12 November 2026 · VRAMfarm team

Wan 2.2 video on a rented GPU: 5B or 14B, how much VRAM and how long per clip

Should you pay for a 48 GB card to make AI video, or is a 24 GB RTX 4090 enough? With Wan 2.2 it depends on which of its two models you run. The 5B model renders a 5-second 720p clip on 24 GB in about 4 to 9 minutes. The 14B model is comfortable on a 48 GB card (on 24 GB it runs only with offloading and patience), but its clips move more naturally and follow the prompt far better. Most people who are disappointed with Wan 2.2 have only seen 5B.

5B and 14B are different models

The names suggest a small and a large version of the same thing. They are built differently.

The 5B model compresses video very aggressively before it works on it, which is why it fits a consumer card and renders fast. The price is detail and motion: hands, faces at a distance and fast movement suffer first. The 14B pair splits each clip between a “high noise” expert that lays out the scene and the motion in the early steps and a “low noise” expert that adds detail at the end. Only one expert has to be on the GPU at a time, so you get the quality of a much larger model at roughly the memory cost of one 14B model.

A practical way to use both: try ideas on 5B, which is cheap and quick, then render the clips you keep on 14B. Don’t judge Wan 2.2 by 5B alone.

Resolution and length: what the models can do

Both models are trained for clips of about 5 seconds. You can ask for more frames (frame counts go in steps of four plus one: 81, 97, 121…), but quality falls off beyond the trained length. Motion starts to loop or drift, and memory grows with every frame. Stretching 14B from 81 to 121 frames gives it 50% more frames to hold and process: several more gigabytes, and well over 50% more time, because attention cost grows faster than the frame count.

For longer scenes, people chain clips instead. Render a clip, take its last frame and feed it to the image-to-video model as the start of the next one. With careful prompts you get 10–20 seconds that hang together, with a visible seam now and then.

Resolution works the same way. 720p is the top of what either model was trained for. 1080p is possible only with an upscaler afterwards, which ComfyUI can run as an extra step. At the other end, 480p is your friend while testing: it is two to three times faster, and a good seed and prompt usually carry over when you re-render at 720p.

What neither model does: sound, readable text in the frame, or several consistent characters across separate clips without help (a LoRA or a reference image).

Minutes per clip on each card

Video time grows with frames times pixels times steps. The 14B app includes 4-step speed LoRAs (lightx2v), which cut the time several times compared with the usual 20–30 steps, at some cost in motion and fine detail.

These are planning ranges from public reports and the cards’ compute, not our own benchmarks. The first clip of a session is slower because the models are loaded into memory. The developers’ own figure for 5B is under 9 minutes on a 24 GB consumer card, which matches the 4090 row.

Which card to rent

For 5B, an RTX 4090 is the default. An RTX 5090 is about 1.3–1.5 times faster and has 32 GB, so it handles longer clips without offloading. The RTX 3090 holds the same 24 GB and runs 5B fine, about half as fast; it is the cheap choice for an overnight queue of drafts.

For 14B, take a 48 GB card or larger:

• L40S: the usual choice. Native fp8 and enough compute for 4–8 minutes per clip.
• RTX A6000: same 48 GB, cheaper per hour, but Ampere has no native fp8, so it is roughly half as fast. Good when nobody is waiting.
• RTX PRO 6000 (96 GB): fast, and enough memory for longer clips or 720p without the speed LoRA.
• H100: the fastest of the common options for full-step renders. Its price per hour is the highest, so it pays off when a series has to be finished today.

A 32 GB RTX 5090 can run 14B in fp8 at 720p for a standard 5-second clip, set up by hand in an empty ComfyUI, with little room to spare. Our ready-made 14B app does not offer it, because longer clips or an extra LoRA push it over the edge. The GPU guide for ComfyUI has the wider comparison.

If your clip runs out of memory, the CUDA out of memory guide lists what to lower first: frames, then resolution, then the batch.

Using the ready-made video app

Choose the video generation app at launch, with 5B or 14B. The app is free; you pay for the GPU time and the traffic of the download. After the server is ready, ComfyUI installs itself and the models arrive in the background: about 17 GB for 5B, about 70 GB for 14B, which can take 10–15 minutes. Those minutes are billed. If you open ComfyUI too early you’ll see a 502 page or “missing model” errors; the dashboard shows when the app is ready.

Then:

1. Open Workflow → Browse Templates → Video and pick the Wan 2.2 template that matches your app: 5B, or 14B text-to-video, or 14B image-to-video. The 14B app has both pairs and their 4-step LoRAs, so either template runs as it is.
2. Write the prompt as a short shot description: subject, action, camera, light. “A woman in a red coat walks towards the camera on a wet street at night, slow dolly back, neon reflections” works better than a list of adjectives.
3. Set width, height and length. Start at 480p for tests.
4. Press Run. Clips land in /workspace/ComfyUI/output as video files; download them from Jupyter or straight from the node.

For image-to-video, upload your photo in the Load Image node. The photo fixes the look and composition; the prompt describes what moves. A photo with the subject roughly where you want it at the end of the clip works best.

The base models of the app are not kept in the saved folder: they download again on the next launch. Your own LoRAs, workflows and outputs in /workspace are kept if you use Save and stop. More in the ComfyUI guide.

What a clip costs

The per-clip cost is the hourly price × minutes per clip ÷ 60, using the middle of each range:

You don’t pay per clip, though, you pay per session. A realistic one for a 30-second ad of six shots on an L40S with the 14B app: about 15 minutes for the download, 20 draft clips at 480p at about 2.5 minutes each, then 8 final renders at 720p (two retakes) at about 6 minutes each. That is roughly 1 h 55 min of GPU time, about €2.01, plus 70 GB of traffic at the machine’s rate.

Two habits keep the bill down. Write and refine prompts on 480p or on 5B, and render 720p only for clips you mean to keep. And queue several clips before you walk away, then stop the server when the queue is empty: billing stops that minute. The pricing page has the current rate of each card, and the cost guide works through images and LLM tokens the same way.