FLUX.1 [schnell] (4-Step Distilled DiT) VRAM Requirements Guide
Ultra-fast 4-step distilled FLUX variant. High aesthetic speed.
>98.2% retention. 10.2 GB weights, leaving +1.5 GB headroom on NVIDIA GeForce RTX 5060 Ti 16GB.
Near-lossless 8-bit precision requiring 23.0 GB VRAM for benchmark evaluations.
Uncompressed FP16 weights taking 38.9 GB. Demands multi-GPU or datacenter accelerators.
Architectural Memory Breakdown & Attention Scaling
Black Forest Labs FLUX.1 [schnell] is a distilled 12B parameter multimodal diffusion transformer (MMDiT) optimized for rapid 4-step local image synthesis. At 4-bit precision (NF4 or GGUF Q4), the transformer weights occupy ~6.8 GB, while text encoders and VAE require 4.5 GB during sampling, fitting inside 12GB–16GB graphics cards within a 11.8 GB footprint.
In ComfyUI, pass `--lowvram` or enable sequential text encoder offload. Quantizing the T5-XXL text encoder to FP8 saves ~4.5 GB of VRAM, allowing FLUX.1 [schnell] to generate 1024x1024 images in 4 steps on 12GB cards.
NVIDIA GeForce RTX 5060 Ti 16GB provides 16.0 GB physical VRAM (+1.5 GB operational margin at 8k context).
Run FLUX.1 [schnell] (4-Step Distilled DiT) Locally
ollama run flux-1-schnell:q4_K_Mvllm serve Black Forest Labs/flux-1-schnell --max-model-len 8192 --gpu-memory-utilization 0.95Frequently Asked Questions
How fast is FLUX.1 [schnell] compared to FLUX.1 [dev]?
FLUX.1 [schnell] requires only 4 sampling steps (versus 28–50 steps on [dev]), rendering identical-resolution images ~6x faster with minimal detail difference.
Can FLUX.1 [schnell] run on an 8GB GPU?
With GGUF Q4 weights and CPU offload enabled in ComfyUI, it can run on 8GB GPUs, though generation time increases from 6 seconds to ~30 seconds.
Test Custom Hardware & Context Windows
Simulate FLUX.1 [schnell] (4-Step Distilled DiT) across all 31 GPUs, adjust context up to 512 tokens, or calculate RunPod break-even costs.
Open FLUX.1 [schnell] (4-Step Distilled DiT) in GPU Garden Matchmaker →