RTX 5070 for ComfyUI: Full CUDA Image Generation at £139/mo

RTX 5070 + ComfyUI: The Short Answer

ComfyUI is a pure PyTorch application. On the RTX 5070, the setup is pip install torch and go — no additional Intel extensions, no ROCm tweaks, no SYCL backends. Everything in the ComfyUI ecosystem that’s built on standard PyTorch works out of the box. That includes the full library of custom nodes: ControlNet preprocessors, AnimateDiff, IP-Adapter, LoRA loading, GGUF loaders, upscalers, and face restoration models.

The RTX 5070’s Blackwell 5th Gen Tensor Cores run BF16 and FP16 matrix operations significantly faster than the equivalent Ada (RTX 40xx) card at the same VRAM tier. SDXL at 1024×1024 generates in ~4–5 seconds with xformers enabled.

Installation and Launch

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

# Install PyTorch with CUDA 12.8 support (Blackwell requires CUDA 12.4+)
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128

# Install xformers for faster attention
pip install xformers

# Install ComfyUI dependencies
pip install -r requirements.txt

# Launch with xformers enabled
python main.py --use-xformers --listen 0.0.0.0 --port 8188

Blackwell GPUs require CUDA 12.4 or higher. If you’re installing on a fresh GigaGPU server, the CUDA driver version will already support Blackwell — use the cu128 wheel index for PyTorch to get the latest CUDA 12.8 build.

To also enable FlashAttention 3 (fastest attention path on Blackwell):

pip install flash-attn --no-build-isolation
python main.py --use-flash-attention --listen 0.0.0.0 --port 8188

Custom Node Compatibility

All custom nodes that compile CUDA extensions will work on RTX 5070 provided they require CUDA 12.4 or earlier. Some older nodes compiled for sm_86 (Ampere) may need a recompile for sm_120 (Blackwell) — most maintainers have updated for Blackwell by mid-2026.

SDXL Workflow

SDXL at ~8 GB FP16 fits comfortably with 4 GB of headroom on the RTX 5070. A standard SDXL workflow in ComfyUI:

• Checkpoint loader → SDXL base (e.g. sdxl_base_1.0.safetensors)
• CLIP Text Encode ×2 → positive + negative prompts
• Empty Latent Image → 1024×1024
• KSampler → Euler a, 30 steps, CFG 7
• VAE Decode → final image

For SDXL + refiner: load the refiner checkpoint in a second KSampler node. The combined memory at runtime (~14 GB) exceeds 12 GB — enable model offloading in ComfyUI (--gpu-only is the default; switch to --normalvram to allow CPU offload of the inactive model). Speed drops slightly but it works.

Single ControlNet with SDXL: ~11 GB total — fits with 1 GB spare. Use --normalvram to give ComfyUI flexibility to manage the budget.

Flux.1 Workflow

See the dedicated RTX 5070 for Flux.1 guide for the full setup. The ComfyUI workflow summary:

1. Install ComfyUI-GGUF custom nodes
2. Load flux1-schnell-Q4_K_S.gguf via GGUF Loader node
3. Load t5xxl_fp8_e4m3fn.safetensors + clip_l.safetensors via DualCLIPLoader
4. Load ae.safetensors VAE via VAELoader
5. Wire to FluxGuidance → KSampler (4 steps, Euler) → VAEDecode

Add --lowvram flag to ComfyUI launch command to reduce peak memory pressure during Flux.1 Dev generation.

Generation Speed

Managing 12 GB

Practical tips for staying within 12 GB in ComfyUI:

• Launch with --normalvram: allows ComfyUI to move inactive models to CPU RAM, preventing OOM when switching between checkpoints.
• Use ModelMemoryManagement node: explicitly unload models between workflow segments.
• Use FP8 checkpoints where available: many SDXL checkpoints have FP8 variants that cut VRAM from ~8 GB to ~4 GB with minimal quality loss.
• Tiled VAE decode: for resolutions above 1024×1024, enable tiled VAE to process the decode in chunks — saves ~2–3 GB peak at 1536×1536.
• Avoid loading multiple large LoRAs simultaneously: each LoRA is small (~0.1–0.3 GB) but they accumulate.