Conversations with Phil Maher / Founder of Stockfilm

What happens when we ask machines to remember?

A family film can teach an image model a vintage look, become a world for a fly simulation, or turn into a city you listen to. Phil builds these experiments from an archive of real home movies. He brings the footage, the results and the questions they leave open.

Explore the six conversations Invite Phil for a talk or interview ↗
An invented Paris street generated with Phil’s published vintage image model
An invented Paris / AI-generated
Actual TimeFly screen capture: archival film beside a simulated fly’s neural activity
A fly simulation watching a home movie
Phil Maher on AI:AM, the daily live show from Nathan Labenz and Prakash Narayanan. Play starts at Phil’s segment.

Latest speaking engagement

AI:AM October 2026 Phil enters at 1:19:38

Measuring experience. Preserving history.

Phil joins Nathan Labenz and Prakash Narayanan to discuss original film as AI training data, experimental face restoration, and TimeFly, his browser-based fly simulation. Explore the full conversation, the hosts’ questions, and the shorter published cuts.

Explore the conversation Transcript, chapters & clips

Host Nathan Labenz has also interviewed, on The Cognitive Revolution

About AI:AM
Phil Maher in conversation with filmmaker Christian Taylor.

Hear Phil in conversation

Documentary First Episode 287 49 minutes

Hear the decisions behind the archive.

Phil joins Christian Taylor to talk about the family reels that started Stockfilm, rescuing fragile film, and why a few seconds of everyday life can make a documentary unforgettable.

Start at 23:36: home movies and storytelling →
Explore the interview Transcript, chapters & clips

Past guests on Documentary First

TLDR / What Phil brings

An archive to show.
Experiments to question.

Phil’s work connects film preservation with hands-on experiments in AI, sound, scientific visualization and archive research. Your audience can see how an image is made, changed, interpreted and counted—and decide what it can actually tell us.

  • Concrete material: source footage, generated scenes, moving comparisons, working demos and published data.
  • Questions with consequences: does a sharper face preserve identity? Can a soundtrack invent a memory? Whose history is missing from a model’s examples?
  • A practitioner’s perspective: Phil can walk through his choices, show what worked and explain what remains uncertain.

Stockfilm preserves 1,600+ reels and has 162,000+ searchable clips. July 2026 archive facts (2026-07-27). Background and sources ↗

Phil Maher speaking about his archive and experiments

Where the questions began

From his grandfather’s films to these experiments.

Phil founded Stockfilm in 2015 after discovering his grandfather’s home movies. Sourcing, scanning, restoring and cataloging footage puts him close to the material: the ordinary details that survive, the records that are missing, and the changes software can make. That daily practice is the starting point for the conversations below.

Read Phil’s story ↗

Six ways into the conversation

Choose a question.
See what comes with it.

Each topic below includes examples, methods and suggested questions. Use them to prepare an interview, shape a panel or choose what to put on a screen. The material stays visible as you scroll.

01 / Generative images, memory & style

Teaching an image model
the look of old Paris.

When an invented scene feels like a memory, what are we recognizing?

Phil trained image models on archival Paris frames and published them on Civitai as ReplicantArmy. The project starts with real film, but its output can be a scene the camera never recorded. A familiar landmark, soft edges and faded tones can make that new image feel as though it belongs in an old collection.

That gives a conversation two directions. One is the practical work: choosing frames, preparing examples and judging outputs. The other is more unsettling: what makes an image feel historically convincing, and how quickly does that feeling turn into belief?

Put the source collection beside the generated gallery. Look at texture, perspective, buildings and people separately. An image may capture a recognizable photographic style while inventing the things within it.

The film that starts the question

Actual archival preview frame: Eiffel Tower from a Paris street
Archival frameEiffel Tower from a Paris street · Clip 123119153
Actual archival preview frame: Arc de Triomphe
Archival frameArc de Triomphe · Clip 123122025
Preview frames from the Paris source collection. These illustrate the collection; they are not a verified list of the model’s training inputs. Paris and 1947 retain their catalog qualifications.

Four scenes generated with the published model

Click an image to inspect its original Civitai post.

AI-generated Eiffel Tower scene with faded sepia tones and soft film-like detail
AI-generated / ReplicantArmyEiffel Tower, reimagined
AI-generated Arc de Triomphe in a softly textured sepia image
AI-generated / ReplicantArmyArc de Triomphe, reimagined
AI-generated vintage-style Paris street scene with a tram and pedestrians
AI-generated / ReplicantArmyA tram on a Paris street
AI-generated Paris bridge with muted tones, grain and soft focus
AI-generated / ReplicantArmyA riverside view

Generated stills illustrate a vintage style. They do not document events in 1947. The published model is a LoRA; its model page contains the release and usage information.

