Retrospective craft notes / 198 notes
Model releases, tool documentation, guild terms, disclosure rules and the older precedents behind AI filmmaking, each read from the shot outward: the workflow, and what the tool does not change.
Historical event dates and source dates are separate from the preparation date of this local edition. Every note is a retrospective draft prepared 16 September 2026; none was published on its historical date.
198 notes
The Cinémathèque française and the Méliès family's own archive describe the 1902 film and its success without naming a formal patent.
Library of Congress records and a commissioned essay detail which King Kong effects were stop-motion, projection, or full-scale mechanics.
Disney's own archive and a 1938 trade report credit a studio team led by William Garity, not one inventor, with the multiplane camera.
A 1945 trade report and a later patent document Dunn's optical printer; the Academy's current award records do not reach back that far online.
A granted 1963 patent and a contemporaneous trade article document Petro Vlahos's sodium-vapor traveling-matte process and its adoption.
American Cinematographer's 1968 coverage and a National Film Registry essay document Trumbull's own slit-scan rig behind 2001's Star Gate.
ILM's and Lucasfilm's own histories agree the effects house was founded in 1975 to close a capability gap for Star Wars.
Garrett Brown's 1977 patent documents the Steadicam's design; its specific Academy Award citation could not be independently opened here.
Library of Congress sources carefully date The Jazz Singer's 1927 premiere and narrow what its synchronized-dialogue first actually covers.
A 1952 trade account and a National Film Registry essay document Fred Waller's three-camera Cinerama system and its wartime origin.
OpenAI's own February 2024 page describes a restricted preview for red teamers and invited artists, not a public launch.
Google's May 2024 announcement shows Veo launched as a waitlisted preview without audio or camera control.
Meta's own blog post and paper describe Movie Gen as a partnered research release, not a shipped tool.
Stability AI's arXiv paper documents a training method and data curation, separate from the model's public release.
Tencent's paper reports a professional-evaluator comparison it designed and ran itself, not an independent ranking.
Alibaba's own repository dates Wan2.2's mixture-of-experts upgrade and its unverified Wan-Bench 2.0 claim.
Runway's announcement and its own changelog date Gen-3 Alpha's rollout and its staged feature additions.
Genmo's own repository, blog post and model card date Mochi 1's release and disclose its hardware and resolution limits.
Luma's own product page and news index date Ray2's subscriber-only rollout and its later feature additions.
ByteDance's Seed team published Seaweed-7B as a research paper and demo page, not a downloadable or licensed product.
Zhipu AI's two open CogVideoX tiers carry different licenses, and only the smaller one is unrestricted for commercial use.
HPC-AI Tech's Open-Sora documents its own training recipe and benchmarks against open models, not against OpenAI's Sora.
Lightricks published LTX-Video's weights under a license that requires a paid agreement above a stated revenue threshold.
Rhymes AI open-sourced Allegro under Apache 2.0 and ranked it against rivals in its own unpublished-methodology user study.
Meta's Emu Video paper describes a two-step image-then-video method and reports preference results Meta gathered itself.
NVIDIA launched Cosmos as openly licensed world-foundation models for robotics and driving, not a filmmaking tool.
Runway's own research post named what Gen-2 could not yet do and set out a long-term world-model research agenda.
Shengshu Technology's Vidu paper reports performance on par with Sora using the developer's own comparison, not a third party's.
MovieLabs's own 2030 Vision materials name founding studios and a technology roadmap, not a certified vendor list.
SMPTE's own pages describe an AI in Media Task Force and annual journal issues, distinct from a ratified engineering standard.
The EBU's 2023 briefing and 2024 guide give public broadcasters a coordination vocabulary, which the EBU frames as voluntary.
NIST AI 100-4 catalogs provenance and detection techniques for synthetic content as guidance, not a binding rule.
The C2PA's own conformance program and versioned specification define what generators and validators must do, and who has actually passed.
Stanford's Foundation Model Transparency Index measures company disclosure, and its authors say so explicitly.
VideoPhy's own 2024 test found generated video followed captions and physics together only 39.6% of the time, with stated limits.
FaceForensics++ names its four manipulation methods and 1,000-video source, boundaries a citation should preserve.
