Make AI Video on an Old 4 GB Graphics Card: Running Wan 2.2 on a GTX 1050 Ti, and the Rule That Keeps It Honest
Most advice for running AI video on a weak computer is advice about what to give up. Use the fast version of the model. Skip steps. Shrink the picture. Accept the mush.
A small project published on November 13, 2026 takes the opposite position, and states it as a rule: "a speed-up that changes the picture is not a speed-up." Its author got Wan 2.2, a free and commercially usable video model from Alibaba's Wan team, running at its official settings on a GTX 1050 Ti with 4 GB of memory, the kind of card that shipped in gaming laptops in 2017. A two-second clip takes about 26 minutes. A five-second clip takes about two and a half hours.
Those numbers sound like a joke until you change the question. Not "can I iterate on this machine?" You cannot. The question is "can this machine render finals while I sleep?" It can, and the recipe proves that the result matches what the unmodified model would have produced.
What the recipe is
The project is called wan22-pascal-4gb, by a developer who goes by alexturkey. It has three parts:
- a tested set of settings for running the 5-billion-parameter Wan 2.2 model that does text-to-video and image-to-video,
- four small patches for stable-diffusion.cpp, a free C++ program by leejet that runs image and video models without the usual Python setup,
- a ready-made Windows build, so you do not need to compile anything.
The output is 832 by 480 at 24 frames per second, either two seconds (49 frames) or five seconds (121 frames). On the author's laptop the card never used more than about 2.7 GB of its 4 GB. The program itself needed about 7.1 GB of regular system memory at five seconds, so close your other heavy apps.
The author is upfront that others got Wan 2.2 onto 4 GB cards first, usually with the distilled "turbo" version that runs in four steps. What this project adds is the full model at the official 50 steps for five-second clips, with every speed-up checked frame by frame.
The rule, and why a creator should care
Here is how the check works, without the engineering. Every finished frame is a PNG file. You can compute a fingerprint of a file, a hash, that changes completely if even one pixel changes. The author rendered a reference clip with the untouched program, then added each speed-up one at a time and compared fingerprints. A patch stayed only if every frame's fingerprint matched exactly.
Four patches passed. Together they cut a two-second render from about 32.6 minutes to about 26.4, and the five-second render by about 13.7 percent, with identical pixels.
The interesting part is what failed. The project's record of everything tried is the most useful document in the repo for anyone making video:
- Skipping steps that seem small saved up to a third of the sampling time. The overall image score looked fine. But the worst region, the fox's fur, lost detail, and on longer clips the face went blurry with ghosting from motion. The author's note: "A softened image can read as 'cleaner' at a glance. It is detail loss."
- The turbo model, per the README, gave "broken, abstract textures" with this version of the engine.
- Fewer steps, a lighter decoder and bigger decoding tiles either damaged the picture or did not fit in memory.
That first point is worth carrying into every tool you use, not just this one. A shortcut that softens fur and faces can look like a cleaner render on a phone screen. You only see the loss on a big monitor or a client's TV, after you have delivered.
The prompt lesson that matters most on a slow machine
When one render costs you most of an evening, a prompt that misses is expensive. The project's prompt notes include a small story that every video creator should read. The short prompt, a red fox walking through snow at golden hour, produced a clip where the fox was in frame "only for the last half second." The rewrite said where the subject sits, which way it faces and what must stay sharp. The author's lesson: "say where the subject is in the frame, what moves, and what must not block the view."
That is shot description, not keyword stuffing. It is how you would brief a camera operator. On fast hardware you can brute-force your way past a vague prompt by rerolling. On a 4 GB card you have to write it like a shot list.
Two smaller craft notes from the same file. The recipe uses Wan's official negative prompt, which is written in Chinese, because the model was tuned with it. And the file must be saved as plain UTF-8 without a byte-order mark, because a stray invisible character at the start "silently changes the prompt."
Put this into practice
This is the routine I would follow with any old NVIDIA card around.
- Draft somewhere fast first. Search results point to a hosted Wan 2.2 5B demo on Hugging Face, and many paid services run the same model. Use those, or a friend's newer card, to find a prompt and composition you like. Do not iterate on the old laptop.
- Download the two zips from the project's v0.1.0 release: the engine and NVIDIA's runtime files. Check the SHA-256 fingerprints listed in the README before unpacking. Install Microsoft's Visual C++ runtime if Windows asks for it.
- Put the models on an SSD. They come to about 10.9 GB. On a spinning hard drive, the author's first step took 329 seconds by itself.
- Run the 17-frame check (
--frames 17). It takes about ten minutes and tells you whether your setup works before you commit hours. - Run the fox test and compare fingerprints. With the default fox prompt and seed 42,
scripts\verify_frames.pychecks your frames against the author's. If they match, your machine reproduces the recipe exactly. If not, you know before you spend a night. - Render your shot at two seconds (
--frames 49). Write the prompt to a UTF-8 file and pass it with--prompt-file. - Run five-second finals overnight, detached, one job at a time, with the browser closed, because the README warns that a browser starves the memory and closing the console window kills the job.
- Keep a log of prompt, seed and frame count for every final. When a render takes hours, your notes are worth more than the render.
Where it falls short
It is slow, honestly slow. About 159 seconds per step at five seconds, so about 2.6 hours per clip. This is a final-render machine, not a sketchbook.
It was measured on one setup. One prompt, one seed, one GTX 1050 Ti, Windows, CUDA 12.8. The author says other GTX 10-series cards were not tried, Linux and Mac were not measured, and the bigger 14-billion-parameter Wan models were not tested. Other cards or drivers can change the fingerprints even when the picture looks the same, so rerun the check.
Step counts depend on clip length. At 121 frames, 20 steps produced a dot-grid pattern across the picture, so five-second clips need the full 50. Other prompts may need different settings; the author calls the recipe "a tested point, not a law."
It tracks one engine version. The patches are built for one specific stable-diffusion.cpp release, and newer releases may need the patches redone.
Resolution tops out at 832 by 480. A smaller 512 by 288 test came out oversaturated with color banding. If you need 720p or more, this is a draft and preview tool, not a delivery tool, unless you upscale afterward.
The license is good, with the usual caveat. The patches and scripts are MIT. The Wan 2.2 5B model files are Apache 2.0, which allows commercial use, and the model card says the Wan team claims no rights over your generated videos as long as you follow its usage terms. Who owns AI output in your country is a separate question, and the author says so too. The README also asks you to label generated videos as AI-generated, which is good practice anyway.
The old laptop gets a job again
There is a plain economic story in this project. A card that a lot of people already own, retired from gaming years ago, can make the same frames a much newer machine would make. It just needs the night to do it.
The craft lesson travels further than the hardware one. Decide what a shortcut costs before you take it, check the frames that matter (faces, fur, fine texture) rather than the average, and write prompts like shot descriptions. If your old card is gathering dust, run the 17-frame check this weekend. You will know in ten minutes whether you have a second render machine.