PHIL MAHER / CONVERSATION NOTES https://stockfilm.com/speaking Updated 11 October 2026 SHORT BIO 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. SIX CONVERSATIONS Teaching an image model the look of old Paris https://stockfilm.com/speaking#vintage-ai-models - What did you want the model to learn: the city, the film stock, or a feeling? - How did you choose the frames, and what did you leave out? - Which details make an invented picture feel old before we inspect its subject? - Can a model preserve a look while getting the history wrong? - Where should a generated image’s label appear when it leaves the model page? Sources: https://stockfilm.com/1947-paris-france https://civitai.com/models/405790/1947-paris-france-replication-sd2021-lora Face restoration: show the change, question the result https://stockfilm.com/speaking#ai-restoration - What made you want to restore these faces in the first place? - Which change looks convincing in a still but becomes questionable in motion? - What would count as evidence that a reconstructed face preserves identity? - How should a documentary label a restoration that supplies missing detail? - When would you choose the softer source over the more appealing output? Sources: https://stockfilm.com/research/ai-restoration https://stockfilm.com/research/ai-restoration/gfpgan https://github.com/TencentARC/GFPGAN Archival data for independent researchers https://stockfilm.com/speaking#research-support - What kinds of real film damage are difficult to represent with synthetic examples? - What should accompany a frame so another researcher can understand its provenance? - How would you avoid neighboring frames from the same reel leaking across a train/test split? - What does a no-reference quality score leave untested? - If the chart and the picture disagree, which additional evidence would you ask for? Sources: https://arxiv.org/html/2608.02346v2 https://stockfilm.com/ai-training#published-research TimeFly: an archive meets a fly simulation https://stockfilm.com/speaking#timefly - Why put a fly simulation and a home-movie archive in the same project? - Which parts of the display come directly from measured data? - What did you have to choose before the wiring could produce an animation? - What would we need to test before calling its movement biologically realistic? - How do you make a scientific visualization compelling without making its claims larger? Sources: https://stockfilm.com/timefly https://stockfilm.com/timefly/methodology https://male-cns.janelia.org/ The Sounds We Lost: sound as interpretation https://stockfilm.com/speaking#the-sounds-we-lost - Why let people encounter the film without added sound first? - What did the close footsteps make you notice that distant traffic did not? - When does a plausible sound become a historical claim? - What does the film tell us about listening distance—and what are we inventing? - Can returning to silence change our understanding more than another sound layer? Sources: https://stockfilm.com/the-sounds-we-lost https://stockfilm.com/the-sounds-we-lost/audio-production.json The Visual Memory Index: who got filmed, what went missing? https://stockfilm.com/speaking#visual-memory - Who had access to a camera, and how could that shape the archive we inherit? - What decisions turn a reel into hundreds of searchable clips? - What is the difference between a catalog estimate and a date you can verify? - How can we tell absent footage from footage we have described badly? - If an AI model learns from this collection, which gaps might it carry forward? Sources: https://stockfilm.com/research/visual-memory-index/2026 https://stockfilm.com/research/visual-memory-index/methodology EVIDENCE AND LIMITS Paris outputs are AI-generated scenes, not records of events in 1947. Older restoration inputs were preprocessed and original settings are unknown; facial details are unverified. Dates and locations follow comparison filenames and need catalog confirmation. Loop-Mamba is independent research; Phil is not an author. Metrics do not certify historical truth. TimeFly models measured wiring with chosen dynamics and controls; it is not a living animal. The Sounds We Lost adds generated interpretations, not recovered sound. Human headphone review and artistic sign-off remain required in the production record. The Visual Memory Index describes one archive. 2026 charts use the immutable v1.0 snapshot, not current inventory. INVITATIONS phil@stockfilm.com https://stockfilm.com/contact?inquiry=speaking Share audience, question, format, date/time zone, location or remote setup, preferred length and Q&A. Budget or honorarium is optional.