Daydream

Daydream

World Models & Interactive Video

Introducing SDXL + IPAdapters

Introducing SDXL + IPAdapters

Introducing SDXL + IPAdapters

Open-source real-time AI video just leveled up.

Today, we’re publicly launching SDXL for StreamDiffusion, bringing the most advanced Stable Diffusion model into a highly controllable, open-source, real-time workflow.

This release delivers HD real-time video generation at over 15 FPS on the Daydream platform, with optimized configurations reaching up to 25 FPS.

And because it is fully open source, anyone can extend, remix, and build on it through the Daydream API or our StreamDiffusion fork.

Creators like DotSimulate are already building SDXL-powered tools and workflows on Daydream, showing what is possible when next-generation models meet real-time performance.


Quality You Can Control

Our open-source stack combines multiple research tracks into one cohesive, production-ready system that enables creators to fine-tune every aspect of image quality, temporal stability, and style.

SDXL

  • 3.5× larger model with expanded attention blocks for richer visuals
  • Native 1024×1024 resolution for sharper, more detailed HD frames
  • Reduced flicker and artifacts for improved frame-to-frame consistency

Image-Based Style Control (IPAdapters)

IPAdapters (Image-Prompt Adapters) let you guide your video’s look and feel using any reference image. They function similarly to LoRAs but allow real-time adjustability.

Two modes:

  • Standard: Apply and blend artistic styles across your video stream
  • FaceID: Maintain character identity and consistency throughout sequences

Technical features:

  • 7+ Temporal Weight Types such as Linear, Ease-in/out, Style Transfer, and Composition
  • Runtime Parameter Tuning to adjust IPAdapter influence live while generating

Fine-Tune Spatial Consistency (Multi-ControlNet Support)

Accelerated HED, Depth, Pose, Tile, and Canny ControlNets give you precise control over spatial and compositional details. You can combine multiple ControlNets in one workflow and adjust the strength of each.

TensorRT Acceleration

  • Real-time inference through NVIDIA TensorRT optimization
  • Stable 15 to 25 FPS performance even with complex multi-model setups

SD1.5 + IPAdapters

Prefer the classic SD1.5 model? We’ve paired it with accelerated IPAdapters for smooth, high-framerate style transfer.


Start Building

Explore the Playground and our API docs, join the Daydream community on Discord, and start creating your own real-time AI video workflows today.