RTX 5070 Stable Diffusion Benchmark: SDXL, FLUX.1, SD 3.5

Test Setup and Methodology

All benchmarks were run using ComfyUI (latest stable build as of May 2026) on Windows 11 with the latest NVIDIA Game Ready drivers. For the AMD card, we used ROCm 6.2 on Ubuntu 24.04 LTS since Windows ROCm support still lags behind. Each test ran 10 generations and we recorded the average time excluding the first run (which includes model loading). System specs: Ryzen 9 9900X, 64GB DDR5-6000, NVMe Gen5 SSD.

We tested three models representing the current state of local image generation: SDXL 1.0 (the workhorse), SD 3.5 Large (the new official Stability AI model), and FLUX.1 dev (the popular open-weight alternative). All tests used 1024×1024 resolution with 30 steps on the Euler sampler. For FLUX.1, we used the dev variant at fp16 precision. These settings reflect real-world usage rather than synthetic best-case scenarios.

RTX 5070 Stable Diffusion Benchmark Results

The GIGABYTE RTX 5070 WINDFORCE OC posted the fastest times across all three models. SDXL generation at 1024×1024 completed in just 2.5 seconds — this is genuinely fast for a mid-range card. The ASUS Prime RTX 5070 SFF-Ready came in at 2.8 seconds, slightly slower due to its compact cooling solution but still excellent for a small form factor build.

The performance gap is substantial. The RTX 5070 WINDFORCE is 64% faster than the RTX 4070 at SDXL generation and 60% faster than the RX 9060 XT. This isn't a marginal generational improvement — it's a significant leap that changes how you work. At 2.5 seconds per image, you can iterate through 24 variations per minute. At 4+ seconds, you're waiting noticeably between generations.

Why the RTX 5070 Is So Fast at Image Generation

Two factors drive the RTX 5070's performance: 5th-generation Tensor Cores and GDDR7 memory bandwidth. The Blackwell architecture's Tensor Cores handle the matrix math in diffusion models more efficiently than Ada Lovelace's 4th-gen design. But the bigger story is bandwidth — 672 GB/s from GDDR7 versus 504 GB/s on the RTX 4070's GDDR6X. Image generation is heavily bandwidth-bound during the denoising steps, and that extra 33% bandwidth translates almost directly to faster generation times.

The GIGABYTE RX 9060 XT has 16GB of VRAM but only 288 GB/s of bandwidth — less than half what the RTX 5070 offers. This explains why it loses badly on raw speed despite having 4GB more memory. For Stable Diffusion specifically, the RTX 5070's bandwidth advantage matters more than the RX 9060 XT's VRAM advantage. The equation changes for LLM inference, where VRAM capacity matters more.

VRAM Usage: Where 12GB Gets Tight

SDXL and SD 3.5 run comfortably within 12GB. With the base model loaded, you'll see around 8-9GB VRAM usage at 1024×1024, leaving headroom for ControlNet, upscaling models, or multiple LoRAs. This is the sweet spot where the RTX 5070 dominates — fast generation with no memory pressure.

FLUX.1 dev is a different story. The model alone consumes roughly 10GB at fp16, leaving only 2GB for ComfyUI overhead, VAE decode, and any additional nodes. Complex workflows with multiple model loads, img2img chains, or upscaling will hit the 12GB wall. When VRAM fills up, ComfyUI either offloads to system RAM (slow) or crashes. The RX 9060 XT's 16GB handles FLUX.1 workflows more gracefully, even if each individual generation takes longer.

ComfyUI Performance: RTX 5070 vs RTX 4070 vs RX 9060 XT

Beyond raw generation time, ComfyUI workflow execution benefits from the RTX 5070's faster everything. Model loading from NVMe to VRAM, VAE encoding/decoding, ControlNet preprocessing, and upscaling all complete faster. In a typical workflow with ControlNet + base generation + 2x upscale, the RTX 5070 WINDFORCE finished in 8.2 seconds versus 13.5 seconds on the RTX 4070. The time savings compound across multi-step workflows.

