Wan 2.2

Wan 2.2 creates short videos from a text-only prompt, or a prompt plus starting images for the first and/or last frames. Unlike LTX-2 or MiniMax H3, Wan cannot generate a soundtrack (but you can use LTX-2’s conditioning mode to add one).

License

Wan 2.2 (T2V-A14B, I2V-A14B and TI2V-5B), its text encoder and the Lightning LoRAs are released under Apache 2.0, which allows commercial use. None of the downloads is gated.

This is a summary, not legal advice. Read the full license on the model’s HuggingFace page before using a model commercially.

The Wan Models

Wan 2.2 ships three transformer variants, plus a shared text encoder and two VAEs. All share the same diffusion-style sampling but differ in size, conditioning, and intended task.

Variant	Task	Params	VAE	Conditioning
T2V-A14B	Text → Video	14B × 2 experts	A14B VAE (16-ch, 8× spatial)	Text only
I2V-A14B	Image+Text → Video	14B × 2 experts	A14B VAE (16-ch, 8× spatial)	Text + reference image (36-channel concat)
TI2V-5B	Text → Video OR Image+Text → Video	5B (single)	Wan 2.2-VAE (48-ch, 16× spatial)	Text, optionally with reference image (first-frame mask blend)

• T2V-A14B generates videos from a text prompt alone. Best motion coherence and prompt-following of the three.
• I2V-A14B locks the first frame to a reference image you supply, and also handles first-to-last interpolation and extension. Best subject-preservation.
• TI2V-5B is the small single-expert variant. It does both text-to-video and image-to-video at substantially lower VRAM, with somewhat less stable long-range coherence.

High-noise and low-noise transformers (A14B variants only)

The A14B models are a mixture-of-experts (MoE) pair — two 14B transformers per variant. The denoise loop swaps between them at a model-defined boundary timestep:

• High-noise expert runs early, when the latents are still mostly noise: composition, layout, broad motion.
• Low-noise expert runs late, when the latents are close to clean: detail and texture.

InvokeAI handles the swap automatically. In the Video panel, a single-file A14B main takes its partner through the Transformer (Low Noise) component slot; a Diffusers A14B main carries both experts internally. TI2V-5B is single-expert — no swap, no low-noise CFG, no second slot.

Lightning (the Wan fast path)

The stock A14B variants need ~40–50 denoise steps. The Lightning distillation LoRAs collapse that to 4 steps at CFG 1 with minimal quality loss — about a 10× speedup. There’s a pair per variant (one per expert); the panel’s Lightning toggle wires both automatically.

Tip

TI2V-5B has no Lightning LoRA. Its smaller size makes each step cheap; run it at 40–50 steps for a similar wall-clock to A14B + Lightning.

Installing Wan models

The Wan 2.2 starter bundle (~56 GB) installs the UMT5-XXL text encoder, both VAEs, TI2V-5B Q4_K_M, T2V-A14B Q4_K_M (high + low) with T2V Lightning (high + low), and I2V-A14B Q4_K_M (high + low) with I2V Lightning (high + low).

Higher-quality Q8_0 quantizations and full Diffusers builds are available a-la-carte in the starter models list.

The UMT5-XXL encoder also comes as a GGUF file: Wan T5 Encoder (UMT5-XXL, GGUF Q8_0) in the starter models (~6 GB, about half the Diffusers encoder), or any other quantization from city96’s umt5-xxl-encoder-gguf. It carries its own tokenizer; select it in the Wan T5 Encoder slot. Lower quantizations change the text conditioning: among the quantized builds, Q8_0 stays closest to the Diffusers encoder, while Q4_K_M deviates by about 21 % (relative L2 error, worst-token cosine 0.40). A UMT5 GGUF installed by an older InvokeAI is listed as a FLUX/SD 3 T5 encoder: use Re-identify model in the Model Manager to make it a Wan T5 encoder.

Tip

On a 12 GB VRAM card: use TI2V-5B — it covers both text-to-video and image-to-video and fits comfortably. The bundle includes it; to skip the A14B downloads, install the Wan 2.2 TI2V-5B (Q4_K_M) starter model on its own, which brings its VAE and the UMT5-XXL encoder. The A14B variants will technically run via aggressive offloading but are slow and prone to OOM.

ComfyUI int8 builds

Single-file A14B and TI2V-5B transformers in ComfyUI’s int8_tensorwise (convrot) format load like any other single file, with the same two-expert setup for A14B. The attention and feed-forward weights stay int8 in memory, which saves memory. By default they are dequantized per layer on each forward pass; int8 Compute can instead run them on NVIDIA int8 tensor cores. LoRAs, including Lightning, are applied alongside the int8 weights rather than merged into them, except that DoRA layers do not run on an int8 transformer, so a DoRA LoRA fails there when it is applied. Files marked only {"convrot": true}, without a format entry, and the int4 builds are refused when you install them.

On one Windows machine with an RTX 4090 (T2V A14B, 832x480, 17 frames), an int8 expert file is about 14.5 GB against about 28.6 GB for a bf16 expert. With both experts and the UMT5 encoder kept in the RAM cache, the int8 build peaked at about 41 GB of RAM. With max_cache_ram_gb: 16 it still runs but reloads the experts for each clip, and peaked at about 30 GB of RAM, because both experts are held in memory during the swap. Windows also counts VRAM toward a process’s committed memory, so that figure reads higher than the RAM in use. int8 mainly saves memory; how much speed it changes depends on whether the model otherwise fits in VRAM. Other systems will differ.