Every week someone asks which GPU to buy for local image generation, and every week the replies argue about model numbers. This generation or last. Ti or Super. Which one benchmarks two frames faster in a game. For diffusion work, most of that argument is noise.
The number that decides what you can run is on the spec sheet, usually in smaller type: video memory. Speed is a slope. A slower card finishes the same image later. Memory is a wall. A model that does not fit either fails, or runs with parts of itself shuttled to system RAM at a cost you feel on every step.
What People Actually Own
Valve's Steam Hardware Survey for September 2026 is the best public snapshot of consumer GPUs. It skews toward gamers, which is exactly the population that buys desktop cards. Its VRAM table:
| VRAM | Share of surveyed systems |
|---|---|
| 6 GB or less | 19.18% |
| 8 GB | 26.71% |
| 12 GB | 13.06% |
| 16 GB | 27.21% |
| 24 GB | 5.50% |
| 32 GB or more | 1.91% |
Smaller tiers (10, 11, 20 and 22 GB, and "other") make up the rest. Two facts stand out. First, 16 GB is now the most common tier, narrowly ahead of 8 GB. Second, 45.89% of systems still have 8 GB or less, and only 7.41% have 24 GB or more. The single most common card in the survey is the RTX 5070 at 5.86%, a 12 GB card.
What the Models Need
Here is the Vision Studio model rack with the VRAM figure our verified model catalog declares for each:
| Model | Job | Catalog VRAM |
|---|---|---|
| Stable Diffusion 1.5 | Text to image | 4.0 GB |
| SD 3.5 Medium | Text to image | about 6 GB |
| SDXL | Text to image | about 8 GB |
| Stable Video Diffusion | Image to video | 8.0 GB |
| AnimateDiff | Text to video | 8.0 GB |
| LTX Video | Text to video | 9.4 GB |
| SD 3.5 Large | Text to image | about 12 GB |
| FLUX.1 [dev] | Text to image, and inpaint via FLUX Fill | 23.8 GB |
The top of that table is simple arithmetic. Black Forest Labs describes FLUX.1 [dev] as "a 12 billion parameter rectified flow transformer". At two bytes per parameter in bf16 or fp16, the transformer weights alone are about 24 GB, before text encoders, the VAE or the working memory a render needs.
Fits, Almost Fits, Does Not Fit
Between "fits" and "does not fit" is a grey zone, and it is where most buyers live. The Diffusers memory guide describes the tools for it honestly:
- Model offloading keeps one major component on the GPU at a time. It is the faster option, with smaller memory savings.
- Sequential CPU offloading moves weights layer by layer. It "dramatically reduces memory usage", and the guide calls it "extremely slow" and often impractical.
- VAE tiling decodes an image in overlapping tiles to cut peak memory, at the cost of possible slight tone variation between tiles.
Offload turns a wall into a slope. It is not free, and a card that needs it for every render will feel slower than its benchmark suggests.
How Vision Studio Handles the Grey Zone
Guessing is the expensive part, so the app does not make you guess. The Model Foundry's hardware fit check sizes each model against your GPU and system memory and labels it as fitting, fitting with CPU offload, over budget, or CPU-only, before you download gigabytes you cannot use.
When a model only fits with offload, it loads with CPU offload and VAE tiling, and an out-of-memory error steps down a fallback ladder instead of failing the job. The engine uses the first NVIDIA GPU through CUDA, Apple Silicon through Metal, or the CPU when there is no GPU.
System RAM Is the Second Wall
Offload does not make memory disappear. It moves weights from the GPU into system RAM, which means a card that relies on offload needs enough RAM to hold what was moved. The Diffusers guide is explicit about the cost of the faster streamed variant of offloading: it can increase CPU memory significantly, and it advises having twice as much system memory as the model size.
The good news is that the market is moving. The same September 2026 Steam survey shows 32 GB of system RAM as the most common configuration at 42.22%, ahead of 16 GB at 37.82%. If you are planning to run large models on a mid-range card, budget for RAM alongside the GPU. The fit check accounts for both: it sizes each model against your GPU and your system memory, not the GPU alone.
Buying Rules That Hold Up
- 8 GB: SD 1.5 and SD 3.5 Medium are comfortable. SDXL, SVD and AnimateDiff sit right at their catalog figure, so expect fallback to engage on larger sizes.
- 12 GB: adds LTX Video with headroom, and puts SD 3.5 Large at the line.
- 16 GB: the whole rack except FLUX.1 [dev], which needs offload. Now the most common tier, and the sensible floor for a new purchase.
- 24 GB: FLUX.1 [dev] at its catalog figure, with the least time spent waiting on offload.
- No dedicated GPU: the engine falls back to the CPU and the fit check labels models CPU-only. It works, and it is the slowest path by a wide margin. Treat it as a way to try the software, not a way to work.
Within a VRAM tier, then compare speed. Across tiers, memory wins: a faster 8 GB card loses to a slower 16 GB card the moment the model you want does not fit.
Buy the Wall, Not the Slope
Gaming benchmarks measure the slope, and they are the wrong benchmark for this job. For local generation, decide which models you want to run, read their memory needs, and buy the memory. The speed will follow. The models will not, if they do not fit.
Vision Studio is free, open-source (MIT), and runs locally on NVIDIA, Apple Silicon or CPU. The model rack, VRAM figures and fit check are documented at vision-studio-x.com/features. Download it at vision-studio-x.com/download.
-Rocky
#VRAM #GPU #LocalAI #FLUX #StableDiffusion #AIHardware #VisionStudio