Let’s make
the next one better.

Start with the guide

The user guide covers the basics: adding media, choosing a task and model, batches, video, and where results are saved. Its troubleshooting section answers the most common questions.

Report a bug

Report bugs and ideas on GitHub Issues. For anything private, email hermes.reisner@gmail.com.

Include your app version, operating system, CPU/GPU, model, input dimensions and exact steps. For video, include format, duration, HDR/SDR mode and output size. Copy diagnostics can help; review it first. A short sample you are allowed to share is useful but optional.

Public reports are visible to everyone. Remove private paths, images and unrelated screen content. Do not post settings files, full queue databases, passwords or signing certificates. Send security reports privately to the contact address.

Start with a short clip

LocalSR is intended for capable hardware. AI processing is not real-time. A short video can take hours; long or high-resolution work can take days. In one completed Linux ROCm test, a five-minute 480×854 video took about 100 minutes to upscale to 960×1708. That result is one workload on one machine, not a speed promise.

CPU and GPU scores are measured separately. An integrated GPU may be slower than the CPU. Free working memory, scene content, output dimensions and model choice all matter. A 16 GB GPU does not guarantee SeedVR2 will fit. Start small and inspect the result.

Known beta limits

• The Mac build (Apple Silicon, macOS 14 or later) is the most tested. The Windows and Linux packages run on the CPU, are unsigned, and have had automated checks only.
• AI can change text, faces and fine details. Keep originals and inspect outputs.
• Two of eleven difficult scanned documents were rejected by the denoiser’s instability checks; nine completed. Rejection avoids saving the detected unstable result, but it does not guarantee every accepted image is artifact-free.
• SeedVR2 video restoration is a Labs feature, exports SDR and needs a lot of memory. Its small VAE tiles can leave seams.
• Experimental HAT HDR preservation exports 10-bit HLG/PQ with colour tags. The models were trained on SDR; HDR perceptual quality is unverified. Live previews are SDR display conversions.
• MOV support depends on codecs, colour metadata, orientation and audio tracks, not the extension alone.
• The full list is in KNOWN_LIMITATIONS.md.

Project availability

LocalSR is a personal project for volunteer testers. The maintainer has limited time alongside university; responses and fixes may take time.