2v_5b_fast_int8-denoiser1.aimodel 2.02 GB, wan22_ti2v_5b_fast_int8-denoiser2.aimodel 2.02 GB, wan22_ti2v_5b_fast_int8-taedecoder.aimodel 19.9 MB, wan22_ti2v_5b_fast_int8-text0.aimodel 2.1 GB, wan22_ti2v_5b_fast_int8-text1.aimodel 2.32 GB, wan22_ti2v_5b_fast_int8-text2.aimodel 2.32 GB, wan22_ti2v_5b_fast_int8-vaeenc.aimodel 107.8 MB	13.02 GB

Clip sizes (the first frame must be the clip's size):

• ios-any-gpu: 384×672 (default), 672×384, 320×576, 576×320

ios-any-gpu needs the com.apple.developer.kernel.increased-memory-limit entitlement: its T2 run (17-frame fixture clips) peaked at 2024 MB on iPhone18,2 (record b23fa5c1). Its stages load one at a time.

On the device ios-any-gpu needs about 37.75 GB of storage: the 13.02 GB download plus about 24.73 GB that Core AI caches as it specializes the assets on first use of each size (measured on iPhone18,2 after clips at 384×672, 672×384, 320×576, 576×320, record b23fa5c1; verification/storage/ios-any-gpu.json).

Third-party code

The tiny decoder (taeDecoder) is converted from madebyollin's taeHV (MIT licence, the module carried unmodified) with the taew2_2 weights (Apache-2.0). Every variant carries taeHV's licence as THIRD_PARTY_LICENSES/taehv-LICENSE.

Verification

Every row is a record in verification/ about exactly these bytes (matched by bundle digest). Reference rows are strict T2 passes of the unquantized export on the same fixture, in verification/reference/.

Variant	Tier	Result	Detail	Device	OS build	Compute	Record
ios-any-gpu	T2	pass	6/6 cases; profile quantized-8bit-relative; fixture 7052e35d5589c9dc	iPhone18,2	24A446	target	b23fa5c1
unquantized reference (not published)	T2	pass	6/6 cases; profile strict-relative; fixture 7052e35d5589c9dc	Mac17,6	26A434	target	fef1c5e9

The quantized profile also requires: T2 strict on the unquantized reference export (met by the reference row).

Use

aether generate wan22-ti2v-5b-fast-int8 --video --image first.png --prompt "The cat turns its head toward the sea" -o clip.mp4

import Aether
import ImageIO

let aether = try Aether()
let generator = try await aether.videoGenerator("wan22-ti2v-5b-fast-int8")
let source = CGImageSourceCreateWithURL(URL(fileURLWithPath: "first.png") as CFURL, nil)!
var request = VideoRequest(prompt: "The cat turns its head toward the sea")
request.firstFrame = CGImageSourceCreateImageAtIndex(source, 0, nil)
let clip = try await generator.video(for: request)
try await clip.write(to: URL(fileURLWithPath: "clip.mp4"))