DeepMind's own pages describe SynthID watermarking Veo output for detection, distinct from full manifest-based provenance.
A filmmaking guide's coverage definition and a video vendor's own shot-term glossary show what a generation brief must still specify shot by shot.
An editing guide's eyeline definition and a vendor's own camera-terms reference show gaze continuity is not the same thing as camera position.
A shot-reverse-shot guide and two vendors' own consistency documentation show that matching an actor's look across cuts still needs correction, not a switch.
An editing guide's cross-cutting definition and a scheduling tool's own scene-tracking features show what parallel action still requires before intercutting.
A long-take guide's editorial function and two vendors' own duration and extend documentation show length is now a stated technical ceiling.
A shutter-angle guide's own formula and a generation vendor's prompt-level motion-blur language show the physical rule has no generated equivalent yet.
A rack-focus guide's lens-based definition and a vendor's own prompt-level focus terms show depth of field is described, not computed, in generated shots.
Two editing guides' J-cut and L-cut definitions and an editor's own split-edit documentation show what generated picture still has to be trimmed to fit.
A continuity guide's 30-degree rule and a vendor's own angle-changing camera terms show a brief must state the angle change no model tracks on its own.
A format guide's standard ratios and a vendor's own ratio-expansion documentation show a generated clip's frame needs a declared step to match delivery.
Foundry's own documentation describes CopyCat as a task-specific trainer, not a generative video model.
The Academy's own ACES documentation describes a color-management standard, not a claim about generation output.
OpenEXR's own documentation traces the format from a 2003 ILM release to Academy Software Foundation governance.
Blackmagic's own manual keeps 'Neural Engine' language to the Color and Edit pages, not to Fusion's Matte Nodes.
Adobe's own archived pages show the sampling tool shipped before CS6, later extended into video as Project Cloak.
ILM's own case materials describe a markerless capture rig named FLUX, distinct from any generative de-aging model.
Papercup's own published case studies claim results for Sky News and Bloomberg's dubbed news channels by name.
Digital Domain's own technology page names Charlatan's method and productions without describing any outside license.
The Visual Effects Society's Gen AI Resources page assembles panel discussions rather than a single ratified policy on generative tools.
American Cinema Editors addresses generative AI through public programming, not a written policy, distinct from its bargaining counterpart.
Motion Picture Sound Editors barred generative AI soundtrack work from Golden Reel eligibility, then wrote narrow exceptions into its rules.
IATSE's own account of a unanimous 2023 NLRB election led to a first Marvel VFX contract two years later.
The DGA's own commercial agreement text carries no AI-specific clause, unlike the 2023 Basic Agreement already on file.
Directors UK's own policy page seeks a statutory AI remuneration right that a March 2026 report has not yet granted.
The Producers Guild's Code of Credits names no AI-specific credit, even as its own conference addressed AI contract risk.
La Biennale di Venezia added a single clause on AI to its 2026 festival regulations, absent from the 2025 edition.
A joint documentary-field update shows applicants asked to describe, not certify, their AI use.
The European Film Market's own wrap-up names AI programming at its 2026 edition, not a binding rule.
A 2026 government report responds to a copyright-and-AI consultation without adopting a new audiovisual rule.
The CNC's own call for projects backed exploratory AI residencies for French screen creators, not a broad AI rule.
The ACTRA-CMPA-AQPM agreement negotiates Canadian rules for AI-generated performers, distinct from any US contract.
A live clause in the agency's Terms of Trade requires applicants to follow its AI guidelines, absent from a static PDF copy.
A Commission call for evidence on AI and culture is linked to Creative Europe but is not a MEDIA funding condition.
The BFI's own funding guidance sets transparency conditions for AI use, distinct from BAFTA's award-eligibility question.
Autodesk's own pages frame 2024's ShotGrid-to-Flow Production Tracking switch as continuity, with one new AI scheduling feature layered on top.
Frame.io's own materials describe Content Credentials as data the platform reads and preserves, not something every connected camera writes.
StudioBinder's own call sheet and shot list pages describe automatic data population, not any feature the company itself calls AI.
ftrack's own documentation describes Actions and webhooks as configurable automation, distinct from a competing vendor's AI scheduling claim.
Yamdu's own disclosure page names the models behind its script-breakdown AI beta and states plainly what those features do not generate.
