Over the past two years, the field of AI video generation has exploded. Proprietary systems like OpenAI’s Sora, Runway Gen-2, and Pika have shown the world what’s possible: turning a few words or an image prompt into seconds—or even minutes—of coherent video.

But the open-source community hasn’t stood still. A wave of open-source video generation models is quickly closing the gap, giving creators access to powerful tools that support text-to-video, image-to-video, and video-to-video workflows.

At SaladCloud, we’re focused on what it takes to run these models in production: how they scale across GPUs, what kind of hardware they require, and what the true cost per minute of generated video looks like. Before we dive into WAN2.2 benchmarks, let’s take a quick look at the open-source landscape.

How Open-Source Video Generation Works

The majority of open-source video generation models today are based on diffusion or flow-matching techniques—extensions of the same methods that power Stable Diffusion, adapted to generate sequences of frames.

Other families of approaches do exist (e.g., GAN-based or autoregressive methods), but diffusion and flow-matching dominate the current open-source landscape due to their scalability and image quality.

These models support a variety of creative workflows:

• Pose-to-video: Animate human motion based on pose sequences.
• Text-to-video (T2V): Generate a clip directly from a text description.
• Image-to-video (I2V): Animate a single image into motion.
• Multiple images to video: Stitch several reference frames into a continuous sequence.
• Frame interpolation: Smoothly generate in-between frames given a start and end frame.
• Video-to-video synthesis: Transform existing video into a new style or look.

Available Open-Source Models

When considering which video model to run in production, three things matter most:

1. License – can you use it commercially?
2. Parameters – how big is the model?
3. Memory footprint – how much GPU VRAM does it need?

Here’s a look at six of the most relevant open-source video models today:

Model	License	Parameters	Memory (approx)	Notes
WAN2.1	Apache 2.0	1.3B / 14B	8–60 GB	Early release; artifact-heavy, scaling issues.
WAN2.2	Apache 2.0	5B / A14B	24-80 GB	Popular, controllable, strong ecosystem.
Skyreels-V2	Permissive	14B	~42 GB	Broad adoption, rich controllability features.
LTX	Non-permissive	2-13B	~13 GB+	Efficient, comes in size variants for trade-offs.
Mochi	Apache 2.0	10B	~57 GB	Innovative, but adoption has slowed.
Nova	Apache 2.0	0.3-1.4B	~30 GB	Stability AI’s latest; balances quality & efficiency.

WAN2.2: The Best Open-Source Video Model Yet

When we covered WAN2.1, it showed the possibility of open-source video generation. WAN2.2, however, delivers what’s more practical: cinematic-quality videos generated on hardware you can afford.

Highlights of WAN2.2 vs WAN2.1

• Architectural leap – WAN2.2 introduces a Mixture-of-Experts (MoE) design, combining the efficiency of smaller models with the fidelity of massive ones. Only 14B parameters activate per step, even though the total capacity is 27B.
• Training data explosion – +65% more images, +83% more videos, and detailed aesthetic labels (lighting, composition, color), leading to cinematic realism.
• Temporal stability – eliminates frame-to-frame flicker, keeps characters consistent, and produces smoother motion.
• Camera & composition controls – built-in support for dolly, pan, crane, handheld modes, motion blur, speed ramping, safe zones, and rule-of-thirds alignment.
• Quality leap – from experimental visuals in WAN2.1 to professional-grade output in WAN2.2.
• Economic shift – WAN2.2 reduces failure rates, improves GPU efficiency, and fundamentally changes the cost-per-video economics of open-source video generation.

WAN2.2 is probably the best open-source video generation model available today, transforming AI video from experimental tech into production-ready infrastructure.

WAN2.2 comes in two primary variants:

• 5B (TI2V-5B) – a balanced model designed for consumer-grade hardware, rapid prototyping, and scalable cloud use.
• A14B – a larger architecture targeting maximum quality, best suited for datacenter or multi-GPU environments.

WAN2.2 – The 5B Model

The WAN2.2-TI2V-5B is the most approachable entry point into WAN2.2:

Key Traits of the 5B Model:

• Efficiency-first – designed to run on GPUs like the RTX 4090 or 5090 without requiring more powerful hardware.
• Balanced quality – delivers sharp visuals and smooth motion, while keeping render times manageable.
• Versatility – works across text-to-video (T2V), image-to-video (I2V), and multi-image generation pipelines.
• Production viability – at 5B parameters, it’s small enough for scale-out deployments, but large enough to achieve professional-grade results.

