Question 1

What is SDXL?

Accepted Answer

SDXL (Stable Diffusion XL 1.0) is Stability AI's 3.5B parameter open-weight image generation model, released July 2023. It was the first open-source model to generate natively at 1024×1024 and established the standard that FLUX and SD 3.x later built upon. It uses a UNet architecture with dual CLIP text encoders and has the largest community ecosystem of any image model — 50,000+ LoRAs and 10+ ControlNet types.

Question 2

How much does SDXL cost on Runflow?

Accepted Answer

SDXL costs approximately $0.003 per image (at $0.00125/compute second, ~2–3 seconds per 1024×1024 image). The Lightning variant is even cheaper at under $0.001/image. SDXL is one of the most affordable image generation models available via API.

Question 3

How does SDXL compare to SD 3.5 Large and FLUX.1 [dev]?

Accepted Answer

SDXL has the largest LoRA/ControlNet ecosystem and most effective negative prompts. SD 3.5 Large (8.1B) has better typography (91% legibility) and uses the newer MMDiT architecture. FLUX.1 [dev] (12B) has the best photorealism and prompt adherence. SDXL is the cheapest and fastest of the three, with the lowest self-hosting requirements (7–8 GB VRAM vs 18 GB for SD 3.5 and 24 GB for FLUX).

Question 4

Can I use custom LoRA weights?

Accepted Answer

Yes. Pass LoRA weights via the loras parameter with a URL to the weights file and a scale (0–4). You can stack multiple LoRAs per generation. SDXL has 50,000+ community LoRAs on Civitai covering styles, characters, concepts, and domains.

Question 5

What is SDXL Lightning?

Accepted Answer

SDXL Lightning (by ByteDance) is a distilled variant that generates 1024×1024 images in just 1–4 steps — under 1 second on A100 GPUs. It maintains near-full SDXL quality with 10–25x faster generation. Available as a separate endpoint on Runflow.

Question 6

Can I use SDXL commercially?

Accepted Answer

Yes. SDXL is released under the CreativeML Open RAIL++-M license, which permits commercial use with no revenue threshold (unlike SD 3.5's $1M limit). You can also self-host — it runs on consumer GPUs with just 7–8 GB VRAM.

Question 7

Do I need to manage GPUs?

Accepted Answer

No. Runflow handles all infrastructure, including GPU provisioning, scaling, and failover. You call our API and receive results — no DevOps required.

Use Case	Assessment	Notes
Portrait	Pending	Portrait quality
Fashion	Pending	Fashion quality
Product	Pending	Product quality
Creative	Pending	Creative quality

Parameter	Type	Required	Description
prompt	string	required	Text description of the image. SDXL works well with both natural language and keyword-style prompts. Dual CLIP encoders handle complex compositional descriptions.
negative_prompt	string	optional	Highly effective on SDXL — describe what to exclude. Common: 'blurry, low resolution, deformed hands, watermark'.
image_size	object \| string	optional	Output dimensions. Presets: square_hd, square, landscape_4_3, landscape_16_9, portrait_4_3, portrait_16_9. Or custom {width, height}.
num_images	integer	optional	Number of images to generate per request. Range: 1–8.
num_inference_steps	integer	optional	Denoising steps. 20–25 for standard quality, 30–50 for maximum fidelity. Lightning variant uses 1–4 steps. Default: 25.
guidance_scale	number	optional	Classifier Free Guidance. 7–9 recommended for SDXL. Higher values (10–15) increase prompt adherence. Default: 7.5.
seed	integer	optional	Reproducibility seed. Same seed + same parameters = identical output.
loras	array[object]	optional	Array of LoRA weights to apply. Each object includes a URL to the weights and a scale (0–4). Stack multiple LoRAs for combined effects.
embeddings	array[object]	optional	Textual inversion embeddings to apply during generation.
expand_prompt	boolean	optional	If true, the prompt is expanded with additional detail via an LLM for improved results.
format	string	optional	Output image format. Options: "jpeg", "png". Default: "jpeg".
enable_safety_checker	boolean	optional	Enable content safety filter.
sync_mode	boolean	optional	If true, returns the image as a data URI.

SDXL API

SDXL family by Stability AI

SDXL

SDXL Lightning

SDXL Inpainting

SDXL Image-to-Image

SDXL ControlNet Union

SDXL Fooocus

Benchmarked for your use case

API Parameters

Technical Specifications

Why run SDXL on the Runflow API

Largest LoRA ecosystem

Mature ControlNet support

Lightning-fast variants

Effective negative prompts

Battle-tested in production

Low VRAM, easy to self-host

Explore other models

FLUX.1 [dev]

Seedream 4.5

Seedream 4

Nano Banana 2

GPT-Image 1.5

Nano Banana Pro

Frequently asked questions