Short answer
For local AI image generation, VRAM matters more than anything else. 16 GB is the sweet spot in 2026: it runs SDXL comfortably and FLUX with FP8 or GGUF quantized weights. The RTX 5060 Ti 16 GB is the best budget pick, the RTX 5070 Ti 16 GB the best value for speed, and the RTX 5090 32 GB the best overall, especially for full-precision FLUX and AI video. A used RTX 3090 24 GB is a strong value alternative. NVIDIA remains the easiest choice because nearly every tool is built on CUDA.
Why VRAM is the most important spec
Image models must hold their weights, the text encoders, the image decoder (VAE) and working data in GPU memory. If they do not fit, tools fall back to offloading parts to system RAM, which can make generation many times slower, or they fail with an "out of memory" error.
More VRAM also lets you:
- Generate at higher resolutions and in larger batches
- Use ControlNet, IP-Adapter and LoRAs at the same time
- Train your own LoRAs
- Run newer, larger models at full quality
- Generate AI video, which needs far more memory than still images
For a deeper explanation, see What is VRAM?.
How much VRAM each model needs
Approximate comfortable VRAM for generating 1024×1024 images in popular tools like ComfyUI or Forge. Memory-saving options can go lower at the cost of speed.
| Model | Size | Comfortable VRAM | Minimum with tricks |
|---|---|---|---|
| Stable Diffusion 1.5 | about 1B | 6 to 8 GB | 4 GB |
| SDXL | 3.5B (base) | 10 to 12 GB | 6 to 8 GB |
| Stable Diffusion 3.5 Large | 8.1B | 16 to 24 GB | 12 GB (quantized) |
| FLUX.1 dev / schnell (FP16) | 12B | 24 GB+ | — |
| FLUX.1 dev (FP8 or GGUF Q8) | 12B | 16 GB | 12 GB |
| FLUX.1 dev (GGUF Q4) | 12B | 8 to 12 GB | 8 GB |
| AI video, small models (e.g. Wan 2.1 1.3B) | 1.3B | 8 to 12 GB | 8 GB |
| AI video, large models (e.g. Wan 2.1 14B) | 14B | 24 to 32 GB+ | 16 GB (quantized, slow) |
| LoRA training (SDXL / FLUX) | — | 16 to 24 GB+ | 12 GB |
Quantization shrinks models with a small loss of quality; the same ideas apply as for LLM quantization.
Our GPU picks
Prices below are NVIDIA's and AMD's US launch prices. Street prices in 2026 are often higher because of memory shortages and AI demand (see why GPU prices are rising), so compare current offers before buying.
| GPU | VRAM | Launch price | Best for |
|---|---|---|---|
| RTX 5090 | 32 GB GDDR7 | $1,999 | Everything: full-precision FLUX, video, training |
| RTX 4090 (used) | 24 GB GDDR6X | $1,599 (2022) | Fast all-rounder if found at a fair price |
| RTX 3090 / 3090 Ti (used) | 24 GB GDDR6X | Used market | Best VRAM per dollar |
| RTX 5080 | 16 GB GDDR7 | $999 | Fast 16 GB card |
| RTX 5070 Ti | 16 GB GDDR7 | $749 | Best value for speed |
| RTX 5060 Ti 16 GB | 16 GB GDDR7 | $429 | Best budget pick |
| RTX 3060 12 GB (used) | 12 GB GDDR6 | Used market | Entry level SD 1.5 / SDXL |
| Radeon RX 9070 XT | 16 GB GDDR6 | $599 | AMD option, Linux users |
Best overall: RTX 5090 (32 GB)
The fastest consumer GPU for AI image generation, with 32 GB of fast GDDR7 and about 1.8 TB/s of memory bandwidth. It runs FLUX at full precision with room for ControlNets and LoRAs, handles larger AI video models and trains LoRAs quickly. Its fifth-generation tensor cores support FP4 for even faster quantized inference. It draws up to 575 W, so check your power supply. Full analysis: Is the RTX 5090 good for AI?.
Best value for VRAM: used RTX 3090 (24 GB)
Older and slower than current cards, but 24 GB of VRAM at used prices makes it a favourite for FLUX, video models and LoRA training. Check the card's condition, cooling and warranty, and expect high power draw.
Best value for speed: RTX 5070 Ti (16 GB)
Fast Blackwell architecture with 16 GB of GDDR7. It handles SDXL quickly and FLUX in FP8 or GGUF formats, which is what most people use day to day.
Best budget: RTX 5060 Ti 16 GB
Make sure you buy the 16 GB version, not the 8 GB one. It is slower than the 5070 Ti, but the extra memory means it can run the same models, just taking longer per image. It is the cheapest new card we recommend for FLUX.
Avoid for AI image work
- 8 GB cards (RTX 5060, RTX 4060, 5060 Ti 8 GB): fine for SD 1.5, tight for SDXL and very limited for FLUX and video.
- RTX 5070 (12 GB): fast, but 12 GB limits newer models. A 16 GB card is a better buy for AI.
NVIDIA vs AMD vs Apple vs Intel
- NVIDIA: the default. ComfyUI, Forge, Automatic1111, kohya_ss and nearly all custom nodes are developed and tested on CUDA first.
- AMD: works with ROCm on Linux and increasingly on Windows, but setup is harder, some extensions do not work and performance per dollar is often lower for AI. See AMD vs NVIDIA for AI.
- Apple Silicon Macs: run Stable Diffusion and FLUX through apps like Draw Things or ComfyUI with Metal. Large unified memory helps fit big models, but generation is usually much slower than a high-end NVIDIA GPU.
- Intel Arc: works with some tools, but software support is the most limited.
Other parts of your PC
- System RAM: 32 GB minimum, 64 GB recommended for FLUX and video workflows.
- Storage: a fast NVMe SSD with plenty of space; model files are often 5 to 25 GB each.
- Power supply: follow the GPU maker's recommendation (1,000 W for an RTX 5090).
Should you rent a cloud GPU instead?
If you generate images only occasionally, or want to try large video models before buying, renting can be cheaper. Consumer cards like the RTX 4090 and 5090 rent for well under a dollar an hour on some GPU marketplaces, and data center GPUs such as the H100 are available by the hour. Compare live prices in our GPU cloud price tracker, and see rent vs buy a GPU and cheapest cloud GPU.
Frequently asked questions
Can I run Stable Diffusion on 8 GB of VRAM?
Yes for SD 1.5 and SDXL with memory-saving settings, and FLUX with heavily quantized GGUF models, but it will be slower and more limited. If you are buying a GPU for AI, choose 16 GB or more.
Is the RTX 5090 worth it over the 5080 for Stable Diffusion?
If you want full-precision FLUX, AI video or fast LoRA training, yes: the jump from 16 GB to 32 GB matters more than raw speed. For SDXL and quantized FLUX only, the 5080 or 5070 Ti is much better value.
Does Stable Diffusion use tensor cores?
Yes. Image generation is dominated by matrix multiplications in FP16, BF16 and FP8, which run on tensor cores. See tensor cores vs CUDA cores.
Is the same GPU good for local LLMs?
Mostly yes. Both workloads reward VRAM and memory bandwidth. Our best GPU for local LLMs guide covers language models specifically.
Sources and further reading
- NVIDIA, GeForce RTX 5090 specifications
- Black Forest Labs, FLUX.1 dev on Hugging Face
- Stability AI, Stable Diffusion 3.5 Large on Hugging Face
- Podell et al., SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis (2023)
- ComfyUI, GitHub repository




Leave a Reply