  1. Source filmA recorded place, with catalog context
  2. Training examplesFrames selected to teach a visual look
  3. Published modelA way to generate with that look
  4. New imageAn invented scene that needs a label

What Phil can unpack

How he approached the Paris models, what he wanted them to learn, and which outputs persuaded or disappointed him. Bring a generated image into the interview and ask him to separate the cues inherited from film from the content the model supplies.

A useful direction

For an AI or design audience, follow the making of an image. For a history or culture audience, follow the moment a visual style begins to stand in for evidence. Both routes can start with the same four pictures.

Browse the source collection ↗Explore Phil’s Paris model ↗ReplicantArmy’s published models ↗

02 / Face restoration, identity & historical evidence

Face restoration: show the change,
question the result.

If a face becomes clearer, have we learned more about the person?

The first reaction to these comparisons can be delight: eyes, a mouth and an expression appear where the input was soft. The next question is harder. Which details were recoverable from the footage, and which are a plausible face supplied by the model?

GFPGAN uses a learned facial prior. That makes it useful for an experiment—and makes it essential to look beyond the sharpest still. Phil’s moving comparisons let an audience watch the detail change as a face turns, blurs or crosses the divider.

Moving comparison for A child in a hat: AI output at left and processed experiment input at right

A child in a hat

Filename label: Harrisburg, Pennsylvania, 1946 · 7.20 seconds

Compare the eyes, hairline and mouth as the green divider crosses the face.

AI output: left of green divider
Processed input: right of green divider
Moving comparison for A child in red: AI output at left and processed experiment input at right

A child in red

Filename label: Hoboken, New Jersey, 1966 · 7.55 seconds

Compare the eyes, hair and smiling mouth.

AI output: left of green divider
Processed input: right of green divider
Moving comparison for Newlyweds in motion: AI output at left and processed experiment input at right

Newlyweds in motion

Filename label: Detroit, Michigan, 1950 · 7.55 seconds

Pause near 2.5 seconds to compare the two sides of the groom’s face.

AI output: left of green divider
Processed input: right of green divider

What Phil can unpack

Why he tried face restoration, what the moving results reveal that a promotional still can hide, and how to show an experiment without letting it replace the source. In the Hoboken example, inspect the cheek and mouth outlines. In the Detroit example, watch the faces turn and compare the blurred frames.

Make the audience part of the test

Play one clip without commentary. Ask what people think has improved. Then replay it and ask what they would need to verify the added detail. The shift from “that looks better” to “what can this prove?” is the conversation.

The accompanying tutorial and starter code let people try the method on their own images or footage. The starter uses a documented GFPGAN v1.4 recipe; it does not claim to reproduce the unknown settings of these older experiments.

All 13 comparisons ↗Tutorial and starter code ↗GFPGAN source ↗How Stockfilm labels evidence ↗

03 / Independent research, real damage & evaluation

Archival data
for independent researchers.

What changes when a restoration model meets damage nobody designed?

Independent researchers evaluated nine restoration methods on 500 frames from Stockfilm in the Loop-Mamba paper. That is a concrete use of the archive: surviving film gives researchers material beyond a simulated scratch or a deliberately blurred clean image.

The archive’s contribution and the model research are separate. Phil can explain the material and the archive’s data practice. The paper’s authors developed and evaluated Loop-Mamba; Phil is not an author.

For a conversation, the interesting step is to move from “a model improved an image” to “how did we decide?” A score can describe visual quality without establishing that a reconstructed object or face is historically correct.

Published use of Stockfilm data

Archival frames evaluated
500
Restoration methods compared
9
Paper version
v2 / 27 Aug 2026
Read the independent paper ↗

Evidence: real-world evaluation, Table 6 and Figure 6. This is research context, not a claim of model authorship or a historical-accuracy certification.

Figure 6 from the independent Loop-Mamba paper: three Stockfilm source frames and enlarged regions compared across restoration methods
Published visual comparison: Figure 6, Loop-Mamba v2. Source: Bai and colleagues, arXiv:2608.02346v2, CC0. The paper supplies the method labels and full-size figure.

One metric, nine different results

BRISQUE, as reported in Table 6. Scale starts at 0 and ends at 40; lower is better on this metric.

  1. NAFNet27.8305
  2. ClearAIR28.6961
  3. Restormer32.5864
  4. HogFormer25.1794
  5. DiffBIR30.4014
  6. AMFDiff20.9117
  7. MambaIR30.6522
  8. MambaIRv230.1474
  9. Loop-Mamba22.3574

The table gives AMFDiff the lowest BRISQUE value, although the paper’s prose calls Loop-Mamba best. TOPIQ and MAN-IQA give different rankings. This chart follows the table; these are perceptual metrics, not measurements of historical truth. Inspect Table 6 ↗

What Phil can unpack

What it takes to describe old film as research material: where a frame came from, what transformations it has already undergone, what the archive knows about it, and what is still an estimate. The same care matters when an attractive benchmark figure travels into an article or presentation.

