The GPU you use for ComfyUI determines which models you can run, how fast images generate, and how much compromise you’re accepting. VRAM is the primary constraint — not compute speed, not memory bandwidth — and the numbers are fairly unforgiving. This guide covers what you actually need at each tier, which specific cards are worth considering, and what to expect if you’re on a tighter budget.
Why VRAM Is the Bottleneck
AI image generation keeps the model loaded in GPU memory throughout the generation process. If the model doesn’t fit in VRAM, ComfyUI either refuses to run or falls back to system RAM — which is orders of magnitude slower. Unlike general gaming where a fast GPU with less VRAM often still works well, for local image generation VRAM capacity is the hard limit.
Compute speed (CUDA cores, shader count) matters for generation speed, but being underpowered on compute just makes things slower. Being underpowered on VRAM makes things not work at all, or tolerable only with heavy quantisation.
VRAM Requirements by Model
- Stable Diffusion 1.5 (512×512): 4GB VRAM minimum. Runs on almost any discrete GPU from the last decade.
- SDXL (1024×1024): 8GB VRAM for the base model at standard settings. 6GB is possible with
--lowvrambut slow. - Pony Diffusion V6 XL: Same as SDXL — 8GB comfortable.
- Flux.1-schnell or flux1-dev (fp8/GGUF quantised): 8GB VRAM, functional but slow. Better at 12GB.
- Flux.1-dev (full fp16/bf16): 16–24GB VRAM. 12GB is tight even with optimisations.
For most users in 2026, 12GB VRAM is the practical minimum for a good all-round experience. 8GB works for SDXL and quantised Flux, but you’ll feel the limits quickly.
RTX 50 Series — Current Generation (Blackwell)
If you are buying new in 2026, the RTX 50 series (Blackwell architecture) is the right starting point. GDDR7 memory delivers significantly higher bandwidth than the GDDR6X in the 40 series, and Nvidia has addressed the VRAM problem that plagued the RTX 4060 Ti — the RTX 5060 Ti ships with 16GB as standard. Generation speeds are noticeably faster than equivalent 40 series cards at the same price point.
RTX 5060 Ti (16GB) — Best Budget Buy
The 5060 Ti is the most important card in the 50 series lineup for most ComfyUI users. 16GB of GDDR7 means you can run Flux.1-dev fp8, SDXL at full quality, and all SD 1.5 workflows without hitting memory limits. Generation speed is meaningfully faster than the RTX 3060. At around £400–£500 new, it is the default recommendation for anyone starting out. Check price on Amazon
RTX 5070 (12GB) — Mid-Range
The 5070 sits in an awkward spot — 12GB of GDDR7 is capable but less than the 5060 Ti’s 16GB. Where it wins is raw compute speed: generation times are noticeably faster, particularly on Flux workflows. If you run ComfyUI heavily throughout the day and generation speed matters more than running the absolute largest models, the 5070 makes sense at around £550–£650. Check price on Amazon
RTX 5070 Ti (16GB) — High-End Sweet Spot
The 5070 Ti combines 16GB GDDR7 with significantly higher compute performance than the 5060 Ti. You can run Flux.1-dev at full quality, SDXL with heavy upscaling nodes, and video workflows (AnimateDiff, Wan2.1) without hitting limits. At around £750–£900 this is the card most serious ComfyUI users should be targeting. Check price on Amazon Check price on Amazon
RTX 5080 (16GB) — Enthusiast
The 5080 offers substantially faster compute than the 5070 Ti with the same 16GB VRAM. The difference shows in video generation and complex multi-model workflows. At around £1,100–£1,300, it is harder to justify over the 5070 Ti unless generation speed is genuinely costing you time. Check price on Amazon
RTX 5090 (32GB) — No Compromise
32GB of GDDR7 and the fastest consumer GPU available. You can run every model at full precision, generate video without batching constraints, and run multiple models simultaneously. At around £1,800–£2,200 it is not for most users, but if ComfyUI is central to your work it removes every limitation. Check price on Amazon
RTX 40 Series — Previous Generation / Used Market
The RTX 40 series remains a strong option on the used market, where prices have dropped significantly since the 50 series launched. If you are comfortable buying used or refurbished, these cards offer good value — particularly the 3060 and 4070 Ti Super where the price-to-VRAM ratio is favourable.
NVIDIA remains the best-supported option for ComfyUI. CUDA support is mature, most custom nodes are tested on NVIDIA first, and the software ecosystem just works.
