Proposal for a New AI Model in Topaz Video AI: Restoration of Corrupted Macroblocks (Dropout / Encoding Errors)

Topaz Video Ideas
1

Proposal for a New AI Model in Topaz Video AI: Restoration of Corrupted Macroblocks (Dropout / Encoding Errors)

1. Executive Summary

Digital video originating from tape‑based formats (DV, Digital Betacam, DVCAM, HDV…) or from faulty encodings often exhibits macroblock corruption, such as:

  • macroblocks frozen across several frames,

  • blocks not updated during camera movement,

  • shifted or misaligned macroblocks during a travelling shot,

  • mosaic‑like artifacts on a few isolated frames,

  • GOP error propagation.

Currently, Topaz Video AI does not include any model specifically designed to correct these structural defects. Existing models (Proteus, Artemis, Iris) may soften or partially hide the issue, but they do not reconstruct missing or corrupted information.

A dedicated model — or a specialized pre‑inference step — would allow restoration of corrupted frames before super‑resolution, resulting in significantly improved output quality.

2. Nature of the Problem

2.1. Origin of the Corruption

  • Digital dropout on tape: loss of data in DV/MPEG‑2 streams, causing frozen macroblocks.

  • Encoding errors: corrupted GOP structure, invalid motion vectors, damaged P/B frames.

  • Decoding faults: occasional read errors during capture.

2.2. Visual Symptoms

  • Macroblocks remaining static while the scene moves.

  • Horizontally or vertically shifted blocks.

  • Missing information in specific regions.

  • Partially reconstructed frames.

  • Artifacts appearing only on 1 to 5 consecutive frames.

2.3. Why Current Models Fail

Topaz models are optimized for:

  • denoising,

  • deblocking,

  • detail enhancement,

  • super‑resolution.

But not for structural reconstruction of partially invalid frames.
They cannot detect that a macroblock is “frozen” or “misaligned”, and therefore do not attempt to correct it.

3. Proposal: New AI Model “Macroblock Repair” (MBR‑Net)

3.1. Objective

Create a model capable of:

  • automatically detecting corrupted macroblocks,

  • reconstructing missing areas using neighboring frames,

  • realigning shifted blocks,

  • synthesizing textures consistent with global motion.

3.2. Technical Approach

The model could combine:

  • Multi‑frame temporal analysis (3 to 7 frames around the corrupted one),

  • Detection of frozen blocks via inter‑frame comparison,

  • Warping + AI‑based inpainting,

  • Motion‑aware synthesis to respect camera movement,

  • Texture propagation from valid frames.

3.3. Implementation Options

Option A: Dedicated Model

Proposed name: MBR‑Net (MacroBlock Repair Network)
Usage:

  • placed before super‑resolution,

  • manually or automatically enabled.

Option B: Integrated Pre‑Inference

A simple checkbox in Topaz Video AI:

“Detect and repair corrupted frames (macroblocks / dropout)”

Functionality:

  • quick scan of the video stream,

  • automatic detection of suspicious frames,

  • repair step executed before Proteus/Iris.

4. Typical Use Cases

  • Restoration of DV / DVCAM / HDV archives.

  • Analog captures encoded in MPEG‑2 with errors.

  • Old YouTube videos with GOP corruption.

  • Family videos with occasional digital dropout.

  • Independent films shot on early digital cameras.

5. Recommended Processing Pipeline

An ideal pipeline inside Topaz Video AI would be:

  1. Automatic detection of corrupted frames
    Inter‑frame analysis, frozen block detection.

  2. Structural restoration (MBR‑Net)
    Reconstruction of invalid macroblocks via AI.

  3. Temporal stabilization
    Correction of micro‑jumps introduced by reconstruction.

  4. Super‑resolution / Upscaling
    Proteus, Iris, or other models.

  5. Detail enhancement / Sharpening
    Final refinement.

6. Why Topaz Should Implement This

  • No equivalent solution exists in consumer‑level tools.

  • High demand from both amateur and professional restoration communities.

  • Perfect complement to existing Topaz models.

  • Significant improvement in final quality, especially for SD sources.

  • Strong differentiation from competitors (DaVinci, Neat Video, etc.).

7. Conclusion

Adding a dedicated model for macroblock corruption repair — or an automatic pre‑inference step — would allow Topaz Video AI to address a real, frequent, and currently unsupported problem.

This model would be especially valuable for:

  • archive restoration,

  • SD/MPEG‑2 source enhancement,

  • recovering damaged family videos.

It is a logical and highly anticipated evolution of the Topaz Video AI suite.

Best regards, Vincent.