# SDXL

Source: https://docs.prodia.com/models/sdxl/

SDXL (Stable Diffusion XL) is the model to reach for when you need fast, cheap image generation with fine-grained control. It produces 1024x1024 images in under two seconds and costs a fraction of a cent per image, making it well-suited to high-volume workflows, experimentation, and applications where iteration speed matters more than absolute fidelity. It also exposes more knobs than most newer models — explicit `steps`, `guidance`, `seed`, and `negative_prompt` parameters give you direct control over the diffusion process.

### Architecture

SDXL is a latent-diffusion model built around a U-Net (\~2.6B parameters) operating in the latent space of a VAE. It uses *two* text encoders in parallel — OpenAI's CLIP ViT-L and the larger OpenCLIP ViT-bigG — and concatenates their outputs to form a richer text embedding (\~3.5B total parameters across the pipeline). This dual-encoder setup is the main reason SDXL produces noticeably better composition and prompt adherence than its 1.5 predecessor while remaining small enough to run on a single consumer GPU.

The model includes an optional refiner stage — a second U-Net specialised for the final denoising steps. Enabling `refiner: true` improves fine detail (skin texture, foliage, fabric) at the cost of extra latency.

### When to use SDXL

- *High-volume generation* — sub-2s ETAs and \~$0.002 per image make SDXL viable for batch jobs, A/B tests, and user-facing prototypes
- *Style-controlled output* — the `style_preset` parameter applies one of 17 baked-in styles (anime, photographic, neon-punk, line-art, etc.) without hand-crafting prompts
- *Negative prompting* — explicit `negative_prompt` support is rare among newer models; useful when you need to exclude specific subjects, artifacts, or qualities
- *Image editing without retraining* — `img2img` and `inpainting` variants share the same backbone, so style and quality stay consistent across the workflow
- *Reproducibility-critical work* — the `seed` parameter combined with deterministic `steps` makes SDXL easy to use for caching, regression tests, and side-by-side comparisons

For higher-fidelity output and stronger prompt following, consider [FLUX.2](https://docs.prodia.com/models/flux-2/). For text-in-image workloads, use [Recraft V4](https://docs.prodia.com/models/recraft-v4/).

### Job types

| Job type                       | Description                                    | ETA    |
| ------------------------------ | ---------------------------------------------- | ------ |
| `inference.sdxl.txt2img.v1`    | Generate an image from a text prompt           | \~1.8s |
| `inference.sdxl.img2img.v1`    | Transform an existing image guided by a prompt | \~1.5s |
| `inference.sdxl.inpainting.v1` | Replace a masked region of an image            | \~1.7s |

### Parameters

*Common to all job types:*

- `prompt` (required) — text description, 3–1024 characters
- `negative_prompt` — qualities or subjects to avoid, 3–1024 characters
- `style_preset` — one of: `3d-model`, `analog-film`, `anime`, `cinematic`, `comic-book`, `craft-clay`, `digital-art`, `enhance`, `fantasy-art`, `isometric`, `line-art`, `low-poly`, `neon-punk`, `origami`, `photographic`, `pixel-art`, `texture`
- `guidance` — classifier-free guidance scale, default `8`. Lower values (5–7) give more creative results, higher values (9–12) follow the prompt more strictly
- `seed` — integer for reproducible output; omit for random
- `steps` — diffusion steps, 1–1000, default `25`. 20–30 is the sweet spot
- `refiner` — boolean, default `false`. Adds the SDXL refiner stage for sharper detail
- `conditioner` — boolean, default `false`. Enables additional conditioning signals
- `noise` — noise scale, default `1`

*txt2img only:*

- `width`, `height` — output dimensions, 512–1536, default `1024`. SDXL was trained at 1024x1024 and works best at that resolution

*img2img and inpainting:*

- `strength` — how much the input image is altered, 0–1, default `0.6`. Lower values stay closer to the input

*Inputs:*

- `img2img` and `inpainting` accept an input image (PNG, JPEG, or WebP, 256x256 to 1536x1536, max 10 MB)
- `inpainting` additionally requires a mask image — white pixels are regenerated, black pixels are preserved

### Prompting tips

- *Lead with the subject:* SDXL pays more attention to the start of the prompt. Put the main subject and key adjectives first, scene details after
- *Combine style preset with prompt:* `style_preset` sets the overall aesthetic — your prompt can still describe subject and composition. Prefer the preset over hand-written style words like "in the style of anime"
- *Use negative prompts surgically:* terms like `blurry, low quality, distorted, watermark, text` are reliable defaults. Avoid stuffing too many terms or you'll wash out the signal
- *Tune `guidance` for tone:* drop to `6` for more creative variation, raise to `10` for tighter prompt adherence. Above `12` images often look over-saturated
- *Stick to 1024x1024:* SDXL was trained at 1024 square. Other supported sizes work but may show composition drift, especially at the extremes (512 or 1536)
- *Enable the refiner for portraits and close-ups:* `refiner: true` notably improves skin, eyes, and fine textures. It adds latency, so leave it off for thumbnails and previews

### Examples

Text to image with a style preset and seed:

```json
{
  "type": "inference.sdxl.txt2img.v1",
  "config": {
    "prompt": "a majestic snow leopard perched on a frozen mountain ledge, sweeping vista of snowy peaks behind, golden hour, sharp focus",
    "negative_prompt": "blurry, low quality, distorted, painting",
    "style_preset": "photographic",
    "width": 1024,
    "height": 1024,
    "seed": 42,
    "refiner": true
  }
}
```

Image-to-image transformation with low strength to preserve composition:

```json
{
  "type": "inference.sdxl.img2img.v1",
  "config": {
    "prompt": "a watercolour painting of the same scene, soft pastel tones, visible brush strokes",
    "style_preset": "analog-film",
    "strength": 0.45,
    "guidance": 7
  }
}
```

Inpainting a masked region:

```json
{
  "type": "inference.sdxl.inpainting.v1",
  "config": {
    "prompt": "a vase of fresh sunflowers on the table",
    "negative_prompt": "blurry, distorted",
    "strength": 0.85,
    "steps": 30
  }
}
```

### Guides

[Generating Images](https://docs.prodia.com/guides/generating-images/) — Step-by-step guide for generating images with Prodia.

[Transforming Images](https://docs.prodia.com/guides/transforming-images/) — Use img2img to transform an existing image.