Cuebric's launch coverage and its 2026 homepage describe a widening pitch for the same AI environment-generation engine.
Entertainment Partners's own pages treat generative AI as a compliance and workforce topic, not a feature of its payroll products.
Move.ai's own documentation describes a multi-camera, GPU-dependent capture system and a self-reported comparison to marker-based rigs.
Disney's own archive credits MAGI, Triple-I, Robert Abel and Digital Effects, not an in-house unit, for Tron's CGI.
ILM's own history says an animator-built stained-glass knight, not a capture system, drove Young Sherlock Holmes' effect.
ILM's own history credits three named artists with the water-tentacle effect that won The Abyss its 1989 visual effects Oscar.
ILM's own account of Terminator 2 credits a morph team, a 3D character team and two more for the T-1000's Oscar-winning effect.
VES's own magazine dates the society's ambitious inception to late 1996, ahead of the founding year most often cited.
The Academy's own citation credits Stephen Regelous's agent system, not a learned model, for the trilogy's battle scenes.
Weta's own account describes a viewfinder feed for on-set direction, not an image-generating system, behind Avatar's pipeline.
The bankruptcy court's own docket dates the Chapter 11 filing to 13 February 2013, eleven days before Life of Pi won its Oscar.
ILM's own interview says a fully digital Tarkin was chosen because the character was judged too iconic to recast.
Weta's own case study credits its character pipeline, not ILM's separate de-aging rig, with the fully digital young Will Smith.
Runway's announcement and changelog date Frames' research reveal separately from its paid-plan availability.
Sber's ai-forever released Kandinsky Video's code and weights under Apache 2.0, a fully permissive open license.
The CAI's 2020 white paper introduced assertions and claims, terms the C2PA later formalized in its own specification.
Foundry's own documentation and blog describe Cattery as a free model library layered onto CopyCat's training node.
Weta FX's own technology pages and SIGGRAPH publication list document named tools without a commercial release.
Equity's own campaign page documents an unresolved fight with producers for AI consent and pay terms, not a settled one.
Telefilm's four-principle AI policy sets disclosure and IP terms; CMPA offers members resources, not a parallel rule.
Rokoko's own product pages name Move.ai in a direct comparison while documenting a separate, camera-free wearable capture method.
Wrapbook's own site separates AI-labeled features from the rules-based payroll automation it does not call AI.
A trade profile and the effects house's own materials describe a motorized rig built to repeat one planned move, not to invent a shot.
American Cinematographer's own account of the 2000 film shows a full color-timing pass done digitally, not the first attempt at the process.
The Academy's own record and a cinematography trade account describe a still-camera array, not a claim about any specific present-day technology.
Max Fleischer's 1917 patent and a Library of Congress record describe a mechanical tracing device, the origin of a name software now borrows.
Bob Sabiston's own account of the software behind Waking Life describes an interpolation aid, not an automated or intelligent system.
ILM's own account and the Academy's 1994 award record show fewer than 60 fully CG shots working alongside physical creature effects.
The Academy's 2003 visual-effects award and Weta's own materials name Gollum as a signature performance-driven character, not a solo automated one.
Avid's own financial history and contemporaneous trade coverage of its 1999 Academy Award frame nonlinear editing as an infrastructure shift.
Digital Domain's own case study and the Academy's 2009 award record describe a composited head-and-body pipeline, not a single generated actor.
Pixar's own history page and Academy coverage of the 1996 ceremony credit John Lasseter's leadership for the first full-length CG feature.
Meta's own paper and blog post document a text-to-video method trained without text-video pairs, framed as research, not a shipped product.
Google's Imagen Video paper and project page detail a cascaded text-to-video method the team states it chose not to release.
Stability AI's announcement and model card document Stable Video Diffusion as a research preview with named limits, not a commercial tool.
Pika Labs' own 1.0 announcement documents a product upgrade, funding, and a waitlist, but not technical specifications for the model.
Runway's help-center documentation scopes Motion Brush to painted regions with directional modifiers, later retired in favor of prompting.
The Zeroscope model card describes a watermark-free fine-tune of Modelscope, released independently rather than as a company product.
DAMO Academy's ModelScope model card documents a downloadable research release, with limits the card itself states plainly.