Follow the evidence, including its disagreements

Ask Phil to start with a damaged frame and work backward to its source record. Then compare the chart with the figure. If a metric, a caption and your own impression disagree, what further test would help? That is a productive question for researchers and for anyone using an AI-enhanced image.

Paper and reported results ↗Published research context ↗Archive dataset datasheet (JSON) ↗

04 / Scientific data, simulation & public understanding

TimeFly: an archive
meets a fly simulation.

How much of a convincing simulation comes from measurement—and how much comes from choices?

A home movie becomes a flat visual world for a simulated fly. The interface shows the film, samples of visual input and a map of neural activity. Move the light, switch the film and watch the model respond.

TimeFly combines a measured fruit-fly wiring diagram with modeled neuron dynamics and an engineered way of turning activity into movement. Those ingredients do different jobs. The connections come from scientific data; the parameters, filtering rules, stimulus and steering interface determine how the experiment runs.

Phil can use the working page to make those layers visible. It is an unusual route into a discussion about open scientific data, browser-based experiments, and the temptation to treat an expressive visualization as a complete explanation.

A real recording of TimeFly: the simulated fly moves over a 1963 home movie while the modeled neural activity changes

12-second screen recording of the working TimeFly page, captured 11 October 2026 (local time). The flashes represent simulated activity. Open the live experiment to move the light and change films.

Wake the fly in the live experiment ↗
  1. MeasuredMaleCNS wiring and annotations from a scientific release
  2. ModeledNeuron dynamics, visual input and filtered connections
  3. EngineeredA light to chase, steering readout and a drawn fly

What Phil can unpack

Why he brought archival film and a connectome together, how to make an experiment inspectable, and which explanations the interface needs. The methodology separates measured anatomy, modeled dynamics and engineered behavior, including the assumptions and omissions.

A demonstration that invites interruption

Start the fly, point to the film and neural map, and ask the room what they think each display proves. Then change the light or the film. The point is to examine the relationship between the input, the model and the animation—and to identify which conclusions require an actual animal experiment.

Try the simulation ↗Methods, assumptions and limits ↗MaleCNS source data ↗

05 / Sound, attention & interpretation

The Sounds We Lost:
sound as interpretation.

The picture stays the same. Why does a different sound make it feel like a different place?

Begin with a short film and no added sound. Then imagine standing near the people, or listening past them to the street. A dry footstep, passing engine or distant mechanical texture can move your attention without changing a frame.

The Sounds We Lost turns that choice into a listening room. Three city studies offer four sound layers apiece, two authored interpretations, a visitor’s mix and a return to silence. The footage survives; the sounds are newly generated and edited for the work.

That distinction is the subject of the conversation. A soundtrack can suggest distance, weather, rhythm or an event outside the frame. Which suggestions feel natural? Which become claims the picture cannot support?

Archival Chicago study: a street with traffic, people and a CHICAGO sign
5.005-second image loop, clip 125593798. Chicago and 1968 are catalog estimates. The excerpt contains a cut between two views; the inspected digital master has no audio stream.
Actual listening-room interface showing Chicago footage, two interpretations and controls for four generated sound layers
The working listening room, captured 11 October 2026 (local time). Open it to hear the layers together, change their levels or return to silence.

Hear what the work adds

These are two individual generated layers from the Chicago study. Listen to the texture, then return to the film. They are not recordings from that street and are not synchronized to the people’s movements.

At walking distance / generated footsteps

Close, dry taps and scuffs. The rhythm, footwear and listening distance are invented.

Passing traffic / generated motor texture

A rounded motor texture approaches and recedes. The footage does not establish the engines’ original sound.

15.015-second edited sound loops. Stable Audio Open 1.0 generation, selected and edited under Phil Maher’s project direction with Codex assistance. The production record documents prompts, seeds and editing; human headphone review and artistic sign-off remain listed as required.

Compare the full sound studies ↗Generation and editing record (JSON) ↗Sound model and license ↗

What Phil can unpack

The choice to leave the films silent at first, the difference between an audible event and a plausible texture, and why omission can be as useful as adding a layer. He can walk through what is visible, what the catalog supports and what the sound designer has imagined.

Try the same scene twice

Hear “At the curb,” then “The city carries on.” Ask which person or object drew attention each time. Finish with no added sound. This can become a film-school discussion about authorship, a museum discussion about interpretation, or a conversation about whether generated media changes what we remember.

Working prototype and independent exhibition proposal. Software availability does not establish a tested physical installation, festival selection or recovered historical audio.

Enter the project ↗Artist’s statement and exhibition proposal ↗

06 / Collection bias, search & the surviving record

The Visual Memory Index:
who got filmed, what went missing?

