SDXS-1B: compact diffusion for GPU

Compact Diffusion Model SDXS-1B: Architecture and Training

SDXS-1B is an image generation model with a 1.6 billion parameter UNet, a Qwen3.5-2B text encoder, and an asymmetric VAE (32 latent channels). It delivers high-quality generation on mid-range graphics cards. The alpha version is published under Apache-2.0 with open-source data preparation and training code.

Development began in December 2024 with experiments on Sana linear transformers. Transition to UNet in February 2025 and parallel work on DiT architectures. By December 2025, a prototype version sdxs-0.8B was trained with a UNet in SD1.5 style, Long CLIP, and a 16-channel VAE on flow matching.

Testing SDXL ideas in 2026 showed performance degradation or slowdown. Return to classic architecture. Adaptation of Flux.2 VAE: conversion to a 32-channel asymmetric variant with fine-tuning on 1 GPU over 2 days. Fine-tuning with 5 targets and normalization to preserve details.

VAE Quality Metrics

VAE sdxs-1b leads in PSNR and LPIPS among 16x models:

8x scale factor
SDXL             | MSE=1.925e-03 PSNR=30.00 LPIPS=0.123 Edge=0.181 KL=32.113
FLUX.1           | MSE=4.098e-04 PSNR=36.06 LPIPS=0.033 Edge=0.083 KL=13.127
FLUX.2           | MSE=2.425e-04 PSNR=38.33 LPIPS=0.023 Edge=0.065 KL=2.160

16x scale factor
Wan2.2-TI2V-5B (2Gb) | MSE=7.034e-04 PSNR=34.65 LPIPS=0.050 Edge=0.115 KL=9.429
sdxs-1b (200Mb)      | MSE=2.655e-04 PSNR=37.83 LPIPS=0.026 Edge=0.066 KL=2.170

Switch to Qwen3.5-2B as text encoder: embeddings from the penultimate layer for better structure. Support for multimodality and multilingualism.

SDXS-1B Architecture

The model includes three components:

• UNet (1.6 billion parameters): twice as large as SD1.5, uniform blocks for balance of anatomy and details.
• Qwen3.5-2B: surpasses CLIP/SigLIP in quality, multimodal capabilities.
• Asymmetric VAE: 32-channel latent, 8× encoder, 16× decoder. Built-in 2× upscaler for 512–768 px resolution without style loss.

The upscaler scales by 2 times, preserving details and style. Useful for tasks prioritizing accuracy: X-rays, art reproductions.

Prompt refinement: automatic improvement for tag-based queries. Support for image-to-image and experimental audio input.

Data Preparation

Dataset: 1–2 million images from Midjourney, Nijjourney (drawings, illustrations, ~10% photos). Captions: danbooru tags + descriptions up to 250 tokens.

Preprocessing:

• Scaling 768–1408 px (step 64 px) for different aspect ratios.
• Training at half the resolution.
• Text limit of 250 tokens, 10% dropout on the encoder.

Training and Hyperparameters

Phased training: low-level features → composition → details. Optimizer AdamW8bit, LR from 4e-5 to 4e-6. LR schedules: 0.5–5.0 by stages. Training on 1–8 RTX 5090, 2–3 months.

Code: open-source scripts train.py, dataset conversion. Usage via Diffusers Hugging Face.

Key Points

• Compactness: 1.6B parameter UNet + 200 Mb VAE, inference on consumer GPU.
• VAE SOTA: PSNR 37.83 at 16x, built-in 2× upscaler.
• Qwen3.5: multimodality, multilingualism without quality loss.
• Alpha version: strong in illustrations, weaker in photos (negative prompt "photo").
• Open-source: Apache-2.0, code on Hugging Face.

Limitations and Prospects

Limitations: dataset focuses on illustrations, photorealism needs refinement. No full topic coverage.

Prospects: dataset expansion, turbo-LoRA, ControlNet, video. Community invited for fine-tuning.

— Editorial Team