LocalForge AILocalForge AI

Stable Diffusion VRAM & Hardware Requirements (Local, 2026)

Every forum thread says "8 GB is enough." Then you load SDXL with ControlNet and watch your system page to disk. This guide exists because most hardware recommendation lists give you vibes instead of measurements. You deserve actual VRAM numbers per model class, tested at real resolutions, so you can buy the right GPU instead of the cheapest one that technically boots.

We're covering three model classes: SD1.5, SDXL, and Flux. Each has different VRAM demands, and the gap between them is larger than most guides admit. We assume Windows + NVIDIA because that path still has the fewest driver headaches in 2026. AMD works via ROCm on Linux. Apple Silicon runs through MLX and MPS backends - expect separate install guides for both.

Below you'll find measured VRAM usage, a GPU comparison table with current prices, and a clear upgrade priority order. No "it depends" - just numbers you can check against your own Task Manager.

The Quick Answer

Key Takeaway - May 2026

Buy an RTX 3060 12 GB if you're on a budget ($170 - $280 used). It runs SD1.5 comfortably, handles SDXL at 1024x1024, and fits Flux GGUF Q5 quantizations. If you want Flux at full quality or run SDXL with heavy ControlNet stacks, get an RTX 4070 12 GB ($500 - $600 new). The RTX 4090 24 GB ($1,600+) is the only card that runs Flux Dev at FP16 without quantization. Or use LocalForge AI for a pre-configured Forge setup - but it can't add VRAM to your card.

GPU VRAM SD1.5 512x512 SDXL 1024x1024 Flux Dev Street Price (May 2026)
RTX 3060 12 GB 4 GB used, ~4 sec 8 GB used, ~18 sec Q5 GGUF only (9 GB) $170 - $280
RTX 4060 8 GB 4 GB used, ~3 sec 8 GB used, ~50 sec Q4 GGUF only (7 GB) $280 - $320
RTX 4070 12 GB 4 GB used, ~2 sec 7.5 GB used, ~10 sec FP8 fits with careful settings $500 - $600
RTX 4070 Ti Super 16 GB 4 GB used, ~2 sec 7.5 GB used, ~8 sec FP8 with more headroom $750 - $850
RTX 4090 24 GB 4 GB used, ~1 sec 7.5 GB used, ~4 sec FP16 fits (22 GB) $1,600 - $2,000

Minimum vs Recommended Specs

The floor to run Stable Diffusion is an NVIDIA card with 6 GB+ VRAM, 16 GB RAM, 50 GB free on an SSD and a modern quad-core CPU; the target for SDXL is 12 GB VRAM and 32 GB RAM. Here is the floor, the target and the full-quality Flux tier side by side. The detail behind each row is in GPU VRAM Reality, RAM and Storage Requirements and CPU Requirements below.

Tier GPU VRAM System RAM Storage CPU What it runs
Minimum 6 GB+ (a 4 GB card can work with optimized builds but expect OOM errors beyond basic txt2img) 16 GB 50 GB free SSD Any modern quad-core (Intel i5 / Ryzen 5 from 2018+) SD1.5 at 512x512
Recommended 12 GB (SDXL's real comfort zone) 32 GB 200 GB+ if you collect models Any modern quad-core (Intel i5 / Ryzen 5 from 2018+) SDXL at 1024x1024 with ControlNet headroom; Flux via GGUF Q5 or FP8 with careful settings
Full-quality Flux 24 GB 32 GB (64 GB only for training) 200 GB+ Any modern quad-core (Intel i5 / Ryzen 5 from 2018+) Flux Dev FP16 (22 - 24 GB)

8 GB cards run base SDXL with zero headroom.

Check These Numbers Yourself

Every VRAM figure on this page is a typical peak, and you can compare it with your own card in a few minutes:

  • Open Task Manager (Performance tab, your NVIDIA GPU) or your GPU monitor of choice and watch dedicated GPU memory while an image generates.
  • Use the same resolution as the figure you're checking, such as 512x512 for SD1.5 or 1024x1024 for SDXL and Flux.
  • Keep the steps and sampler the same between runs so only one thing changes at a time.
  • Match the precision (FP16, FP8 or a GGUF quant) to the line you're comparing against.
  • Read the highest value reached during the run, not the idle reading before or after it.

GPU VRAM Reality

Here's what actually gets consumed at inference time. These are typical peak figures - your idle usage will be lower.

SD1.5 at 512x512 (the lightweight class):

  • Base model load: ~4 GB VRAM total (1.7 GB weights + 0.2 GB VAE + 0.2 GB text encoder + activations and overhead)
  • With ControlNet: add 1 - 2 GB depending on the preprocessor
  • Batch size 2: roughly doubles activation memory, pushing toward 6 - 7 GB
  • Bottom line: any 6 GB+ card handles SD1.5 without tricks. A 4 GB card can work with optimized builds but you'll fight OOM errors on anything beyond basic txt2img.