Deforum's own documentation describes a scripted, frame-by-frame repaint of Stable Diffusion, a precedent still used in node-based pipelines.
D-ID's own product page and license agreement document a bounded avatar-video tool and a contractual, not technical, consent gate.
Synthesia's own policy and feature documentation require a recorded consent step before any real person becomes a reusable avatar.
OpenAI's own December 2024 announcement and system card state Sora Turbo's resolution, duration and safety limits at first public release.
OpenAI's own announcement separates Sora 2's stated capabilities from the consent and moderation controls built into its new Sora app.
OpenAI's own support notice states the exact dates and data terms for winding down the Sora web, app and API products.
Google's own December 2024 posts state Veo 2's benchmark method, resolution, and named limitations at that release.
Google's own announcement documents Flow's ingredient-to-scene workflow and its stated plan limits, not confirmed film credits.
Runway's own help article and pricing page state Gen-4.5's duration cap, cost, and its still-limited input types.
Runway's own research page states what Act-One takes as input and what it does not claim to replace.
Luma's own June 2024 launch page states Dream Machine's generation speed, shot length, and named limitations.
Alibaba's own repository states Wan2.1's parameter sizes, hardware needs, and internal benchmark method.
Runway, Luma, and Veo each document named camera controls for generated video, and Runway's own pages describe where that control still breaks down.
Runway documents up to three references per generation, and Luma states plainly that a single image leaves identity consistency incomplete.
Luma and Veo both document a first-and-last-frame feature, and neither vendor's documentation calls the interior motion predictable.
An ASC visual analysis shows how kept screen direction reads as continuity, a discipline vendor documentation of generated shots does not describe as automatic.
Runway's own prompting documentation gives a concrete structure for the subject, action, camera, and style fields a shot list already tracks.
Trade-press accounts of camera matching and CG-to-plate handoffs describe the craft work of cutting a generated shot next to live footage.
Luma's own production guidance recommends locking wardrobe, geography, and lighting decisions before requesting a shot, framed here as an editorial checklist.
Vendor documentation on environment and reference conditioning does not claim a generated establishing shot will match a live-action scene filmed later.
Runway's own Act-Two documentation requires an unbroken view of the face, and an independent study finds pose-transfer models often garble the action anyway.
A cutaway does not need continuous performance, which is why vendor documentation and an editors' guild account both treat it as lower risk.
Adobe's own account of two commissioned shorts names which tasks moved to Firefly and which stayed with the director.
ComfyUI's own documentation ties reproducibility to saved workflow JSON and seeds, not to the community nodes built on top of it.
Runway's own developer documentation sets per-second API pricing, moderation rules, and failure handling distinct from its consumer app.
The Protopica creators' own account ties a Creole-language short to a documented generate-then-curate keeper ratio.
Boords' own product pages position AI-generated storyboard frames as a draft entry point into an otherwise unchanged sign-off workflow.
Lightricks' own product page ties LTX Studio's scene generation to a script, concept, image, or video a user supplies first.
Netflix's own Eyeline announcement credits The Eternaut but does not document the generative workflow behind that claim.
Wonder Dynamics' 2023 site and pricing page describe an 80-90 percent automation figure alongside scene exports built for further finishing.
Adobe's help documentation caps Premiere's Generative Extend at seconds and mutes rather than invents spoken dialogue.
Blackmagic's own pages describe IntelliScript, IntelliSearch, and Magic Mask as assistive tools inside one finishing suite.
Topaz's documentation distinguishes models that preserve source intent from ones that generate new detail on upscaling.
Adobe's 12 February 2025 announcement states Firefly trains only on licensed and public-domain content, not customer files.
Runway's 25 July 2025 research page documents Aleph as an in-context editor for uploaded video, not a text-to-video generator.
Autodesk acquired Wonder Dynamics in 2024 on undisclosed terms; the tool now ships as Autodesk Flow Studio.
Runway's help articles trace object isolation and removal from a dedicated Inpainting tool to prompt-driven Edit Studio.
Adobe's help documentation describes Roto Brush 3's AI tracking model and the manual corrective strokes it still expects.
Flawless's own pages describe TrueSync's lip-resync process and the consent mechanism it says a production should use.