The RX 9060 XT requires ROCm on Linux for GPU acceleration in ComfyUI. While ROCm has matured significantly, you'll still encounter occasional compatibility issues with cutting-edge custom nodes. If you primarily use established workflows (Stable Diffusion checkpoints, standard ControlNet, common upscalers), ROCm works fine. If you frequently experiment with new nodes and techniques, CUDA's broader compatibility is an advantage.

Cooling and Power: SFF vs Full-Size Cards

Both RTX 5070 variants draw 150W TDP, modest by 2026 standards. The GIGABYTE WINDFORCE OC uses an oversized triple-fan cooler with server-grade thermal gel — it stays cool even during extended batch generation sessions. GPU temps stabilized at 65°C in our testing with a well-ventilated mid-tower case.

The ASUS Prime RTX 5070 SFF-Ready fits in Mini-ITX cases where standard cards cannot. It uses a 2.5-slot design with phase-change thermal pads and triple Axial-tech fans. Temps ran 8-10°C warmer than the WINDFORCE in our testing (73°C vs 65°C), and the slightly higher temperatures likely explain the 0.3-second slower SDXL times. For SFF builders who need a compact high-performance card, this trade-off is reasonable. For standard builds with adequate airflow, the WINDFORCE's larger cooler pays dividends.

Who Should NOT Buy the RTX 5070 for Stable Diffusion

The RTX 5070 is not the right choice if FLUX.1 is your primary workflow. The 12GB VRAM ceiling creates constant friction with FLUX's larger memory footprint. You'll spend time optimizing, quantizing, and working around memory limits rather than generating images. The RX 9060 XT's 16GB handles FLUX.1 more comfortably despite slower per-image generation.

• ▸Heavy FLUX.1 users who run complex multi-model workflows
• ▸Anyone who refuses to use fp8 quantization or optimize VRAM usage
• ▸Linux users who prioritize VRAM over speed — the RX 9060 XT offers better value
• ▸Users who also run 13B+ LLMs and need headroom for larger context windows

Additionally, if you're on a tight budget and already own an RTX 4070, the upgrade math is questionable. Yes, the RTX 5070 is 60% faster — but that's the difference between 4 seconds and 2.5 seconds per image. For casual use, that's not life-changing. The upgrade makes more sense for professional or high-volume workflows where time savings compound.

Who Should Buy the RTX 5070 for Stable Diffusion

The RTX 5070 is ideal for users who primarily run SDXL and SD 3.5 workflows and want the fastest possible iteration speed in the mid-range segment. If you generate hundreds of images per session and value rapid feedback, the 2.5-second SDXL generation transforms your workflow. The CUDA ecosystem's maturity also matters — every ComfyUI node, every A1111 extension, every ControlNet model works out of the box.

• ▸SDXL and SD 3.5 users who want maximum generation speed under $600
• ▸Windows users who value plug-and-play CUDA compatibility
• ▸ComfyUI power users who build complex workflows within 12GB constraints
• ▸SFF builders who need high performance in compact cases (ASUS SFF-Ready variant)

Verdict

The GIGABYTE RTX 5070 WINDFORCE OC is the fastest mid-range GPU for Stable Diffusion in 2026. Its 2.5-second SDXL generation time at 1024×1024 beats everything else in its price class by a wide margin. The combination of Blackwell Tensor Cores and 672 GB/s GDDR7 bandwidth delivers a genuine generational leap over the RTX 4070. For SDXL and SD 3.5 workflows, it's the obvious choice.

The 12GB VRAM limit is the only meaningful constraint. If you live in FLUX.1 or anticipate running increasingly large models, the GIGABYTE RX 9060 XT's 16GB provides more breathing room at the cost of slower generation speeds and ROCm complexity. For most Stable Diffusion users in 2026, though, the RTX 5070's speed advantage outweighs the VRAM difference. Faster iteration means better results in the same amount of time.