SDXL at 1024x1024 (the mainstream class):

  • Base model load: ~8 GB VRAM (5.2 GB weights + 1.6 GB text encoder + 0.2 GB VAE + activations)
  • With ControlNet + hires fix: peak can hit 10 - 12 GB
  • With VAE tiling enabled: drops peak to ~6 - 7 GB at the cost of some speed
  • Bottom line: 8 GB cards run base SDXL but leave zero headroom. The moment you add ControlNet or a hires pass, you OOM. 12 GB is the real comfort zone for SDXL.

Flux Dev at 1024x1024 (the heavy class):

  • FP16 (full precision): 22 - 24 GB VRAM. Only RTX 4090 and workstation cards fit this.
  • FP8 quantized: 12 - 16 GB VRAM. RTX 4070 12 GB fits with careful settings.
  • GGUF Q8: 12 - 16 GB. Similar to FP8, slightly better quality retention.
  • GGUF Q5: 8 - 10 GB. Runs on RTX 3060 12 GB.
  • GGUF Q4/NF4: 6 - 8 GB. Fits 8 GB cards. Noticeable quality drop on fine details.
  • Bottom line: Flux at full quality requires a $1,600 GPU. Quantized Flux on a $200 used RTX 3060 looks surprisingly good. Q5 is a smaller step down from FP16 than most people expect.

Recommended GPUs by Budget

Under $200 - RTX 3060 12 GB (used):

  • TDP: 170W
  • Handles: SD1.5 natively, SDXL at 1024x1024, Flux via Q5 GGUF
  • SDXL speed: ~15 - 25 seconds per image at 1024x1024 (28 steps)
  • Why it wins: 12 GB of VRAM at the lowest price point. Nothing else under $250 gives you 12 GB. The memory bandwidth is slower than 40-series cards, so generation takes longer - but it finishes without OOM crashes.

$280 - $320 - RTX 4060 8 GB (new):

  • TDP: 115W (great for small builds and laptops)
  • Handles: SD1.5 natively, SDXL at 1024x1024 with no headroom, Flux via Q4 GGUF only
  • The catch: 8 GB sounds like "enough for SDXL" until you add ControlNet. Then it isn't. The RTX 3060 12 GB at $180 used is a better SD card despite being older. Buy the 4060 only if power efficiency or warranty matter more than VRAM.

$500 - $600 - RTX 4070 12 GB (new):

  • TDP: 200W
  • Handles: everything the 3060 does, 2 - 3x faster, plus Flux FP8 that fits with careful settings
  • SDXL speed: ~10 seconds per image at 1024x1024 (28 steps)
  • Why it wins: fast enough that SDXL iteration feels responsive. Flux FP8 fits with careful settings instead of barely squeezing in. This is the sweet spot for serious local generation in 2026.

$750 - $850 - RTX 4070 Ti Super 16 GB (new):

  • TDP: 285W
  • Handles: Flux FP8 with room for ControlNet, SDXL with aggressive hires workflows
  • Who needs it: creators stacking Flux + ControlNet + IP-Adapter in the same pipeline. The extra 4 GB over the RTX 4070 prevents OOM when workflows get complex.

$1,600+ - RTX 4090 24 GB (new):

  • TDP: 450W
  • Handles: Flux Dev FP16 natively, SDXL in ~4 seconds, everything without compromise
  • SDXL speed: ~4 seconds per image at 1024x1024 (28 steps)
  • Reality check: this is a luxury purchase for local AI. Q8 on an RTX 4070 gets close to full-quality Flux at a much lower price. Buy the 4090 if you also train models or do video generation.

RAM and Storage Requirements

System RAM:

  • 16 GB: works if you close your browser and don't multitask. SD1.5 and basic SDXL workflows fit.
  • 32 GB: the real recommendation. Keeps things stable when you run a browser, Discord, and a local LLM alongside your image gen UI. RAM is $40 - $60 for a 16 GB DDR4 stick - don't cheap out here.
  • 64 GB: only needed if you're training models or running multiple AI tools simultaneously.

Storage:

  • Minimum: 50 GB free on an SSD. Forge + one SDXL checkpoint + one Flux GGUF = ~25 GB. You'll want breathing room.
  • Comfortable: 200 GB+ free if you collect models. Five SDXL checkpoints + three Flux variants + LoRAs add up fast.
  • SSD vs HDD: as a rough estimate, model load times drop from 30 - 40 seconds on a spinning disk to 3 - 8 seconds on NVMe. That's the difference between "I'll wait" and "I'll switch models mid-session." PCIe Gen 3 NVMe is fine - as rough estimates, Gen 4 gives ~35% faster loads but Gen 5 adds almost nothing (8 - 10% improvement). Your own drive and model size will shift these numbers.
  • HDD as overflow storage: fine for archiving models you don't load often. Don't run your active UI from a hard drive.