ElevenLabs launched automatic AI dubbing in October 2023; its documentation names 90-plus languages and cloning limits.
The DGA's own 2023 announcements describe AI contract language reached before two allied unions settled their disputes.
IATSE's own bargaining and ratification statements describe AI displacement protections in its 2024 Basic Agreement.
SAG-AFTRA's own bulletin sets consent and pay terms for AI and digital replicas under one interactive-media contract tier.
SAG-AFTRA's own bulletins define a digital replica against a wholly synthetic performer and attach different consent rules.
The editors guild's own materials describe indemnification and displacement protections for AI in post-production.
The ASC's own committee materials describe an AI subcommittee built to assist cinematographers, not replace camera-based work.
The Producers Guild's own one-sheet turns one conference session into ten AI due-diligence questions for producers.
SAG-AFTRA's own strike order names unenforceable AI language as the reason its video-game performers walked out in 2024.
The AFM's own releases show AI terms secured in a 2024 film contract and still being sought in 2026 recording talks.
C2PA's own specification defines what a Content Credentials manifest records and what it does not prove.
The European Commission's own Article 50 guidance limits AI marking and disclosure duties to specific roles and contexts.
YouTube's own policy names three triggers for labeling altered or synthetic video and what counts as a minor edit.
OpenAI's and Scarlett Johansson's own dated statements each document their side of a synthetic-voice dispute.
The Partnership on AI's own framework names consent and disclosure practices while stating plainly it audits no one.
Tribeca's own materials credit five named filmmakers for Sora Shorts while leaving its competition status unaddressed.
The Content Authenticity Initiative's own materials scope its tools to inspecting attached credentials, not detecting missing ones.
New York Film Academy's own program pages describe generative AI coursework built around traditional directing and editing skills.
VBench's own paper defines sixteen separately measured dimensions and states what it does not yet cover.
EvalCrafter's own paper names three limitations in its prompt set, motion metrics, and annotator pool.
The Frechet Video Distance paper scopes its metric to statistical similarity, not narrative or creative quality.
The Sora System Card documents red-teaming, mitigations, and physics failures the lab's own technical report admits.
DeepMind's Veo page names resolution, clip length and example counts for its benchmarks, and one named limitation.
The DFDC dataset paper reports 2,114 competing teams and states detection remained extremely difficult.
Artificial Analysis discloses its Elo methodology and vote counts but not how its voters are recruited.
An editorial checklist for reading vendor comparison claims, built from what named pages disclose or omit.
Physics-IQ's authors report their best-scoring model reached only 29.5 percent against real-world physical variance.
Runway's AI Film Festival names credited filmmakers and winners across years, which is public evidence of use, not of tool autonomy.
A framework for telling a credited production from an artist's account from a vendor's demo reel before drawing conclusions.
The directors' own statement, per Variety, narrows a viral AI controversy to three brief interstitial images.
Asteria and Moonvalley's own statements describe licensed data and a hybrid pipeline, ahead of any released feature.
Ad Age's reporting and Coca-Cola's own comments detail the studios, models and criticism behind the 2024 AI holiday ads.
Metaphysic's own account of Here describes a real-time, on-set de-aging system built for one credited production.
Metaphysic's own materials describe on-set tools and a CAA partnership as company statements, not third-party audits.
Trade coverage and Vertigo Films' own posts trace Critterz from a 2023 short to a Cannes market pitch.
An editorial method for comparing video models by fixing prompts, seeds and settings before judging results.
Tencent's own repository documents HunyuanVideo's parameter count, hardware needs, and its own benchmark method.
Katalist's own pages describe a script-to-shot engine now marketed chiefly for advertising variants, not film shot lists.
Entertainment Partners' scheduling tools and a generative script-to-scene feature both document a scene inventory step, but not a link between them.
Bectu's own 2025 statements pursue AI transparency and consent through consultations rather than a bargained contract clause.
BAFTA's own film and television rules add a generative AI disclosure clause without setting an eligibility threshold.
BBC R&D's own report of its Verify team's first C2PA-credentialed article states what the standard does and does not yet do for video.
Topaz's documentation names specific failure modes of its diffusion-based Starlight video models and a control for it.
Reporting and the guild's own page show a tracking process running against admitted off-the-record AI use.
Try another category or clear your search.
