What's next for Refocused? Our roadmap and a pricing update

Hi everyone,

Refocused has been out since late July, and we've shipped nine updates since then. Between them we fixed more than 50 bugs, redesigned the Enhance and Refine workflows, and rebuilt setup and the Windows installer. We added support for AMD Radeon GPUs and NVIDIA's RTX 50 series, along with film grain restoration and lossless FFV1 output. Regenerate-Q also got a lot faster, first with INT8 engines and then with block-sparse attention, which now runs on Apple Silicon, AMD and NVIDIA.

Thank you to everyone who bought a license in these early months, and to everyone who installed it and tried it out. I know that the barrier to even installing a new application is not trivial, so I just want to let you know that I appreciate it.

With that said, I wanted to share the upcoming development roadmap, as well as our more permanent pricing structure.

Development and what's coming

Bug Fixes and Application Updates - Ongoing

While most of our time will go into Model Suite V2, we will continue to put out updates with fixes for the existing models. A stable application is our top priority, so any new issues will get our full attention as they are reported.

Model Suite V2 - Late December 2026

We've been working on training and testing various models and architectures since late September. We've made enough progress to tentatively target the end of 2026 for the full release, and to share that our Model Suite V2 will include these models:

• Regenerate2 - An updated model architecture and new training pipeline for our diffusion-based video super resolution model.
• Regenerate2-Q and Regenerate2-Turbo - Planned quantized and distilled versions of Regenerate2 with faster inference and lower VRAM requirements.
• Mosaic - A deinterlacing model for VHS, DV and broadcast footage.
• Gimbal - A stabilization model for shaky, handheld, or action footage.

Part of our work for Regenerate2 was training a new image encoder, so we will be including a set of models for enhancing images in addition to the video models. Specifically, we intend to have tools for:

• Image super resolution
• Image focus editing

A small caveat: since these models were developed for internal use while training Regenerate2, the image models are not the main priority of V2.

Deinterlacing and Stabilization

We are happy with the results from our deinterlacing and stabilization models, and have started porting them into our inference engine and optimizing them for each platform. While we are targeting the end of 2026 for the complete V2 release, we will include beta versions of the models for Founder License holders as soon as we are able to.

We will make note of these beta features in the release notes, and we estimate Founder License holders will see these new models later in November.

Regenerate2, Regenerate2-Q, Regenerate2-Turbo

We've made some architectural decisions that will allow Regenerate2 to run with significantly less VRAM than our original Regenerate model, so it will be available across all operating systems we support. Our initial results show both better detail and improved consistency across frames when compared to our first generation, but the training run is still in progress.

We will provide more detailed results once the training is complete, but we are targeting two distinct levels of distillation and quantization for users with lower-end GPUs, or for footage that does not require as much restoration, so we will most likely have both a Regenerate2-Q and a Regenerate2-Turbo version of the model.

Our current estimate is that Mac users may see the beta Regenerate2 models in early December 2026, with NVIDIA and AMD support coming closer to the end of December 2026. And as always, this is subject to change. While early training runs have turned up some promising designs, the hard part is always the edge cases. Training the Regenerate models takes thousands of H100 GPU hours, and we always have to strike a balance between consistency from frame to frame and convincing textures: pushing one tends to cost the other, and we usually only see the results after many hours of training.

Pricing

Training video models is slow, expensive work. A video model has to learn more than what a sharp frame looks like. It also has to learn how detail should hold steady from one frame to the next, and that takes a lot of footage and a lot of GPU time.

We also only train on footage we've licensed. That costs more and limits how much footage we can use, but it means the models you run weren't built on anyone's work without their permission. And because everything runs on your own computer, not in a data center, there's a hard limit on how large the models we release are.

To keep developing Refocused and training new models, we're changing our pricing on November 3.

The Founder License

First, thank you for working through the kinks with us. Everyone with a license bought before November 3 has a Founder License: our full license, with the renewal price locked at what you paid. It works the way your license works today:

• Nothing you have changes. Your current models and every app update are yours for good, along with every model released within a year of your purchase.
• Renew at $64.99, for good. After the year of included model updates is finished, you can always renew at your original price to get every model released since, plus another year of new models.
• V2 is included, regardless of when you purchased it.

This license is our thank-you to early adopters. We will never take away a model you've had access to, and we hope our continued work earns your business for years to come.

Until V2 ships

For new customers, we now offer two options:

Our target for V2 is the end of December 2026. The Founder License is only on sale until V2 ships, and its holders get beta access to new models as soon as they are available.

Once V2 ships

When V2 is released, we'll stop offering the Founder License, replace it with the Full License, and add a V2 Suite:

Whichever you choose, every option is still a one-time purchase. There's no subscription, and nothing stops working if you never pay again. Each license covers up to 5 devices and comes with a 14-day money-back guarantee. The free version with a watermark isn't changing.

Thanks again for being here early.

Jacob Stein
Founder, Refocused