Wiki pages Tasks, architecture, inference
Wan 2.2 model wiki
A source-checked map of the Apache-2.0 Wan2.2 family: official task checkpoints, DiT MoE and Wan-VAE figures from the paper/README, and the community inference stack (LoRA, FP8, GGUF, VRAM, async APIs). Later Wan 2.5+ native audio is documented as a product line — not as 2.2 weights.
Checked against Wan-Video/Wan2.2 README · Aug 28, 2026
Open 2.2 models T2V · I2V · TI2V · S2V · Animate
Official figures MoE, VAE, S2V, VACE, VRAM
License Apache 2.0 on 2.2 weights
Search the Wiki
17 articles
Family
Which open checkpoint exists, and what later closed Wan products added.
Family
Wan model family
Map Apache-2.0 Wan2.1 / Wan2.2 checkpoints against later closed Wan 2.5–3.0 products, and see which task each weight actually covers.
Read articleTasks
T2V, I2V, FLF2V, reference, speech, V2A, and native audio-video.
Tasks
Text-to-video
Wan2.2-T2V-A14B and TI2V-5B turn a prompt into 480P or 720P silent video. Official generate.py tasks, sizes, and VRAM floors.
Read article
Tasks
Image-to-video
I2V-A14B and TI2V-5B animate a still. Aspect follows the input image; a prompt is optional if Qwen-VL extends it.
Read article
Tasks
First–last frame video
Dedicated open FLF2V is Wan2.1-FLF2V, not a Wan2.2 checkpoint. How ComfyUI books two keyframes and what 2.2 I2V does instead.
Read article
Tasks
Reference-to-video
Identity and motion from reference images or a drive video: Wan2.1-VACE (R2V / V2V / MV2V) and Wan2.2-Animate-14B.
Read article
Tasks
Speech-to-video
Wan2.2-S2V-14B turns a portrait plus speech or song into lip-synced cinematic video. Official pipeline figures and generate.py flags.
Read article
Tasks
Video-to-audio
Wan2.2 T2V/I2V emit silent picture. Foley and scores come from a separate V2A model such as ThinkSound, then mux.
Read article
Tasks
Native audio-video
Joint picture + dialogue + SFX is a Wan 2.5+ product feature. The 2.2 open T2V/I2V weights do not generate a soundtrack.
Read articleArchitecture
DiT backbone, high/low-noise MoE experts, and Wan-VAE compression.
Architecture
Diffusion Transformer
Wan is a flow-matching DiT over 3D VAE latents, not a U-Net. Patch tokens, UMT5 cross-attention, and timestep AdaLN.
Read article
Architecture
Mixture of Experts
A14B uses two ~14B experts (27B total, 14B active). High-noise layouts early steps; low-noise refines after an SNR switch.
Read article
Architecture
Wan-VAE
Wan2.2-VAE compresses T×H×W by 4×16×16 (rate 64) and still beats several 4×8×8 VAEs on PSNR / SSIM / LPIPS.
Read articleInference
Prompt extend, LoRA, quantization, VRAM, and async job APIs.
Inference
Prompt extension
Official generate.py can expand a short prompt (or an image) with DashScope Qwen or a local Qwen2.5 / Qwen2.5-VL.
Read article
Inference
Video LoRA
Low-rank adapters on Wan DiT. MoE usually needs a high-noise + low-noise pair. Do not drop T2V LoRAs onto Animate.
Read article
Inference
Video quantization
From official dtype casts to community FP8, GGUF, distillation and cache. What each path actually changes in the graph.
Read article Inference
FP8 vs GGUF
FP8 stays a dense GPU matmul. GGUF is a block-quant weight file for low-VRAM loaders. They solve different memory problems.
Read article
Inference
Video model VRAM
Official floors: A14B ≥ 80 GB single GPU; TI2V-5B ≥ 24 GB with offload. Where T5, DiT, activations and VAE sit in the stack.
Read article
Inference
Video API async jobs
Wan sampling takes minutes. DashScope, Gradio Spaces, Replicate and fal all return a job id and poll — do not hold HTTP open.
Read articleSources
Open weights stop at Wan2.2 (plus Wan2.1 FLF2V / VACE). Wan 2.5, 2.6, 2.7 and 3.0 native audio-video are API products without public 2.2-style checkpoints.