Hello Topaz Team,
First of all, thank you for building such an impressive set of video enhancement tools. As a long-time user of AI-powered products and someone working professionally in Product Management, I wanted to share an idea that I believe could unlock an entirely new category of high-fidelity video restoration.
High-Level Concept
Current video upscaling models are generally limited by the information available within the source footage itself. However, in many situations—especially with old home videos, VHS recordings, heavily compressed footage, or older movies—the facial information available in the original frames is simply insufficient for truly accurate reconstruction.
My proposal is a new restoration workflow that combines:
1. Face Detection & Tracking
2. Identity-Aware Reference Assignment
3. Reference-Guided Restoration
4. Identity Preservation Safeguards
The objective is not face replacement or face swapping, but rather helping the AI reconstruct missing facial details more accurately when the original footage lacks sufficient information.
-–
Core Feature: Face Detection & Reference Assignment
When a low-quality video is imported, the system would automatically:
- Detect all faces appearing in the footage
- Track those faces throughout the timeline
- Group appearances belonging to the same individual
- Present the detected individuals to the user
The user could then select each detected person individually and assign reference materials specifically to that person.
For example:
Detected Person #1 → Reference Set A
Detected Person #2 → Reference Set B
Detected Person #3 → Reference Set C
This face-to-reference mapping becomes the foundation for the restoration process.
-–
Feature 1: Personal Video Reference Mode
Intended Use Cases
- Family archives
- VHS tapes
- Childhood recordings
- Old smartphone videos
- Historical personal footage
Workflow
The user selects “Personal / Home Video Mode”.
For each detected individual, the user can provide:
- High-quality photos
- Recent photos
- Videos of the same person
- Multiple viewing angles
- Reference clips showing different facial expressions
Example:
Detected Person #1 → Father
Detected Person #2 → Mother
Detected Person #3 → Child
The AI can then use those references to better understand:
- Facial structure
- Facial proportions
- Eye shape
- Nose and mouth characteristics
- Facial expressions
- Age-related facial details
The result would be a significantly more accurate reconstruction than relying solely on the degraded source frames.
-–
Feature 2: Commercial Footage / Movie Reference Mode
Intended Use Cases
- Movies
- TV shows
- Commercial footage
- Historical film restoration
- Low-quality digital transfers
Workflow
The user selects “Movie / Commercial Footage Mode”.
For each detected face, the user could:
- Upload actor reference images manually
- Upload actor reference videos manually
- Provide IMDb references
- Select references from approved datasets
- Allow the system to retrieve actor references from approved sources
Example:
Detected Person #1 → Robert De Niro
Detected Person #2 → Al Pacino
Detected Person #3 → Joe Pesci
The AI could use those references to better understand:
- Facial anatomy
- Expression patterns
- Eye details
- Skin characteristics
- Identity consistency
- Fine facial features lost in compression
This could dramatically improve restoration quality when dealing with older movies, compressed streaming sources, DVDs, VHS transfers, or damaged archival footage.
-–
Identity Preservation Layer (Important)
I believe this feature should be explicitly designed as a restoration system, not as a face replacement system.
The references should act only as guidance when source information is insufficient.
The original performance, acting, timing, expressions, and identity captured in the footage should always remain preserved.
The purpose is:
✓ Restore
✓ Reconstruct
✓ Recover lost detail
Not:
✗ Replace identity
✗ Swap faces
✗ Change actors
-–
Suggested Safety Mechanisms
Personal / Home Video Mode
Only references belonging to the same private individual should be allowed.
Examples:
✓ User uploads multiple photos of their father to restore their father.
✓ User uploads additional videos of their grandmother to restore their grandmother.
✗ User uploads a celebrity reference to restore a family member.
-–
Movie / Commercial Footage Mode
Only references matching the detected actor should be allowed.
Examples:
✓ Robert De Niro scene → Robert De Niro references
✓ Al Pacino scene → Al Pacino references
✗ Robert De Niro scene → Arnold Schwarzenegger references
✗ Al Pacino scene → Brad Pitt references
-–
Identity Similarity Verification
An additional safeguard could be an AI-powered identity matching system.
The system could estimate how closely a provided reference matches the detected individual and assign a confidence score.
For example:
- 90–100% Match → Accepted automatically
- 70–90% Match → Warning shown to the user
- Below 70% Match → Reference rejected
This would help ensure that references are being used for identity-preserving restoration rather than identity substitution.
-–
Why This Could Be Valuable
I believe this would create a new category of “Reference-Guided Restoration” where restoration quality is enhanced not only by the information contained within the source footage but also by trusted reference materials associated with the detected individual.
Potential benefits include:
- More accurate face restoration
- Better identity preservation
- Improved facial expression reconstruction
- Greater frame-to-frame consistency
- Reduced hallucinations and artifacts
- Significantly better results on old and degraded footage
- A unique differentiator within the video enhancement market
Thank you for taking the time to read this suggestion.
These ideas come from my professional background in Product Management and Product Ownership. I genuinely enjoy thinking about product design, user workflows, and opportunities to improve already great products. If any part of this concept is useful or sparks further discussion internally, I would be happy to know it contributed in some way.
Best regards,
Madalin Mije
Product Owner
