Black Forest Labs FLUX.1 [dev] VRAM Requirements Guide
State-of-the-art open diffusion transformer. Combines a 12B MMDiT core with T5-XXL and CLIP-L text encoders for photorealistic generation.
>98% perplexity retention. Standard Ollama/llama.cpp quantization.
Virtually identical to FP16 (~99.9% fidelity). Demands high-capacity VRAM.
Uncompressed weights. Requires multi-GPU clusters or datacenter cards.
Architectural Memory Breakdown & KV-Cache Scaling
Unlike autoregressive LLMs, FLUX.1 [dev] does not allocate a token KV-cache; instead, memory is dominated by three distinct components: the 12B Multimodal Diffusion Transformer (`MMDiT`) flow backbone (~23.8 GB in BF16, ~11.9 GB in FP8, or ~6.8 GB in NF4/Q4_K_S), the 4.7B `T5-XXL` + `CLIP-L` dual text encoders (~9.5 GB in FP16 or ~4.9 GB in FP8), and high-resolution VAE latent decoding activations (~2.5 GB at 1024x1024). Sequential encoder offloading lets ComfyUI evict T5-XXL before denoising begins, fitting FP8 FLUX comfortably on 16GB–24GB GPUs and NF4/GGUF Q4 on 12GB cards.
Hitting OOM during the final VAE decode step in ComfyUI or Forge? Replace the standard `VAE Decode` node with `VAE Decode (Tiled)` (tile size `512`) and use `t5xxl_fp8_e4m3fn.safetensors` to prevent peak activation spikes.
Provides sufficient headroom for weights, 8k+ context KV-cache, and CUDA runtime buffers without Out-Of-Memory (OOM) crashes.
Run Black Forest Labs FLUX.1 [dev] Locally
# Run via ComfyUI: git clone https://github.com/comfyanonymous/ComfyUI# Use Diffusers pipeline: pipe = FluxPipeline.from_pretrained('black-forest-labs/FLUX.1-dev')Frequently Asked Questions
Why does unquantized BF16 FLUX.1 [dev] spill slightly on a 24GB RTX 4090?
Loading the 23.8 GB BF16 transformer plus 9.5 GB FP16 T5-XXL simultaneously exceeds 24GB; using `flux1-dev-fp8.safetensors` on RTX 4090/5080/5090 keeps everything in VRAM at maximum Ada/Blackwell FP8 speed.
Can an 8GB GPU run FLUX.1 [dev] with GGUF Q4_0?
Yes, ComfyUI-GGUF `flux1-dev-Q4_K_S.gguf` (~6.8 GB) works on 8GB cards by offloading the T5-XXL encoder entirely to system DDR4/DDR5 RAM (`--lowvram`), adding ~4–8 seconds per prompt encode.
Test Custom Hardware & Context Windows
Simulate your exact GPU, calculate token throughput speeds, or compare cloud rental break-even costs.
Open in GPU Garden Matchmaker →