Benchmark Results – 5B with 1280×704 resolution

We ran the 5B model on SaladCloud, using both:

• Secure datacenter GPUs (L40S) — provisioned in blocks of 8 GPUs, which is why they deliver better economics for long video jobs.
• Consumer GPUs (RTX 4090 & RTX 5090) — great for flexible scaling and one-off runs.

All benchmarks used the official WAN2.2 repo and weights provided by WAN, with no CPU offloading or custom optimizations:

GPU Type	Resolution	Avg Gen Time (5s)	Time per 1-min Video	Cost per Min (High Priority)	Cost per Min (Batch Priority)
L40S (Secure)	1280×704	13.5 min	162 min	$2.19	$0.87
4090	1280×704	31.8 min	381.6 min	$2.42	$1.65
5090	1280×704	16.3 min	195.6 min	$1.73	$1.14

With the 5B model, you can create a full minute of 1280×704 video for under $1 using L40S Secure GPUs with batch pricing. If you need to scale to hundreds or even thousands of instances SaladCloud consumer GPU’s might be a great option as well.

Video generated using WAN2.2 5B model on rtx5090. 1280×704

“Medium shot, stainless modern kitchen, pancakes cooking while steam rises, warm backlight, subtle parallax from foreground utensils, 35mm lens”

WAN2.2 – The A14B Model

The A14B variant of WAN2.2 pushes video generation to its absolute limits. With 14B active parameters per step and a Mixture-of-Experts architecture, it produces some of the sharpest, most cinematic open-source video today.

While trying to run the model benchmark we faced a challenge: the official model requires 80+ GB of VRAM, making it impossible to be hosted on most GPUs.

To make benchmarking possible, we used the DFloat11 compressed release. This version applies Huffman coding to the exponent bits of BFloat16 weights, taking advantage of their high compressibility. The algorithm performs on-the-fly weight decompression directly on the GPU, which drastically lowers memory requirements while preserving performance.

You can learn more in DFloat11’s research paper on hardware-aware compression.

Benchmark Results – A14B

We ran the compressed A14B model on SaladCloud, across Secure datacenter GPUs: l40s and A100s.

GPU Type	Resolution	Avg Gen Time (5s)	Time per 1-min Video	Cost per Min (High Priority)	Cost per Min (Batch Priority)
L40S	1280×704	55.1 min	11.02 h	$8.93	$3.53
A100	1280×704	59.7 min	11.93 h	$11.34	$4.77
L40S	480p	14.42 min	2.88 h	$2.34	$0.92
A100	480p	14.45 min	2.89 h	$2.75	$1.16

WAN2.2 A14B sets the quality bar for open-source video generation. At 1280×704, it is still compute-heavy (10+ hours per 1 min clip), but at 480p it becomes viable for production workflows. For broadcast or cinematic projects, this is the best model up to date.

Video generated using WAN2.2 A14B model on L40S. 720×480

“A cinematic scene of a mountain climber reaching the snowy summit”

Video generated using WAN2.2 A14B model on A100. 1280×720

“A cinematic scene of a couple walking under cherry blossoms in the wind”

The Economics of WAN2.2 on SaladCloud

Beyond speed and fidelity, the real breakthrough with WAN2.2 on SaladCloud is cost efficiency.

Video generation is notoriously expensive, and for most teams, the economics have been the biggest barrier to actually using these models at scale. By running on Salad’s distributed GPU’s, you get access to both consumer GPUs (4090/5090) and Secure datacenter GPUs (L40S, A100) at prices far below traditional cloud providers which makes SaladCloud the most affordable way to produce a minute of high-quality video with WAN2.2 today.

Best-Case Costs per Minute of Finished Video

• WAN2.2 5B – as low as $0.87 per minute (L40S, batch pricing @ 1280×704)
• WAN2.2 A14B – as low as $0.92 per minute (L40S, batch pricing @ 480p)

To put that in perspective, generating the same output on a traditional GPU cloud often costs two to five times more. For anyone looking to scale beyond short clips into long-form or professional production, this cost advantage is a game-changer.

Interested in running your workload on SaladCloud Secure (H100s, A100s, L40S, and more)? Check out SaladCloud Secure.