Best GPUs for Running FLUX.1 Locally

The Short Answer: How Much VRAM Does FLUX.1 Need?

Black Forest Labs describes both FLUX.1 [schnell] and FLUX.1 [dev] as 12-billion-parameter rectified-flow transformers. The official BF16 checkpoint files are about 23.8GB before the text encoders, VAE, activations, and other runtime memory are considered. That makes 24GB a practical starting point for running the original weights with fewer compromises, not a universal promise that every workflow will remain entirely in VRAM.

Smaller cards can still run FLUX.1. Quantized weights reduce model memory, and CPU offloading moves inactive components into system RAM. Those techniques are useful, but they trade simplicity and usually speed for lower VRAM use. A 12GB or 16GB card should therefore be evaluated as part of a complete workflow—model precision, resolution, batch size, ControlNet or LoRA usage, and available system RAM all matter.

FLUX.1 Schnell vs. Dev: The Important Differences

FLUX.1 [schnell] is the more approachable local option because Black Forest Labs designed it for one to four sampling steps and released it under Apache 2.0. FLUX.1 [dev] targets higher-quality experimentation and uses a separate non-commercial model license. The official model card says generated outputs may be used for personal, scientific, and commercial purposes subject to that license, but the model weights themselves are not offered under Apache 2.0.

What Matters Most When Choosing a GPU

1. Usable VRAM and workflow complexity

VRAM determines how much of the pipeline can stay on the GPU. Additional conditioning models, LoRAs, upscalers, higher resolutions, and larger batches increase memory pressure. If two cards have similar pricing, the card with more VRAM gives you more flexibility before components must be offloaded.

2. Software support

NVIDIA remains the easier route for many local image-generation users because CUDA is the default target for PyTorch and many ComfyUI nodes. AMD officially supports the RX 9060 XT in current ROCm compatibility documentation, but supported operating systems and framework versions must match the matrix. Check the current matrix before buying or upgrading drivers.

3. System RAM and storage

CPU offloading shifts memory pressure rather than eliminating it. A lower-VRAM build benefits from generous system RAM and fast NVMe storage, especially when loading multiple large model components. Offloading can make a workflow fit, but it is not equivalent to keeping the complete pipeline in GPU memory.

Reviewed GPU Comparison

The figures above are manufacturer specifications, not FLUX.1 generation benchmarks. Board-partner clocks, power limits, cooling, model precision, ComfyUI version, and workflow design can all change real generation time. We do not rank these cards by seconds per image without a controlled, reproducible FLUX.1 test run.

Best Reviewed GPU for the Easiest Setup

GIGABYTE RTX 5070 WINDFORCE OC 12G

Choose the GIGABYTE RTX 5070 WINDFORCE OC 12G if your priority is broad CUDA compatibility and a straightforward ComfyUI setup. It has 12GB of GDDR7 memory on a 192-bit interface, producing 672 GB/s of theoretical bandwidth. NVIDIA lists 6,144 CUDA cores and a 250W reference total graphics power for the RTX 5070.

Its limitation is capacity: the official FLUX.1 BF16 checkpoint is larger than the card's VRAM. Plan on a quantized model, component offloading, or both. This is a good choice for users who accept those optimizations and value software compatibility over maximum memory headroom.

Best Reviewed GPU for More VRAM

GIGABYTE RX 9060 XT GAMING OC 16G

Choose the GIGABYTE RX 9060 XT GAMING OC 16G when 16GB of VRAM is more valuable to you than the CUDA ecosystem. AMD specifies 16GB of GDDR6, a 128-bit interface, up to 320 GB/s of memory bandwidth, and 160W typical board power for the reference RX 9060 XT.

The extra 4GB can reduce offloading in some quantized workflows, but 16GB still does not hold the complete official BF16 pipeline. AMD lists the RX 9060 XT in its ROCm support documentation; verify the exact OS, driver, PyTorch, and ROCm combination you intend to use. This recommendation is strongest for users comfortable validating and maintaining that software stack.

Best Reviewed GPU for a Compact Build

ASUS Prime RTX 5070 SFF-Ready 12GB

The ASUS Prime RTX 5070 SFF-Ready is the compact option in the reviewed group. ASUS specifies a 2.5-slot, 304 × 126 × 50mm card with 12GB GDDR7, 6,144 CUDA cores, and a 192-bit memory interface. It has the same fundamental 12GB FLUX.1 limitation as other RTX 5070 cards, so the reason to choose it is case compatibility rather than extra model capacity.

When a 24GB-Class GPU Is the Better Choice

If you want fewer compromises with the original BF16 weights, frequent high-resolution work, multiple conditioning models, or large batches, a 24GB-class GPU is the safer target. It still may not keep every component and activation in VRAM at every setting, but it gives the 23.8GB transformer checkpoint far more room than a 12GB or 16GB card.

The AI Desk does not currently have a verified 24GB FLUX.1 benchmark in its reviewed catalog, so this guide does not name a speed winner in that class. Treat any purchase decision as workload-specific and look for tests that disclose model version, precision, resolution, steps, batch size, software version, and whether CPU offload was enabled.

Memory-Saving Techniques That Actually Change the Fit

• ▸Quantize the transformer: FP8 or lower-bit community variants can substantially reduce weight memory, with possible quality or compatibility trade-offs.
• ▸Use model or sequential CPU offload: Diffusers can move inactive components to system RAM, reducing VRAM use at the cost of transfer overhead.
• ▸Keep batch size at one: batching increases activation memory quickly.
• ▸Reduce resolution while testing: validate the workflow before moving to larger outputs.
• ▸Add LoRAs and ControlNet nodes one at a time: measure memory after each component instead of assuming the base model result will hold.

Who Should Not Buy These Cards

• ▸Do not buy a 12GB RTX 5070 if you expect the full official BF16 FLUX.1 pipeline to remain entirely in VRAM.
• ▸Do not buy the RX 9060 XT solely for AI workloads before confirming that your intended operating system and framework versions appear in AMD's current ROCm support matrix.
• ▸Do not choose from theoretical bandwidth alone; capacity and software compatibility often decide whether the workflow is usable.
• ▸Do not rely on affiliate-page generation-time claims that omit model precision, steps, resolution, and offload settings.
• ▸Do not pay a compact-card premium unless your case genuinely needs the smaller form factor.

Verdict

For the simplest path among the GPUs reviewed here, choose the RTX 5070 WINDFORCE OC 12G and plan on quantization or CPU offload. For more VRAM headroom, choose the RX 9060 XT 16G only after confirming a supported ROCm configuration. For a compact CUDA build, choose the ASUS Prime RTX 5070 SFF-Ready.

If your goal is the original BF16 checkpoint with fewer memory-management compromises, skip this 12GB–16GB tier and shop for a 24GB-class GPU. The honest answer is not that one reviewed card “wins” every FLUX.1 benchmark; it is that each card makes a different trade between VRAM, software support, form factor, and workflow complexity.

Sources and Verification Notes

• ▸Black Forest Labs: FLUX.1 [schnell] model card
• ▸Black Forest Labs: FLUX.1 [dev] model card
• ▸NVIDIA: GeForce RTX 5070 official specifications
• ▸AMD: Radeon RX 9060 XT official specifications
• ▸AMD: current ROCm compatibility matrix
• ▸Hugging Face Diffusers: reducing memory usage