Open Source

Spoiler: the best image model in the world is now free. Black Forest Labs dropped Flux Pro weights overnight with an Apache-style license and the timeline is on fire.

The Setup

Flux Pro was the paid-API-only tier of the Flux family. Better prompt adherence than Midjourney, cleaner hands than SDXL, and now it lives on HuggingFace next to its dev and schnell siblings. The diffusers integration shipped same day.

pip install -U diffusers transformers accelerate
huggingface-cli login
huggingface-cli download black-forest-labs/FLUX.1-pro

The Money Pattern

Three lines and you're generating production-grade plates. For my Aidxn Design portfolio drops I've already swapped out the Midjourney workflow — Flux Pro into Affinity Designer, retouch, ship. No more "API down" panic at 2am.

import torch
from diffusers import FluxPipeline

pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-pro",
    torch_dtype=torch.bfloat16,
).to("cuda")

img = pipe(
    "studio portrait of a brutalist concrete sculpture, soft rim light, 35mm film grain",
    guidance_scale=3.5,
    num_inference_steps=28,
    width=1536, height=1024,
).images[0]

img.save("plate_01.png")

The Catch

24GB VRAM minimum at full bf16. M4 Mac users — yes, that's me — are stuck with the 8-bit quants via mflux or waiting for the MLX port. The schnell distilled variant runs on 16GB but you give up the prompt fidelity that made Pro worth open-sourcing in the first place.

The Verdict

This is the SDXL moment for 2026. Within a week there will be Flux Pro LoRAs for every aesthetic, ComfyUI workflows wrapping it, and a Replicate template charging you for what's literally free. Pull the weights now. Train your own LoRA on your portfolio. Stop paying for image gen.