Phase 2: Image Generation · 50 min · Python · fal.ai API · Hugging Face Diffusers

The FLUX Model Family

FLUX.1 Pro for quality, FLUX.1 Dev for open-weights flexibility, FLUX.1 Schnell for speed — knowing which to use when is image generation engineering.

Hiring signal: FLUX model selection reasoning (Pro vs Dev vs Schnell) is a standard question in generative media interviews — it tests whether you understand the speed-quality-cost trade-off.

What you will learn

Generate images with FLUX.1 Pro (API, highest quality), FLUX.1 Dev (open weights, non-commercial), and FLUX.1 Schnell (open weights, 1-4 step, Apache 2.0)
Choose the right FLUX variant for speed, quality, or cost
Understand FLUX.2 improvements: better prompt adherence and text rendering
Compare FLUX vs SDXL: FLUX for quality and speed, SDXL for ecosystem (LoRAs, ControlNets)

The Problem

A team uses FLUX.1 Pro for every image generation. At 10,000 images/day, they're spending $500/day on API calls ($0.05/image). They could use FLUX.1 Schnell at $0.003/image ($30/day) for 90% of their use cases and reserve Pro for hero images. But they don't know the difference between the variants, so they overpay by 16x.

FLUX model selection is a standard interview question because it tests whether you understand the speed-quality-cost trade-off — the core engineering decision in image generation.

What you'll build
Generate the same prompt with FLUX.1 Schnell, FLUX.1 Dev, and FLUX.1 Pro. Measure generation time, cost, and compare output quality. Write a decision guide for when to use each variant.

The FLUX Model Family

FLUX.1 is developed by Black Forest Labs (BFL), founded by the original Stable Diffusion creators. There are three main variants:

FLUX.1 Schnell

The speed variant — distilled for 1-4 step generation.

Property	Value
Architecture	DiT (Diffusion Transformer)
Steps	1-4 (distilled)
Guidance	None needed (guidance_scale=0.0)
Speed	~2 seconds (1024x1024)
Cost (fal.ai)	$0.003/image
License	Apache 2.0 (commercial use)
Weights	Open on Hugging Face
Best for	Bulk generation, rapid prototyping, pipelines

# FLUX Schnell via fal.ai
import fal_client
result = fal_client.subscribe("fal-ai/flux/schnell", arguments={
    "prompt": "a cat on a windowsill",
    "num_inference_steps": 4,
    "guidance_scale": 0.0,
    "image_size": {"width": 1024, "height": 1024},
})

FLUX.1 Dev

The quality variant — open weights for self-hosting.

Property	Value
Architecture	DiT
Steps	20-50 (standard diffusion)
Guidance	3.5 (CFG needed)
Speed	~8 seconds (1024x1024)
Cost (fal.ai)	$0.025/image
License	Non-commercial (research only)
Weights	Open on Hugging Face
Best for	Research, self-hosting experiments, quality comparison

FLUX.1 Pro

The production quality variant — API only.

Property	Value
Architecture	DiT (larger than Dev)
Steps	20-50
Guidance	3.5
Speed	~12 seconds (1024x1024)
Cost (fal.ai)	$0.05/image
License	Commercial (via API)
Weights	Closed (API only)
Best for	Hero images, print quality, final production output

FLUX.2

The next generation (announced 2026):

Better prompt adherence (especially long, complex prompts)
Improved text rendering in images
Better human anatomy (fewer extra fingers)
Faster distilled variant

You need to generate 5,000 product images/day for an e-commerce site. Which FLUX variant should you use?

FLUX.1 Schnell is the right choice: $0.003/image × 5000 = $15/day (vs $250/day for Pro), 4 steps (~2s/image), and Apache 2.0 license allows commercial use. Product images don't need Pro-level quality — Schnell is more than sufficient. Reserve Pro for hero/marketing images where quality matters more than cost.

Unlock the full lesson

You've read the first 2 sections. The rest of this lesson covers Decision Framework, FLUX vs SDXL: When to Use Each, Key Takeaways, What's Next — plus a hands-on lab, quiz, and project artifact.

Create a free account to unlock Phase 0 and Phase 1 of every course — no credit card.