RTX 5070 for Stable Diffusion: SDXL, Flux.1, and ComfyUI on 12 GB GDDR7

RTX 5070 for Stable Diffusion: The Quick Answer

Yes — SDXL runs great, Flux.1 runs quantised, SD 1.5 is trivial. The RTX 5070 is a full-CUDA Blackwell card with 12 GB GDDR7 and ~672 GB/s memory bandwidth. For image generation, that bandwidth figure matters: VAE decode and attention cache reads are bandwidth-bound, and GDDR7’s fast per-pin speed compensates for the 192-bit bus width.

The limitation is VRAM capacity. Flux.1 Dev at full FP16 needs ~24 GB — it won’t fit. Flux.1 with a Q4 GGUF transformer compresses to ~7 GB, and the full pipeline (T5 encoder ~5 GB, CLIP ~0.5 GB, VAE ~0.3 GB, working memory at 1024×1024 ~4 GB) lands around 11–12 GB total — fitting with care. For pure SDXL workflows, 12 GB is plenty.

For the full GPU comparison for image generation see best GPU for Stable Diffusion. For a deeper Flux.1 breakdown see RTX 5070 for Flux.1.

VRAM Table: What Fits on 12 GB

ComfyUI Setup

Standard CUDA setup — no special flags needed on the RTX 5070:

git clone https://github.com/comfyanonymous/ComfyUI
cd ComfyUI

# Standard PyTorch CUDA install
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
pip install -r requirements.txt

# Launch (auto-detects CUDA GPU)
python main.py --listen 0.0.0.0 --port 8188

All standard ComfyUI custom nodes work — ControlNet preprocessors, IP-Adapter, AnimateDiff, GGUF loaders, xformers. The RTX 5070’s 5th Gen Tensor Cores are automatically used by PyTorch for BF16/FP16 matrix operations. Enable xformers for faster attention:

pip install xformers
python main.py --use-xformers --listen 0.0.0.0 --port 8188

Automatic1111 Setup

git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui
cd stable-diffusion-webui

# Standard CUDA launch
./webui.sh --xformers --listen --port 7860

The --xformers flag enables efficient attention on CUDA cards and gives ~10–20% speed improvement on SDXL. For the RTX 5070 also consider --opt-sdp-attention which uses PyTorch’s native scaled dot-product attention path (often faster on Blackwell with FlashAttention 3):

./webui.sh --opt-sdp-attention --listen --port 7860

Flux.1 on 12 GB

Flux.1 needs careful VRAM management on 12 GB. The recommended approach:

Flux.1 Schnell Q4 GGUF via ComfyUI GGUF loader:

• Transformer (Q4 GGUF): ~7 GB
• T5-XXL text encoder: ~5 GB — this is the problem. T5 alone is 5 GB.

Wait — T5 at FP16 is ~5 GB. If you load the T5 encoder in FP8 (using the t5xxl_fp8_e4m3fn.safetensors variant), it drops to ~2.5 GB. Combined with Q4 transformer (~7 GB) + CLIP (~0.5 GB) + VAE (~0.3 GB) + working memory (~2 GB at 1024×1024): ~12.3 GB. This is the approach that makes Flux.1 work on 12 GB:

• Use Q4 or Q8 GGUF transformer (via ComfyUI GGUF loader)
• Use T5 FP8 (t5xxl_fp8_e4m3fn.safetensors) instead of FP16
• Stay at 1024×1024 max resolution
• For Flux.1 Schnell: 4 steps, so working memory pressure is lower — fits more comfortably

ControlNet and IP-Adapter

On 12 GB, SDXL + single ControlNet is the safe combination:

• SDXL + depth ControlNet: ~10.5 GB ✓
• SDXL + canny ControlNet: ~10.5 GB ✓
• SDXL + IP-Adapter (standard): ~9.5 GB ✓
• SDXL + ControlNet + IP-Adapter: ~12–12.5 GB — very tight, may OOM at larger resolutions
• SDXL + 2× ControlNet: ~13–14 GB — won’t fit on 12 GB

For stacked multi-ControlNet workflows, the RTX 3090 at £159/mo or Arc Pro B60 at £129/mo (both 24 GB) are better fits.

Speed Estimates

Speed estimates based on Blackwell Xe2 compute specs and community benchmarks. Actual times vary with scheduler and exact model. The RTX 5070’s high GDDR7 bandwidth makes it fast at SDXL — faster than any 16 GB GDDR6 card in this price range.

RTX 5070 vs RTX 5060 Ti for Image Gen

For pure speed on SDXL, the RTX 5070 wins. For Flux.1 and multi-ControlNet stacking, the RTX 5060 Ti’s extra 4 GB is meaningful. See the full RTX 5070 vs 5060 Ti comparison.