3.3 Hardware: How Much VRAM You Really Need

Hardware is where the field's advice is least useful, because the people giving it are running hardware they were given or can write off against a channel that monetizes the review.

The good news is that the honest answer is cheap to state. What matters is a single number — VRAM — and the tier it puts you in. Everything else is secondary.

[AS OF 2026-10] Card prices and cloud rates both move. The tier boundaries below are what practitioners report, not what datasheets promise. [UNVERIFIED]

The number that decides everything

VRAM is the ceiling on what you can load at once. The model, the text encoder, the VAE, the conditioning inputs, and the intermediate latents all live in it simultaneously. When you run out, software either offloads to system RAM — which works and is several times slower — or fails.

Three practitioner statements capture the current reality better than any spec sheet:

• 24 GB is the realistic working floor for local video generation and for LoRA training on modern base models. Not a comfortable floor; the entry point.
• 16 GB does not suffice for that work. It runs image models with offloading and it will not train the identity artifact you need.
• 12 GB is enough for Z-Image, SDXL, and Klein-class models, and "maybe FLUX with quantization" — which is the honest way to describe the entry tier.

The pattern to notice: the tier that runs images is not the tier that runs your business. The difference between an image-capable machine and a persona-capable workflow is the training run and the video render, and both of them live one tier up.

What each tier actually runs

The 24 GB row deserves the details, because it is the configuration this book assumes. On a 4090-class card with 90 GB of system RAM, a 6-second image-to-video render takes roughly one to two minutes. [UNVERIFIED] That is the number to hold onto: not "can it run," but how long a single asset costs you, because your publishing cadence is a function of it.

Quantization: the boundary nobody mentions

Quantized model formats — GGUF and its relatives — compress weights to fit more model into less VRAM. Q8, Q6, Q5, Q4, each smaller and each less faithful.

The rule worth writing on the wall: below Q8, quality degrades noticeably, and it degrades first in exactly the detail you cannot afford to lose — fine facial structure, skin texture, hair edges. [UNVERIFIED]

That matters more for this book than for general image generation, because a persona is judged on whether the face holds together. A quantization level that produces a beautiful landscape and a slightly waxy face is a net loss for a persona pipeline.

Three practical positions:

• Q8 is the floor for identity-critical generation. If you need a lower quantization to fit, you are on the wrong tier for that job.
• Lower quantization is fine for drafts, backgrounds, b-roll, and thumbnails — work where the face is not the subject.
• Quantization is not a substitute for renting a bigger card for one afternoon. A training run is $2 to $14. Spending a month fighting a fragile setup to avoid that is a bad trade.

The supporting hardware people forget

VRAM is the headline, and it is not the whole machine.

System RAM. 32 GB is the practical minimum alongside 24 GB of VRAM. 90 GB or more becomes relevant when models offload or split across VRAM and RAM; the 6-second video benchmark above assumes that headroom. Under-specifying RAM is the most common reason a build underperforms its card.

Disk. Model weights are large: a full-precision FLUX-class checkpoint is in the 20-plus GB range, video models are comparable or larger, and you will accumulate several of each as you experiment. Add output archives — thousands of rendered assets at a few megabytes each runs into hundreds of gigabytes quickly. A 2 TB NVMe drive is the practical floor, and chapter 3.6 describes the structure that keeps it from becoming a landfill.

Electricity. A 450-watt system running four hours a day is roughly 1.8 kWh daily, which is somewhere around $0.30 to $0.50 per day at typical residential rates. [AS OF 2026-10] Not decisive, and worth knowing when you compare against cloud pricing.

Two hardware paths that look better than they are

Two configurations come up in every community thread, and both have a catch worth knowing before you spend money.

A high-VRAM gaming laptop. It runs the software and it is not equivalent to a desktop card with the same number on the box. Sustained render throughput is governed by thermals and power limits, and a laptop that benchmarks well for ninety seconds may throttle through an hour-long training run. Portability is real. So is the difference in sustained speed. [UNVERIFIED]

A used enterprise or workstation card with 24 GB. This is the cheapest legitimate route to the working tier, and it is a real option for a solo operator. The trade is in the details: blower cooling is loud, power draw and connector requirements are higher than a consumer card, and driver support on older generations can lag new frameworks. It is a good deal for someone who will tolerate the noise and check the framework compatibility before buying. It is a bad deal for someone who wants a machine that behaves like an appliance.

Apple Silicon, which deserves its own note. Unified memory means a machine with 24 GB or more of it can hold models that would not fit in a comparably priced discrete card, and the efficiency is genuinely good. The limits are elsewhere: the tooling ecosystem is less mature, training support is narrower and slower than on CUDA, and video work on the open-weight models is experimental at best. [UNVERIFIED] It is a viable machine for image generation and a poor machine for the training-heavy part of chapter 4.5. If you already own one, start there and rent CUDA for training. If you are buying specifically for this work, buy the tier table instead.

Renting versus buying

The cloud rates, first:

• RunPod: roughly $0.50 to $2.00 per hour for 4090-class and above, depending on tier.
• vast.ai: commonly cited around a third of that, with substantially more setup friction — the community's own description of the experience is that configuration is a pain.
• Hosted LoRA training — fal.ai and similar — around $2 at the base price for a training run. [AS OF 2026-10]

Now the arithmetic. A 24 GB card or a machine built around one costs somewhere in the region of $1,600 to $3,500. At $0.50 per hour, that sum buys 3,000 to 7,000 GPU-hours — which is months of continuous use, or years of two-hour days.

That comparison looks decisive until you add the terms it leaves out:

• Latency. Renting means uploading model weights or waiting for a container image, every session. Local means the model is already loaded. For a workflow built on small iterations, the friction is the cost.
• Volume. 20 GB downloads happen over your connection, repeatedly, on rented hardware.
• Privacy. Training data stays at home or it does not.
• Downtime. No vendor outage, no rate limit, no price change.

The honest decision rule: