What runs on the machine you have

The question everyone asks first

Can I run this on my laptop? The honest answer is: some of it, and more than you would expect, and the constraint is video memory rather than anything else.

Rough figures, for the diffusion family:

• 4 GB of video memory runs the older 512-pixel models comfortably, with community optimisations.
• 6 to 8 GB runs 1024-pixel models such as SDXL and its many fine-tunes.
• 12 GB runs most current open models in quantised form.
• 24 GB runs the largest open image models at full precision and makes video generation possible.

Apple Silicon Macs use shared memory and work well from 16 GB upward. A machine with no discrete graphics card can still generate on the processor alone, at roughly one to three minutes per image rather than seconds — slow, and entirely usable for someone producing a few images a day.

The techniques that lower the requirement

Quantisation. Model weights are stored at reduced precision — 8-bit or 4-bit instead of 16. A 12-billion-parameter model that needs 24 GB at half precision fits in 8 to 12 GB quantised, with a quality loss that is measurable and, at 8-bit, usually invisible. Community-distributed quantised files are standard and free.

Sequential offloading. Parts of the model are moved between system memory and video memory as needed. Slower, and it lets a large model run on a small card at all.

Tiling. The decoder is the memory spike at high resolutions. Processing it in tiles removes the spike at the cost of a little time.

Smaller and distilled models. A well-chosen 2-billion-parameter model produces better results than a badly-driven 12-billion one, and runs four times faster.

On a phone

This is worth stating because much of this readership has a phone and no laptop. Local generation on a phone is real: apps exist that run small diffusion models entirely on the device, offline, for free, taking tens of seconds per image on a recent handset. Quality is behind the large hosted models and ahead of what was available on a workstation three years ago.

For everything else, hosted free tiers are the route, and the practical advice is to treat free generation quota as a scarce resource: plan the image, use a fixed seed, and do the assembly and finishing in Photopea, which runs in a phone browser.

Choosing between local and hosted

The trade is not really about money, since both have free options. It is about four things:

Reproducibility. Local models do not change under you. Hosted ones do.

Privacy. A client's unreleased product photograph uploaded to a hosted service has left your control, and many contracts prohibit exactly that. Local generation never sends the image anywhere.

Control. Local tools expose every parameter, allow arbitrary conditioning combinations and custom fine-tunes, and impose no content filter — which is both the advantage and the responsibility.

Quality and convenience. The largest hosted models are ahead of the best open ones on several axes, and they require no maintenance.

Most working people use both deliberately: hosted for exploration and for the things the big models do best, local for client-confidential work, for anything that must be reproducible, and for volume.

Getting started without spending anything

A concrete route, in order:

1. Install ComfyUI or a WebUI. Both are free and both have one-click installers now.
2. Download one well-regarded open checkpoint suited to your memory.
3. Learn the five settings from this module: steps, guidance, seed, denoise strength, and one conditioning map.
4. Add Krita with its diffusion plugin if you prefer painting to node graphs.
5. Use free Colab or Kaggle notebook time for the occasional job too large for your machine, and for training a LoRA.

The honest caution: local setup consumes an evening the first time, and dependency problems are real. The payoff is a toolchain that costs nothing per image, works offline, keeps your work private, and does not change without your permission. For anyone doing this regularly, that is worth an evening.

One more consideration people rarely weigh, and should. Generation uses electricity, and a graphics card under load draws a few hundred watts. A single image is a trivial amount — comparable to running a desk lamp for a few minutes — and a habit of generating two hundred images to pick one is not trivial, either in cost or in the emissions behind it. The discipline that improves your results is the same one that reduces the waste: think before generating, fix the seed, change one thing, and stop when the picture is right rather than when the batch finishes. Careful work happens to be the cheap kind.