
FPS Drops Troubleshooting for DDR5 build
DDR5 build FPS drops needs a methodical troubleshooting pass. Review drivers, thermals, power settings, background apps, cables and game settings before replacing parts in an SA gaming setup.
Read moreWhen you generate an AI image locally, your GPU runs inference, which fits 12-24GB of VRAM and takes seconds per image. Training a full diffusion model needs hundreds of GPUs, though LoRA fine-tuning on consumer RTX hardware is a practical middle ground.
Type a prompt, wait a few seconds, and an image appears on your screen. That moment is your graphics card running inference, not training. The distinction matters because it decides what hardware you actually need. Generating art locally is a job a single consumer card can do comfortably. Building the model that does the generating is an entirely different scale of problem, and conflating the two leads people to wildly over-buy or wrongly assume their setup cannot cope.
When you generate an AI image at home, your GPU is doing inference, which fits comfortably in 12 to 24GB of VRAM and takes seconds per picture. Training a full diffusion model from scratch needs dozens of datacentre GPUs and tens of gigabytes of memory each. The practical middle ground for enthusiasts is LoRA fine-tuning, which runs on a consumer RTX card.
Inference is the act of using a model that already exists. The weights are fixed; your card simply pushes your prompt through them and produces an output. For image generation this is light work by modern standards. Standard square images can run on as little as 4 to 6GB of VRAM, and a card with 12GB or more handles higher resolutions and faster batches without fuss. This is why a single mainstream GPU is all most people ever need to make art locally.
The reason it feels quick is that nothing is being learned. The model is not changing, so there is no need to hold the extra memory that learning demands.
Training is where the model's weights are created or rewritten, and the memory cost explodes. A full fine-tune that updates every weight in a diffusion model can demand roughly 40 to 80GB of VRAM, which puts it squarely on datacentre cards or multi-GPU rigs. Building a model from scratch goes further still, into the realm of large clusters running for extended periods.
That is the figure that scares people off, but it is also the figure most home users never need to meet. You are almost always running someone else's trained model, not making your own from nothing.
Between pure inference and full training sits LoRA fine-tuning, and this is where consumer hardware earns its keep. A LoRA trains a small adapter, often just 50 to 300MB, that teaches an existing model a new style, character or subject without touching the bulk of its weights. It is genuinely achievable on a 12GB card in an afternoon, which means a capable desktop can do real customisation, not just generation.
As a rough guide, lighter models train a LoRA on around 8GB, while the larger current models want 12GB at minimum and are more comfortable with 16 to 24GB. Techniques like gradient checkpointing and mixed precision trim the memory cost further at a modest speed penalty. A well-specified AI creator PC from Evetech covers both fast inference and LoRA work in one machine. If you are choosing the card itself, the GPU best sellers show which options are pulling their weight for this kind of workload.
Less than most people expect. Standard images can run on roughly 4 to 6GB, and a card with 12GB or more handles higher resolutions and faster generation comfortably. Inference is light because the model is fixed and nothing is being learned.
Training a full model from scratch is impractical at home, since it needs tens of gigabytes of VRAM per card and often many cards. What you can do at home is LoRA fine-tuning, which adapts an existing model and runs on a single consumer GPU.
A LoRA trains a small add-on adapter while leaving most of the base model untouched, keeping memory needs low enough for consumer cards. Full fine-tuning rewrites every weight and needs far more VRAM, typically datacentre-class hardware.
Mostly quicker. A stronger card with more VRAM produces images faster and at higher resolutions, but the visual result is set by the model and your prompt, not the card. Extra VRAM matters more for training and large batches than for image quality alone.
Whether you want to generate art in seconds or train your own LoRA styles, the right balance of GPU and VRAM makes it painless. Explore the AI creator PC range at Evetech and build a machine that handles both.
Less than most people expect. Standard images can run on roughly 4 to 6GB, and a card with 12GB or more handles higher resolutions and faster generation comfortably. Inference is light because the model is fixed and nothing is being learned.
Training a full model from scratch is impractical at home, since it needs tens of gigabytes of VRAM per card and often many cards. What you can do at home is LoRA fine-tuning, which adapts an existing model and runs on a single consumer GPU.
A LoRA trains a small add-on adapter while leaving most of the base model untouched, keeping memory needs low enough for consumer cards. Full fine-tuning rewrites every weight and needs far more VRAM, typically datacentre-class hardware.
Mostly quicker. A stronger card with more VRAM produces images faster and at higher resolutions, but the visual result is set by the model and your prompt, not the card. Extra VRAM matters more for training and large batches than for image quality alone.