Budget: RTX 3060 (12GB) — approx. £250–£300 used Check price on Amazon
The RTX 3060 ships with 12GB VRAM in its standard form (the 8GB variant exists — avoid it). For the price, it’s one of the most VRAM-efficient options available. You can run Flux.1-dev fp8, SDXL at full quality, and all SD 1.5 variants without issue. Generation speed is adequate — not quick, but entirely usable for most workflows.
It’s an older card (Ampere architecture, 2021) so raw compute speed is behind current-gen equivalents, but the 12GB VRAM makes it punch above its price bracket for image generation specifically.
Mid-range: RTX 4070 (12GB) — approx. £500–£550 new Check price on Amazon
The RTX 4070 offers a significant performance uplift over the 3060 at a higher price. Ada Lovelace architecture, better power efficiency, faster generation times. The 12GB VRAM is the same as the 3060 so the model support is identical, but you’ll generate 60–80% faster. For users generating many images regularly, the speed difference is meaningful.
The RTX 4070 Ti Super (16GB) Check price on Amazon at around £700–£750 is worth considering if your budget stretches — the extra 4GB VRAM opens up full-precision Flux comfortably and removes the need for quantised models.
High-end: RTX 4090 (24GB) — approx. £1,700–£1,900 new Check price on Amazon
The RTX 4090 is the current consumer NVIDIA benchmark. 24GB VRAM means full-precision Flux.1-dev with headroom to spare. Generation speed is in a different league to everything below it. If budget is no object and you’re serious about local image generation, this is the card — but the price premium is significant and most users get 80–90% of the utility from an RTX 4070 Ti Super at less than half the price.
Used RTX 3090s (24GB) can be found for £600–£800 and offer similar VRAM capacity to the 4090 with slower compute speeds. Good value if you care more about VRAM than generation speed.
AMD GPU Support
AMD GPUs can run ComfyUI but with caveats:
- Linux (ROCm): AMD’s ROCm compute platform works reasonably well on Linux for supported cards (RX 6000 and 7000 series). Performance is close to NVIDIA equivalents, and most PyTorch operations work. Some custom nodes may have issues.
- Windows (DirectML): AMD support on Windows uses DirectML, which is slower and less compatible than ROCm. Expect slower generation and occasional node compatibility problems. Manageable, but inferior to NVIDIA on the same OS.
An RX 7900 XTX (24GB) on Linux with ROCm is a genuinely competitive option — 24GB VRAM at around £800–£900 new. For Windows users, NVIDIA is significantly less hassle.
Apple Silicon
Macs with M1, M2, M3, or M4 chips use unified memory — the same pool serves both CPU and GPU. This means an M2 Pro with 32GB has 32GB available for ComfyUI, compared to a dedicated GPU where you’re limited to the card’s VRAM alone.
Performance is solid. An M3 Pro with 18GB handles Flux.1-schnell fp8 at a usable pace. An M3 Max with 36GB or 48GB competes with mid-to-high-end NVIDIA cards for Flux workloads. Apple Silicon Macs are a legitimate option, particularly if you’re buying a Mac anyway — the image generation capability is a bonus rather than a separate purchase.
CPU-Only: Last Resort
ComfyUI runs CPU-only with the --cpu flag, but generation times are brutal — 5–15 minutes per image for SD 1.5, longer for larger models. It’s useful for testing that software works, or for very occasional use where you’re not in a hurry. Don’t rely on it as a primary setup.
Summary: Which Card to Buy
- Minimum viable (8GB): RTX 3070, RTX 4060 — SDXL works, Flux is compromised
- Best value (12GB): RTX 3060 12GB (used), RTX 4070 — all models, Flux fp8 comfortable
- Recommended (16GB): RTX 4070 Ti Super — full-precision Flux, no compromises
- No-compromise (24GB): RTX 4090, RTX 3090 (used), RX 7900 XTX (Linux)
For model-specific requirements, see Best Models for ComfyUI in 2026. For getting started once your hardware is sorted, see the Windows installation guide or Mac installation guide.
Future-Proofing Your Purchase
Model sizes are increasing. Flux.1 requires more VRAM than SDXL, and future architectures will likely push requirements higher. If you are buying a GPU specifically for ComfyUI and plan to use it for two or more years, buying 4GB more VRAM than you currently need is consistently good advice. The extra cost at purchase is almost always less than the cost of upgrading early.
For integrated graphics or laptop iGPUs, ComfyUI will technically run but is not a practical setup for regular use. The VRAM limitations (typically 2–4GB shared with system RAM) mean only the smallest quantised models load without errors, and generation times are impractical. If you are on a laptop, check whether your model has a dedicated NVIDIA GPU alongside the integrated graphics — ComfyUI will use the dedicated card if present.
For a full index of every ComfyUI guide on Serverman, see the ComfyUI complete guide and hub.
