Wan AI is Alibaba’s open video-generation model family. The public ecosystem expanded quickly from Wan 2.1 into Wan 2.2, with text-to-video, image-to-video, editing-oriented tasks, ComfyUI integration and models designed to reach consumer GPUs.
Wan is useful to follow because it shows two parallel trends in open video AI: larger models pursuing cinematic quality and smaller/hybrid models trying to make high-resolution generation more practical.
Wan 2.1
Wan 2.1 was released as an open suite of video foundation models with public inference code and weights, under the Apache 2.0 license — one of the more permissive licenses in this space, though the repository is explicit that users bear full responsibility for ensuring their own usage complies with applicable laws. The official repository covers multiple tasks including text-to-video, image-to-video and later extensions for first/last-frame generation and video creation/editing.
The public lineup includes 14B models and a smaller 1.3B text-to-video model. Alibaba’s own published evaluation claims Wan2.1 “consistently outperforms existing open-source models and state-of-the-art commercial solutions” across 14 major dimensions and 26 sub-dimensions, using 1,035 internal prompts — that’s the model authors’ own benchmark, not an independent one, and the underlying numeric scores aren’t published alongside the claim.
Consumer-GPU option
The Wan 2.1 repository states that its 1.3B text-to-video model requires about 8.19 GB of VRAM for the documented setup and targets 480P as its recommended stable resolution, with the repo’s own reference figure citing roughly 4 minutes for a 5-second 480p clip on an RTX 4090.
This is significant because many open video systems historically required high-memory GPUs. The smaller Wan checkpoint gives users a way to experiment locally without immediately moving to workstation-class hardware.
What it costs in practice
Independent testing on cloud GPU infrastructure gives a fuller picture than the single reference figure above. A benchmark run on SaladCloud (methodology: 50 unique prompts, RTX 4090 and RTX 5090 GPU pools, 480p and 720p) measured both generation time and per-minute cloud cost for the 1.3B model:
| GPU | Resolution | Time (5s clip) | Cost/min (batch) |
|---|---|---|---|
| RTX 4090 | 480p | 5.3 min | $0.24 |
| RTX 4090 | 720p | 40 min | $1.80 |
| RTX 5090 | 480p | 2.4 min | $0.15 |
| RTX 5090 | 720p | 33.9 min | $2.20 |
Note the gap between this figure (5.3 minutes at 480p on a 4090) and the reference implementation’s own ~4-minute figure — independent runs on shared cloud infrastructure routinely land a bit slower than a vendor’s own reference benchmark, which is a normal and expected kind of discrepancy rather than a red flag. The same testing found the 14B model, running with DFloat-11 compression on an enterprise L40S GPU, took roughly 16 minutes per 5-second clip at 480p and 66.3 minutes at 720p — a useful reminder that the 14B checkpoint is not a casual single-GPU proposition even with compression.
Text-to-video
Wan text-to-video models generate clips from a prompt. Larger checkpoints support higher target resolutions, while the smaller model prioritizes accessibility.
As with every video model, prompt adherence should be tested against real shot requirements: subject identity, action, camera movement, temporal stability and text rendering can behave differently across prompts.
Image-to-video
Wan also includes image-to-video checkpoints. Image conditioning makes the family useful for product videos, stylized first-frame animation and scenes where composition needs to start from an approved visual.
ComfyUI and Diffusers
Wan 2.1 was integrated into ComfyUI shortly after release and also gained Diffusers support. These integrations matter because they let the model participate in larger workflows instead of requiring users to work only through the reference script.
A ComfyUI graph can combine Wan with image preprocessing, control signals, LoRAs, upscaling and post-processing — see our ComfyUI video workflows guide for how Wan’s own VRAM-saving options (FP8, GGUF quantization, text-encoder offload) plug into a node graph in practice.
Wan 2.2
Wan 2.2 is a major upgrade to the family, also released under the Apache 2.0 license. The official project introduces a Mixture-of-Experts (MoE) architecture for video diffusion, using specialized expert models across different denoising timesteps.
The project also emphasizes expanded training data, stronger motion generation and more explicit cinematic/aesthetic control. Officially, Wan 2.2 claims superior performance against leading closed-source commercial models on an internal “Wan-Bench 2.0” suite — again, a vendor claim rather than a third-party score.
| Wan 2.2 variant | VRAM (single GPU) | Notes |
|---|---|---|
| T2V-A14B / I2V-A14B | ≥80GB | Multi-GPU (FSDP + DeepSpeed Ulysses) supported |
| TI2V-5B | ≥24GB (RTX 4090) | Memory-reducible with offload flags |
| S2V-14B | ≥80GB | Multi-GPU supported |
Wan 2.2 hybrid TI2V
One of the most practical Wan 2.2 releases is a 5B hybrid text-and-image-to-video model. The official project describes it as supporting both T2V and I2V at 720P and 24fps, with a design intended to run on consumer hardware such as an RTX 4090 at roughly 24GB VRAM per the table above.
That makes the family relevant not just for research comparisons but for actual local workflow planning.
MoE: why it matters, in concrete numbers
Mixture-of-Experts architectures activate specialized parts of a larger model for different portions of the computation. Wan 2.2’s A14B models use a two-expert design: a high-noise expert active during early denoising to establish overall layout, and a low-noise expert engaged later for detail refinement, switching between them at a signal-to-noise-ratio threshold reached during the denoising process.
The practical payoff is in the parameter math: the A14B models hold 27B total parameters, but only roughly 14B are active in any single inference step — which is where the “A14B” name comes from. The goal is to increase model capacity without paying the full compute cost of activating every parameter at every step, the same trade-off that motivates MoE designs in large language models.
Wan for cinematic generation
Wan 2.2’s training and labeling strategy puts more emphasis on lighting, composition, contrast, color tone and other visual attributes. That does not guarantee a perfect “cinematic” result from every prompt, but it explains why the release focuses strongly on aesthetics and complex motion.
Wan vs LTX
The two families now emphasize different strengths:
- Wan: broad video foundation-model suite, MoE development, high-definition hybrid generation and multiple task variants.
- LTX-2: synchronized audio-video generation, keyframe workflows, IC-LoRA controls and deep ComfyUI integration.
Wan vs HunyuanVideo
Both families provide open weights and local workflows, but checkpoint sizes, supported tasks and hardware profiles differ — HunyuanVideo-1.5’s 8.3B DiT design and its stated 14GB minimum VRAM (with offloading) sit closer to Wan’s TI2V-5B tier than to Wan’s 80GB-class A14B models. For a practical comparison, run the same prompts and source frames at comparable output settings rather than relying on vendor demos.
Benchmark claims vs independent testing
It’s worth being explicit about which numbers in this article come from where. Alibaba’s “outperforms” claims for both Wan 2.1 and Wan 2.2 are the model authors’ own internal evaluations (1,035 prompts for 2.1; the internal Wan-Bench 2.0 suite for 2.2) — neither release publishes the underlying numeric scores in its public repository, which makes them directional marketing claims rather than reproducible benchmark results. The VRAM figures and the SaladCloud timing/cost table above are independently reported, with their test conditions stated, and are a better basis for a hardware or budget decision than the vendor’s own quality claims.
What to test before choosing Wan
- Does your GPU fit the exact checkpoint?
- Do you need text-to-video, image-to-video or a specialized editing task?
- Is ComfyUI or Diffusers part of your production stack?
- What resolution and clip length are actually stable?
- What are the current license terms for the checkpoint?
- Do you need LoRA or fine-tuning support?
- What will a realistic clip actually cost you in GPU time, not just whether it fits in VRAM?
About the figures in this article
The Signal did not run the benchmarks cited above. The VRAM figures come from the official Wan 2.1 and Wan 2.2 repositories; the generation-time and cost figures come from independent SaladCloud testing with its stated methodology (50 prompts, specific GPU pools, specific resolutions); and the “outperforms” quality claims are Alibaba’s own, clearly labeled as such rather than presented as independently verified. Treat every number here as a snapshot under specific test conditions rather than a universal result.
Related The Signal coverage
- Wan 3.0: Alibaba’s New AI Video Model
- Open Source AI Video Generators
- ComfyUI Video Workflows
- LTX Video
- HunyuanVideo-1.5
- Open Source AI Image Generators
