SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis

By Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, Robin Rombach. International Conference on Learning Representations, 2023

5,094 citations recorded by the source corpus.

We present SDXL, a latent diffusion model for text-to-image synthesis. Compared to previous versions of Stable Diffusion, SDXL leverages a three times larger UNet backbone: The increase of model parameters is mainly due to more attention blocks and a larger cross-attention context as SDXL uses a second text encoder. We design multiple novel conditioning schemes and train SDXL on multiple aspect ratios. We also introduce a refinement model which is used to improve the visual fidelity of samples generated by SDXL using a post-hoc image-to-image technique. We demonstrate that SDXL shows drastically improved performance compared the previous versions of Stable Diffusion and achieves results competitive with those of black-box state-of-the-art image generators. In the spirit of promoting open research and fostering transparency in large model training and evaluation, we provide access to code and model weights at https://github.com/Stability-AI/generative-models

Subfields: Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/d7890d1906d95c4ae4c430b350455156d6d8aed9

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

Citations in the corpus

Janus: Decoupling Visual Encoding for Unified Multimodal Understanding and Generation (2024, neutral): “53 and an accuracy of 61 %, surpassing text-to-image generative models such as DALL-E 2 [63] and SDXL [59].”
PixArt-Σ: Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation (2024, methodology): “3, we compare PixArt-Σ with open-source models SDXL [37] and PixArt-α [5], our method enhances the realism of portraits and boosts the capacity for semantic analysis.”
Autoregressive Model Beats Diffusion: Llama for Scalable Image Generation (2024, methodology): “Reviewing the literature on image generation in the year before 2024, we identify three keys to existing advanced models [Peebles & Xie 2023; Podell et al. 2023; Xue et al. 2023; Chen et al. 2023b,c; Betker et al. 2023; Li et al. 2024; Esser et al. 2024]: 1) well-designed image compressors, 2)…”
GaussianEditor: Swift and Controllable 3D Editing with Gaussian Splatting (2023, positive): “Given I , M , and y , the process begins with generating a 2D inpainted image I M y , utilizing a 2D inpainting diffusion model as per [34].”
Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (2024, positive): “They achieve remarkably successful results in challenging tasks such as zero-shot text-to-image (Podell et al., 2023; Saharia et al., 2022; Esser et al., 2024) or text-to-video (Polyak et al., 2024; Brooks et al., 2024) generation.”
Transfusion: Predict the Next Token and Diffuse Images with One Multi-Modal Model (2024, methodology): “In fact, recent work fuses representations from multiple off-the-shelf encoders to enhance performance [Podell et al., 2023, Esser et al., 2024b].”
T2I-Adapter: Learning Adapters to Dig out More Controllable Ability for Text-to-Image Diffusion Models (2023, methodology): “In addition, we compare our T2I-Adapter and ControlNet on larger SD i.e. , SDXL (Podell et al. 2023), which has 2.6B parameters.”
Qwen-Image Technical Report (2025, neutral): “Image generation models—encompassing both text-to-image generation (T2I) (Rombach et al., 2021; Wu et al., 2022; Liang et al., 2022; OpenAI, 2023; Podell et al., 2023; Chen et al., 2024c; Li et al., 2024b; Esser et al., 2024; BlackForest, 2024; Gao et al., 2025; Gong et al., 2025; Cai et al., 2025)…”
Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets (2023, methodology): “Borrowing from large-scale image model training [13, 64, 66], we introduce a systematic approach to curate video data at scale and present an empirical study on the effect of data curation during video pretraining.”
Improved Distribution Matching Distillation for Fast Image Synthesis (2024, methodology): “Our generators are trained by distilling SDXL [57] and SD v1.”
Exploiting Diffusion Prior for Real-World Image Super-Resolution (2023, neutral): “Given the 16 LR images with the three”
VBench: Comprehensive Benchmark Suite for Video Generative Models (2023, methodology): “As shown in Figure 6 (a), T2V models significantly underperform in Multiple Objects and Spatial Relationship compared to T2I models (especially SDXL [81]), which highlights the need to enhance compositionality ( i.e ., correctly composing multiple objects in the same frame).”
BLINK: Multimodal Large Language Models Can See but Not Perceive (2024, methodology): “Specifically, we first generate synthetic images using Stable Diffusion XL [63], employing COCO captions [46] as prompts.”
Zero123++: a Single Image to Consistent Multi-view Diffusion Base Model (2023, methodology): “…SyncDreamer [13] and our Zero123++ on four input images, including one image from the Objaverse dataset with large uncertainty on the back side of the object (an electric toy cat), one real photo (extinguisher), one image generated by SDXL [15] (a dog sitting on a rocket) and an anime illustration.”
A Tale of Two Features: Stable Diffusion Complements DINO for Zero-Shot Semantic Correspondence (2023, neutral): “However, their performance differs on textureless examples.”
An Image is Worth 32 Tokens for Reconstruction and Generation (2024, positive): “After the first training stage with proxy codes, we optionally have the second “decoder fine-tuning” stage, inspired by [10, 50], to improve the reconstruction quality.”
Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models (2025, positive): “Recently, several inﬂuential text-to-image works [6, 8, 21] have explored higher-dimensional tokenizers compared to the widely adopted VAE in Stable Diffusion [31, 34], as these tokenizers offer improved detail reconstruction, enabling ﬁner generative quality.”
InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation (2023, methodology): “7B, generate 512 × 512 images, then these images are interpolated to 1024 and refined by SDXL-Refiner to get high-resolution images.”