ComfyUI local AI computer guide
How to run ComfyUI AI workflows locally with 8 GB VRAM
We ran ComfyUI and Wan 2.2 on a computer with 48 GB of system RAM and an 8 GB RTX 4060. It made usable local video, slowly, without a top-of-the-line graphics card.
Our test computer
48 GB system RAM
This is the computer's main memory. It gives ComfyUI room to move model parts away from the graphics card and keep large work files open.
8 GB GPU memory
This is the fast memory on the graphics card. We used an NVIDIA RTX 4060. It is limited in capacity, so ComfyUI must move data in and out while a large job runs.
Wan 2.2 video model
We used the open source Wan 2.2 image/text-to-video AI model to test a real video workflow.
What you can build on one local workstation
Images and style tests
Make pictures, change colors and clothing, build moodboards, or use small style add-ons called LoRAs. They are good for exploring a look, but not perfect for keeping one character exactly the same.
Video and cleanup
Create short clips from image or text, add frames for smoother motion, enlarge output, make masks to alter videos, remove backgrounds, and join several parts into one scene.
Batch media work
Resize, convert, or clean up many files.
Why high RAM matters for ComfyUI
The same computer can make images, turn them into short videos, enlarge pictures, create masks, cut out objects, add frames, join layers, and convert files in batches. Extra RAM keeps large models and temporary files ready for the next job.
One high-RAM computer can cover that work without sending every job to a paid cloud model.
How ComfyUI works
ComfyUI looks like a board of boxes joined by lines. Each box does one step. One may load a model. Another reads your words. Another makes an image. The last box saves the file.
The best simple description is: Photoshop node graph meets alchemy lab meets synthesizer rack. It is powerful, but it can feel strange at first. There are many knobs, model names, seeds, masks, samplers, and add-ons. A copied workflow may break after one missing file or update.
Much of ComfyUI is learned by trying a workflow, looking at the output, and changing one setting at a time.
How to start with Wan 2.2 on 8 GB VRAM
- Install or update ComfyUI from the official download page.
- Open Workflow, then Browse Templates, then Video.
- Choose Image to Video or Text to Video workflows from Wan 2.2 templates. The 5B version is the sensible place to begin on an 8 GB card.
- Download the model files listed by the template and place them in the folders shown in the guide.
- Keep the starting size and frame count low.
- Write a short, clear request. We found that running the prompt into Qwen 3.5 greatly improved the output quality. Use the following prompt for Qwen: "Expand this into a Wan 2.2 video prompt."
- Write a negative prompt to block unwanted generation habits.
- Download any VAE and LoRA files required by the workflow. The files we encountered ranged from 1–14 GB.
- If it works, change only one item at a time. Raise the frame count or picture size in small steps.
ComfyUI's official Wan 2.2 guide says the 5B model should fit well on 8 GB VRAM with its built-in memory offload. Your run time will change with the graphics card, output size, frame count, model files, and other software using memory.
How 8 GB reaches beyond 8 GB
When a job does not fit in GPU memory, ComfyUI can keep model parts in system RAM and move them to the GPU as needed. This is called offload. The GPU is the workbench; system RAM is the shelf beside it.
Other tools can also help:
- Smaller model files: quantized models trade some accuracy and detail for a smaller memory footprint.
- Tiling: the computer works on small parts of an image instead of the whole image at once. Seams or added run time are possible.
- Queued renders: let slow jobs run one after another, often overnight.
- Graph changes: remove unneeded steps and unload a model after its part of the job is done.
- Small previews: test the idea at a low size before spending time on the final output.
This does not make an 8 GB card perform like a 24 GB card. But it lets a more careful setup perform and make useful outputs.
A five-second clip from a six-minute workflow
This five-second video of a spoon falling onto a cheesecake took about six minutes to generate on our test computer with 48 GB of system RAM and an 8 GB RTX 4060. We used Qwen to expand and improve the prompt before running the ComfyUI workflow.
This is a useful example of a practical local workflow: generate a strong five-second starting clip in about six minutes, then use a video-edit workflow to correct the release and contact continuity while preserving the successful motion and splash. The RTX 4060 is the kind of card many people can realistically afford, and the result is already close enough to improve rather than restart.
Where local video still looks weak
Local image-to-video can misunderstand your prompts badly. Longer clips also give the model more time to drift away from the first image.
But a new idea, a clear story (with Qwen editing your video prompts), and a short clip can matter more than perfect realism. The output is also private and easy to share.
Who should choose more RAM?
Choose a high-RAM machine when you want one local computer for many creative jobs, keep several large models, or plan to work with video and large temporary files. For our 8 GB GPU test, 48 GB of system RAM gave the workflow useful room. A 64 GB machine is an easier target to find and gives more space for future jobs.
Choose a larger GPU first when speed is the main goal, you need high-resolution video every day, or you do not want to tune low-memory workflows. More system RAM expands what can be attempted. More GPU memory usually makes those jobs easier and faster.
What the test proved
Our 48 GB machine moved among image creation, video, cleanup, masking, and file conversion on an 8 GB graphics card. More GPU memory would make the same jobs faster and easier.