Phase 2: Image Generation · 50 min · Python · Hugging Face Diffusers · Civitai

Stable Diffusion XL & SD 3.5

SDXL's ecosystem moat — thousands of LoRAs, ControlNets, and community models — is why it still matters alongside FLUX.

Hiring signal: Knowing when to use SDXL over FLUX (ecosystem, LoRAs, ControlNets) demonstrates practical production reasoning, not just 'newest is best' thinking.

What you will learn

Generate images with SDXL at 1024x1024 native resolution
Use SD 3.5 Large for improved text rendering and compositional accuracy
Understand the SDXL ecosystem moat: years of community LoRAs, checkpoints, embeddings, ControlNets
Choose between SDXL and FLUX: SDXL for custom styles via LoRA and ControlNet, FLUX for quality and speed

The Problem

A team switches entirely to FLUX.1 for image generation. Then a client says "we need images in our specific brand style." The team tries prompt engineering — it doesn't work. They try FLUX LoRAs — there are only a few dozen available. They realize SDXL has 10,000+ community LoRAs on Civitai, including one that matches the client's brand exactly. They switch back to SDXL for this use case.

FLUX has better base quality, but SDXL has an ecosystem moat built over years. Knowing when to use SDXL over FLUX demonstrates practical production reasoning, not just "newest is best" thinking.

What you'll build
Download a community LoRA from Civitai, load it with SDXL, and generate images with and without the LoRA to demonstrate style transfer. Then compare with FLUX.1 Schnell to show when SDXL+LoRA beats FLUX.

SDXL: The Ecosystem Model

Stable Diffusion XL (SDXL) was released by Stability AI in July 2023. While FLUX has better base quality, SDXL has something FLUX doesn't: years of community-built extensions.

SDXL Ecosystem

Resource	Count	Source
Community LoRAs	10,000+	Civitai
Fine-tuned checkpoints	1,000+	Civitai
ControlNet models	20+ types	Hugging Face, Civitai
Textual inversions	500+	Civitai
IP-Adapter variants	10+	Hugging Face
Community workflows	1,000+	ComfyUI

This ecosystem is SDXL's moat. If you need a specific style — anime, photorealistic, watercolor, 3D render, architectural — there's likely a LoRA for it on Civitai.

SDXL Technical Specs

Property	Value
Architecture	UNet
Native resolution	1024x1024
Parameters	~3.5B (UNet)
VRAM (fp16)	~8GB
Default sampler	DPM++ 2M Karras
Default steps	25-30
Text encoders	Dual CLIP (ViT-L + ViT-bigG)
Token limit	77
License	CreativeML Open RAIL-M

from diffusers import StableDiffusionXLPipeline
import torch

pipe = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16, variant="fp16",
)
pipe.to("cuda")

# Generate with SDXL
image = pipe(
    prompt="a majestic mountain landscape, professional photography",
    num_inference_steps=25,
    guidance_scale=7.0,
    height=1024, width=1024,
).images[0]

Unlock the full lesson

You've read the first 2 sections. The rest of this lesson covers Loading Community LoRAs, SD 3.5 Large, When SDXL Beats FLUX, Production Hybrid Approach, 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.

Key terms in this lesson

LoRA (Low-Rank Adaptation)