When we search an archive—or train a model on one—whose view of the past are we getting?

The Visual Memory Index makes Stockfilm’s collection visible as data. Its versioned report describes what survives by decade and place, what people search for and where the searchable material falls short of demand. The figures and aggregate downloads let readers examine the claims.

A collection can contain a great deal of footage and still give a narrow view of history. The first chart below shows the concentration by decade. The second distinguishes having a catalog field from having a physical-reel trace. Neither quantity, by itself, establishes that a date or location is correct.

What survives, by decade

Share of 162,603 searchable clips in the 2026 v1.0 snapshot. A clip is a catalog item, not a unique person or reel.

  1. 1930s4.8%
  2. 1940s6.5%
  3. 1950s21.1%
  4. 1960s41.9%
  5. 1970s19.7%
  6. 1980s4.7%

The 1960s account for about 42% of the searchable catalog. The bars use the whole catalog as the denominator; clips outside the listed decades are not shown.

A catalog field is not the same as a trace

Share of the searchable catalog. Dates and locations can be estimates; these bars describe metadata coverage, not verification rates.

  1. Has a catalogued date98.7%
  2. Has a catalogued location98.9%
  3. Traced to a physical reel46.1%

Published 15 July 2026, report stockfilm-vmi-2026-v1.0. These are immutable release figures, rather than current live inventory. Read the versioned report ↗

What Phil can unpack

The choices between a reel on a shelf and a searchable result: selection, scanning, segmentation, description, dating and grouping. Those decisions affect what a researcher sees—and which examples an image model receives.

Connect this topic back to the experiments

Return to the Paris images or restored faces and ask what their source examples could not contain. Then ask what more digitization would solve and what it would leave untouched. Making an archive easier to search does not make its surviving record representative.

The report publishes aggregate data with small-cell suppression. It excludes raw search queries and identifiers. Hosts can use the downloadable figures and CSVs to prepare questions without treating a visitor’s query as a personal history.

Report, figures and aggregate downloads ↗Definitions, denominators and limits ↗Figures for your program ↗

Build the discussion around your audience

Three suggested paths.
Room to follow a new question.

You do not need to cover every project. One example examined closely can open a better conversation than a tour of six demonstrations.

AI, memory and trust

Begin with an invented Paris image. Watch a restored face. Finish by asking what the Visual Memory Index says about the examples a model learns from.

Start with the Paris images ↗

Making research understandable

Compare the restoration chart with its figure. Run TimeFly. Ask which observations come from measured data and which come from the design of a model.

Start with the research evidence ↗

Film, sound and interpretation

Watch a city without added sound. Hear two interpretations. Return to the archive’s gaps and ask whose presence, viewpoint and memories survived.

Start in the listening room ↗

Talks, panels, lectures & interviews

Plan for a conversation
people can inspect.

Choose the audience and the central question first. Phil can discuss which films, results and live experiments suit the format, your available time and the room.

For an audio interview, describe one image or change carefully and give listeners the project links. For a visual talk, plan a screen or screen share. The sound studies benefit from an audio check and a quiet comparison.

Share the proposed length, Q&A time, location or remote setup, and whether a live browser demonstration is practical. The recordings and screenshots on this page also provide a way to show the work.

For hosts & organizers

An introduction
and material to prepare.

Copy a short bio or event introduction. The speaker notes include all six subjects, suggested questions and source links. The portrait is available for web use; the press kit provides archive background.

Download Phil’s portrait ↗Download conversation notes (TXT) ↗Open the press kit ↗Phil on LinkedIn ↗

Phil Maher, founder of Stockfilm

Phil Maher founded Stockfilm in 2015 after discovering his grandfather’s home movies. He preserves and catalogs real 8mm and Super 8 film, then builds experiments with the archive: vintage image models, face restoration comparisons, a fly-brain simulation and imagined city soundscapes. His conversations connect the footage, methods and results with questions about memory, historical evidence and what machines add to an image.

Event introduction

Phil Maher is the founder of Stockfilm, an archive that began with his grandfather’s home movies. His experiments put old footage into image models, face restoration, a fly-brain simulation and a listening room. Today he’ll share the footage and results, and ask what we preserve, what we invent and how those choices change our understanding of the past.

Bring a question

Invite Phil for
a talk or interview.

Tell Phil who your audience is and what you want to explore. Include the format, proposed date and time zone, location or remote setup, and preferred length. Add a budget or honorarium if you have one.

Start a conversation with Phil ↗phil@stockfilm.comOr choose a subject for the invitation:
Inventing old Paris ↗Restoring a face ↗Testing on real damage ↗A fly watches a home movie ↗Listening to silent film ↗The history outside the frame ↗
Archive practice ↗The preservation commitment ↗Footage on screen ↗