Stockfilm Research · AI restoration
A face comes into focus. A blurred expression becomes readable. What did the model change—and what did it invent? Compare 13 experiments from the Stockfilm archive, then try the workflow on your own images or video.
The experiments
Harrisburg, Pennsylvania · 1946
A smiling child outside a house. Distinct facial detail appears while the film texture and softer clothing remain visible.
Historical comparison · 7.20 seconds · exact model settings unrecorded
Choose a film. The first five cover different faces, lighting and motion.
Year and location labels follow the source filenames. These experiment inputs are preprocessed footage, rather than untouched film scans. Generated facial detail can differ from a person’s actual appearance.
Run it yourself
Our starter kit keeps the source file, restores detected faces with GFPGAN, and writes a new output. Follow the notebook or use the local image and video commands.
The tools
Face reconstruction, background enhancement and video handling do different jobs. These are the official projects and papers behind the workflow and its alternatives.
Uses a learned facial prior to reconstruct plausible detail in damaged or low-resolution faces. Our starter kit uses the v1.4 checkpoint.
Apache 2.0 repository with listed third-party exceptions. Check the notices for the components and weights you use.
An optional tool for the rest of the picture: backgrounds, edges and texture. Evaluate it separately from face reconstruction.
BSD 3-Clause repository. The starter recipe keeps background enhancement off.
Reads the input, extracts frames, preserves the supported frame rate and reassembles the result with audio. It does not invent facial detail.
LGPL or GPL terms depend on the build and its components.
An alternative to compare on the same input. Its fidelity control balances reconstruction quality and resemblance to the input.
S-Lab terms include non-commercial conditions; commercial use requires contacting the contributors. A research reference, outside this starter recipe.
Sequence-based restoration is a separate research track: RealBasicVSR and BasicVSR++ use neighboring frames. They are research references rather than supported steps in this starter kit.
What you are watching
These historical comparisons were supplied as GFPGAN experiments. The original checkpoint, software revision and restoration settings were not recorded in the files we inspected. The tutorial documents a new recipe; it does not claim to reproduce the old renders exactly.
The matched inputs are compressed working files. Most are 4K HEVC; the Detroit input is 256×144. These are not untouched film-scan masters.
The matching 4K files show the face-restoration result. The moving-wipe comparisons reveal it on the left or top. Exact historical settings remain unrecorded.
Neighboring 8K files record another enhancement step and conversion to 59.94 fps. They are later derivatives, outside the primary comparisons shown here.
Research derivatives, kept separate from production masters. These experiments examine generative enhancement. They do not change Stockfilm’s preservation policy for licensed production footage. Read Archive Integrity.
Sharper eyes, teeth or skin texture are generated interpretations. Judge the full shot, apparent age, expressions and motion—not just a striking paused frame. Doubled facial outlines, abrupt softness, blurred hands and surviving film damage are visible in this collection.
It produces plausible facial detail using learned patterns. It can change identity, expression, apparent age or texture. Preserve your source and label processed versions, especially when working with historical evidence.
The starter recipe works on detected faces. Scratches, dust, background blur and moving hands may remain. General enhancement and film-damage repair require separate evaluation.
The supplied MP4s already contain a moving wipe. Pause or seek to inspect the boundary. The independent input and output files have timing differences, so a draggable video comparison needs verified alignment first.
The comparisons are available to inspect this research. The small tutorial samples are provided for following the workflow. Production reuse of archive footage requires the appropriate Stockfilm license; the code license is separate. Read the licensing guide.
Use the same input, frame size and crop. Record the checkpoint and settings. Compare the full frame and face at native size, then watch motion for changing features or flicker. Without a clean reference, image sharpness alone does not establish accuracy.
Published 2026-10-07 · Stockfilm Research · Historical renders and current tutorial results are documented separately.