CPU Requirements

Your CPU barely matters for image generation. The GPU does 95%+ of the work during inference.

  • Minimum: any modern quad-core (Intel i5/Ryzen 5 from 2018+)
  • Where CPU matters: ControlNet preprocessing, dataset preparation, and some post-processing pipelines. A faster CPU saves 1 - 2 seconds per image on those tasks.

Laptop vs Desktop

Desktop wins on every metric except portability:

  • VRAM: desktop GPUs get full VRAM allocations. Laptop RTX 4070s often have 8 GB instead of desktop's 12 GB.
  • Thermal throttling: laptops start throttling under sustained GPU loads. A 30-minute SDXL session will run slower on a laptop than specs suggest because the cooler can't keep up.
  • Power: laptop GPUs run at lower TDP (80 - 115W vs 170 - 200W for desktop equivalents).
  • Cost: a desktop RTX 3060 12 GB system can be built for ~$500 total. A laptop with equivalent VRAM starts at $1,000+.

If you must use a laptop: plug in the power adapter (battery mode halves GPU clocks), set Windows power plan to "High Performance," and verify your gen UI is using the NVIDIA GPU (not Intel integrated) via Task Manager.

Upgrade Priority Order

If you're building or upgrading specifically for local image generation, spend money in this order:

  1. GPU VRAM - the single biggest factor. Going from 8 GB to 12 GB unlocks SDXL ControlNet workflows and Flux quantizations. Going from 12 GB to 24 GB unlocks full-precision Flux.
  2. SSD - if you're still on a hard drive, a $40 NVMe gives you the biggest quality-of-life improvement per dollar.
  3. RAM to 32 GB - prevents page file thrashing when multitasking. A $50 upgrade that eliminates mysterious slowdowns.
  4. PSU - a reliable 650W+ unit prevents shutdowns under sustained GPU load. Don't pair a 450W TDP GPU with a 500W PSU.
  5. CPU - upgrade last. Almost any modern quad-core is fast enough for inference.

Bottom Line

The RTX 3060 12 GB used ($170 - $280) is the best value for local Stable Diffusion in 2026. It handles SD1.5, SDXL, and quantized Flux. The RTX 4070 12 GB ($500 - $600) is the sweet spot if you want speed and Flux FP8 with careful settings. Stop buying 8 GB cards for AI work - the $100 you save now costs you every model class released after 2024.

What to Do Next

FAQ

Is 8 GB VRAM enough for Stable Diffusion in 2026? +
For SD1.5, yes - it uses ~4 GB at 512x512. For SDXL, technically yes but you have zero headroom for ControlNet or hires fix. For Flux, only Q4 GGUF quantizations fit. An RTX 3060 12 GB at $180 used is a better investment than any 8 GB card.
Can I run Stable Diffusion on CPU only? +
You can, but SD1.5 at 512x512 takes 5 - 10 minutes per image on CPU versus 3 - 5 seconds on a GPU. SDXL on CPU can exceed 30 minutes. It's useful for testing but not for iterating on prompts.
Is 12 GB VRAM enough for Flux? +
For quantized Flux (GGUF Q5 - Q8 or FP8), yes. Flux Dev at full FP16 precision needs 22 - 24 GB. The quality difference between Q5 and FP16 is smaller than you'd expect - most users won't notice it in casual generations.
Does Apple Silicon work for Stable Diffusion? +
Yes. M1/M2/M3 Macs run SD through MLX or MPS backends. An M2 with 16 GB unified memory handles SDXL. Performance is roughly 2 - 3x slower than a similarly priced NVIDIA desktop GPU. Follow Apple-specific install guides - Windows instructions won't apply.
NVMe or SATA SSD for model storage? +
NVMe loads checkpoints in 3 - 8 seconds versus 10 - 15 seconds for SATA SSD versus 30 - 40 seconds for HDD. PCIe Gen 3 NVMe is enough - Gen 4 helps ~35% but Gen 5 adds almost nothing for model loading.
Should I buy an RTX 4060 or RTX 3060 for Stable Diffusion? +
RTX 3060 12 GB used. It has 50% more VRAM than the RTX 4060 8 GB and costs less. The 4060 is faster per-clock but 8 GB VRAM is a hard ceiling for SDXL + ControlNet and Flux. VRAM matters more than speed for local AI.