Oct 26, 2023

Replay v0.2 is here

One month ago, we announced Replay v0.1: a photorealistic, buttery smooth text-to-video generative model that lets you create photorealistic videos from simple text prompts. Today, we’re announcing Replay v0.2: an improved model with better quality and more controllability.

The Genmo Replay Journey

Genmo's most recent product is Replay, a new, high-quality AI video generator available to the public for free. Replay is available at https://genmo.ai with an easy to use and fast interface. And since launching v0.1, users have responded in an incredible way.

Still, generative AI video is in its infancy, and reality sets a very high standard for our research. We’re a tiny team of academics and artists, but we move quickly to improve. Replay is one of the fastest improving AI models today. Here's a brief recap of our recent progress:

  • September 21st: Launched Replay publicly for the first time [demo, blog]

  • September 22nd: Introduced Camera Controls to provide more flexibility to users [demo, blog]

  • September 26th: Added a Replay FX preset library to apply visual filters and effects [demo]

  • October 2nd: Expanded the FX library, with FX Pack 2 [demo]

  • October 5th: Prompt suggestions [tutorial]

  • October 6th: Real-time streaming video previews so you can see videos as they render [demo]

  • October 26th: Replay v0.2 with image prompts, 2.7x resolution and 3x length

What’s New in Replay v0.2

Replay v0.2 has improved generation capabilities and quality. In addition, the model has a new mechanism for control via image prompts.

Image prompts: With our latest model update, users can transform an image into a video. Replay now supports a single visual style reference, to generate a matching video. While text prompts are incredibly powerful, it's often difficult to communicate a particular shot layout and style. Image uploads allow finer control over video composition and content than is possible with words alone. After you drag and drop an image or photo created in any tool, Replay will convert the image into a representation that the model can understand and use it to guide generation. This was a top-requested feature from our community, so we're delighted to open up access immediately.

Higher resolution & improved quality: Video clips created with Replay v0.2 have over 2.7x the pixels of Replay v0.1. Resolution is now consistently greater than 2K, whereas v0.1 created in 720p-1080p. Users can expect fewer AI-generated artifacts like distorted faces, blurry details or saturated results. Results will be crisper, more detailed, and have sharper lines throughout.

Longer videos: Generate up to 6 seconds of video, 3x longer than Replay v0.1's limit of 2 seconds. With longer videos, you can tell more complete stories. Users have the option to select between 2 second, 4 second and 6 second videos in the settings menu beneath the prompt box on our homepage.

Batch mode: Queue up and create up to 3 videos in parallel with one click. No need to wait for each video to finish before starting the next.

How is this accomplished?

Replay v0.2 builds upon our foundational research in generative models that has defined the field. We leverage conditional generative models to precisely control video outputs while maximizing visual quality and coherence. Concurrently with a focus on quality, Replay has optimized model efficiency, allowing our team to scale access to AI technology broadly and responsibly. While quality and efficiency may seem contradictory, optimizations in memory efficiency are part of the reason Replay v0.2 is able to scale to longer lengths and higher resolution.

Tips for using Replay v0.2

Replay can produce very high fidelity videos, but is still early technology. It can take some experimentation to find the best results. Here are some tips for getting the best results:

  • Generate several options: Input images won't be perfectly reproduced in the output video, so we recommend generating several videos and choosing your favorite. There is no limit on the number of videos you can create, so you can keep clicking "Submit" until satisfied.

  • Get creative and try different styles: Replay v0.2 is specialized for photorealistic styles, but does very well at animation, line drawings and more. You can convert a drawing into a full animated clip with an image prompt.

  • Adjust the preserve setting: Images can control a generated video to varying degrees. Sometimes, you only want to set a broad composition — for example, by setting the coarse layout and color palette for the scene. In other cases, creators want to preserve the details of their input image closely. Along with Replay v0.2, we have exposed a preserve control slider in the settings menu.

Closing thoughts

We're excited to see what you create with Replay v0.2! Let us know if you have any other suggestions for improving the Replay experience, and share your creations broadly.