How much VRAM to generate AI images?
Graphics memory decides which models, resolutions and processes a machine can run. Why the size of the model does not tell the whole story, the real question (the complete workflow), and the way forward without buying a GPU.
VRAM is the memory directly accessible to the graphics card; it has to hold the model, its text encoders, the VAE, the LoRAs, the intermediate data and the images, and a complete workflow consumes far more than a simple generation. The right question is not "how much VRAM for this model" but "how many resources for my complete workflow", because numerical precision, quantisation, resolution and software change everything. At Axiris, every model tested is documented with the hardware and the conditions of the test; and for anyone who does not want to buy a GPU, a machine in Switzerland, rented by the hour and delivered ready to use, exists.
Hardware is part of choosing the model
When a model runs locally, its hardware requirements are part of the workflow. The amount of graphics memory available decides which models, resolutions and processes a machine runs comfortably, or not at all.
VRAM
VRAM is the memory directly accessible to the GPU. It has to hold the model itself, its text encoders, the VAE, the LoRAs, the intermediate data, the images or latents, and the other components of the workflow.
A complete pipeline (generation, then transformation, then retouching by area, then upscaling) therefore consumes much more memory than a simple generation.
The size of the model does not tell the whole story
Real requirements also depend on numerical precision, quantisation, image resolution, the number of images produced at once, the software, offloading to system memory, the other models loaded at the same time and the type of workflow.
It is therefore risky to reduce the question to "how much VRAM does this model need?". The real question is: how many resources does my complete workflow require?
An order of magnitude, to give you a sense of scale: the official model card of Z-Image Turbo, read in August 2026, states that it runs in about 16 GB of graphics memory for one generation in its reference configuration. See Models of the moment. That is the publisher's figure, for a single model, outside any workflow. Video requires much more.
Speed and precision: two regimes
The same machine can serve two uses. For exploration, speed matters: a distilled model, which produces an image in few steps, lets you iterate without breaking the rhythm. For finalisation, quality matters: more steps, sometimes another model, the final resolution. A well-built workflow does not do everything at maximum quality; it reserves precision for what has been validated.
Optimising
ComfyUI and other environments offer strategies to adapt loading to the hardware: reducing precision, offloading part of it to system memory, loading models one after the other. Each one trades memory for speed or quality. When a generation fails for lack of memory, it is the first thing to read in the console: Troubleshooting ComfyUI.
What Axiris states
When we document a model we have tested, we state the hardware used and the conditions of the test. That makes it possible to tell a theoretical requirement from real production feedback, and to know whether a result carries over to your machine. The method: Comparing AI image models.
Without buying a GPU
Hardware is not a prerequisite. For a creative, you start on the studio's machine, without investing. For a company, a dedicated machine in Switzerland, rented by the hour and delivered ready to use, with the system, ComfyUI and Nova installed, gives the control of local without the purchase or the administration: A sovereign GPU in Switzerland, by the hour. Choosing between the ways: Local AI model or API?.
Contact request: a first conversation, free and with no commitment. Every practice is different; it is by listening to you that we propose what suits you.