When it is actually worth it
There are four situations where running locally gives you a real advantage. If you work with sensitive material that must not leave your machine. If you make a very large number of images and the vendor invoice is a noticeable monthly line item. If you train or modify your own model and what the vendors offer is too narrow for that. Or if content filters block you in work that is entirely legitimate — fine-art nudes, medical illustration, historical subjects.
In every other case the vendor route is faster, cheaper and better quality. That is worth saying plainly: running locally is not the default, it is a decision. Why it still matters to everyone that the option exists at all — Why Open Weights Matter covers that in a separate post.
What VRAM decides
With local generation a single number decides almost everything: how much memory your graphics card has. Not the processor, not the RAM, not the disk.
- The largest open-weight image models (around thirty billion parameters) want something like ninety gigabytes of video memory at full precision — that is not a home category.
- In a thrifty mode the same models start at around sixty-four gigabytes, but you are still talking about a professional card.
- The smaller variants of a few billion parameters, on the other hand, run on an eight-to-ten gigabyte card, and that is everyday hardware.
- On a strong home card the smaller, accelerated models hand you an image in one or two seconds — faster than what the vendors deliver.
The quantisation that makes this manageable
The numbers are frightening, but there is one simple technique that shifts an order of magnitude: storing the weights at lower precision. A half-precision build uses roughly forty percent less memory in exchange for a barely perceptible loss of quality.
In practice the community publishes several sizes of every popular model — the same model is available as a thirteen-gigabyte, heavily compressed build and as a sixty-four-gigabyte, full-precision one. The rule of thumb: take the largest one that fits on your card, and leave a gigabyte or two free as working space for the generation itself.
The licence most people misread
This is the most common and the most expensive misunderstanding in the field. "Open-weight" does not mean free for commercial use. A recently released model family shows exactly this: the smaller, few-billion-parameter variant came with a fully permissive licence, while its larger sibling is non-commercial. Same name, same announcement, two different legal situations.
Before you use anything in business work, open the model card and read the licence — specifically the variant you downloaded, not the family. It is also not rare for a much-discussed new model to never ship downloadable weights at all, only a hosted service; in that case the word "open" lives in the press release and nowhere else.
What the same thing costs in the cloud
To have something to measure against, it is worth knowing what the same work costs at a vendor. Today's prices run roughly as follows for a single image about a thousand pixels wide.
- The cheaper, fast models are around three to four cents.
- The middle tier is around seven cents.
- The top models are around thirteen cents.
- A rented, ready-made environment where you can run the open models without buying hardware starts at around twenty dollars a month.
Do the arithmetic for your own case: five hundred images a month in the middle tier is thirty-five dollars a month. At that rate a serious graphics card takes years to pay for itself — local generation is rarely the cheaper answer, it is the more independent one.
What the setup takes
The most widespread local environment is a node-based interface where you assemble the steps of generation out of boxes: load the model, encode the prompt, sample, save the image. It is unusual at first, but that is exactly where its power comes from — the steps hidden inside vendor interfaces are each separately accessible here, so you can build any workflow you want.
The learning curve is a few hours, and the community shares finished workflows you can load. What to count on, though: local generation does not mark your images by itself the way the big vendors do. If you publish them, the marking is your job — Watermarks and Provenance sets out what that takes.
Common misconceptions
- "Plenty of RAM is enough, the graphics card is not critical." It is. You can run from system memory, but orders of magnitude slower — an image takes minutes instead of seconds.
- "Open models are just as good as the closed ones." In text rendering and instruction following, typically not. In style and freedom they are often better.
- "Once I have downloaded it, I can do anything with it." Only what the licence allows — and that differs from variant to variant.
- "Running locally is free." The generating is; the hardware and the electricity are not. At low volume the service is cheaper.
What to take away
Local image generation is a real, working route in 2026 — but it is not about money. It is worth it for privacy, independence, freedom to train and unlimited experimentation; if you simply want good images fast, it is not. On which vendor tool suits which job, Choosing an AI Image Tool gives you the comparison.