We are pleased to announce that Copernicus Computing has completed the industrial research phase of the FrameLift AI project – a customized animation upscaling system using artificial intelligence and advanced ESRGAN models. This is a key milestone that confirms the technology’s maturity and readiness to transition into the experimental development phase (TRL 6).
Why FrameLift AI?
Producing animation in native 4K resolution involves enormous time and hardware costs. From the beginning, our goal was to develop a hybrid production model in which:
- animations are rendered at lower resolution (HD),
- then upscaled to 4K quality using customized AI models,
- without losing visual quality or frame consistency.
Key Goals of the Industrial Research Phase
In the last quarter of 2025, we focused on:
- radically reducing client model training time,
- maximizing GPU resource utilization,
- maintaining rigorous quality parameters (PSNR ≥ 35 dB, SSIM ≥ 0.90).
Instead of increasing system complexity (e.g., multi-GPU or custom scripts), we opted for precise optimization of configuration parameters – a stable, scalable solution that is fully transparent to the end user.
What Did We Achieve?
1. Training Optimization
Our approach was to resolve two contradictions. On one hand, the advantage of rendering is full alignment between the achieved result and the artist’s vision, but unfortunately this comes with high computational precision requirements. On the other hand, using AI algorithms significantly speeds up the process, but even in solutions like upscaling, it can generate errors. However, as a large render farm, we have an unusual capability – we can not only generate an image through upscaling. We can also render the image at target resolution and verify how our algorithm performed. This means we can not only choose the best possible algorithm, but also create an algorithm customized to a specific scene – and this is what makes our project unique. Thanks to this approach, during the industrial research phase, we achieved a match indistinguishable to the human eye between normal rendering and AI generation based on a customized model. Furthermore, we managed to reduce training time for a single model to approximately 37 minutes, without any image quality degradation.
2. Real Time Savings in Production
The most important measure of success was testing on actual production scenes. Comparing native 4K rendering with the FrameLift AI pipeline showed:
- average time savings: 44.7%,
- in the most demanding scenes – over 60% reduction in production time.
The more complex the scene, the greater the advantage of our solution.
3. Quality Stability
Every optimization underwent rigorous validation:
- all scenes exceeded the PSNR ≥ 35 dB threshold,
- SSIM remained above 0.90,
- no artifacts, flickering, or loss of detail were observed.
This means that production acceleration does not come at the cost of quality.
What’s Next?
Completion of the industrial research phase means FrameLift AI is entering the next development phase. In 2026, we plan:
- dynamic weight selection between base model and client model,
- further reduction of frame flickering,
- closed beta tests with animation and 3D visualization studios.
Summary
Closing the industrial research phase is not just a formal milestone for us, but above all solid proof that AI can truly change the economics of animation production. FrameLift AI combines:
- speed,
- quality,
- and the scalability required in professional rendering pipelines.
In upcoming posts, we will share progress from the implementation phase and first experiences from beta testing partners.