Industries / Media, arts and recreation / Media and entertainment / Film production
Visual effects
Visual effects (VFX) is the work of making images that could not be filmed and joining them to filmed footage so that each shot reads as one picture.
7
Milestones achieved, from 1895 to 2025.
3
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1895
Stop, swap and splice
In Edison's 18-second film The Execution of Mary, Queen of Scots, the director Alfred Clark stopped the camera as the axe rose and swapped a mannequin in for the actor; the Public Domain Review calls it one of the first uses of editing for special effects. In Paris the stage magician Georges Méliès built his own camera and glass studio, and the Library of Congress credits him with pioneering tricks such as double exposure, black screens and forced perspective, which places near and far objects so that they seem the same size.
Credit
Alfred Clark (Edison Manufacturing Company); Georges Méliès (Star Film)
1940
Shooting actors against blue
A travelling matte is a mask that changes every frame, letting an actor filmed against a plain blue screen be cut out and laid over another background. Alexander Korda's The Thief of Bagdad used blue screen and won the special effects Oscar, with photographic effects credited to Lawrence Butler. Petro Vlahos later refined the method with his "color difference" process, which cut the halo around hair, smoke and glass, and AWN reports that he shared a 1965 Academy award for it with Ub Iwerks and Wadsworth E. Pohl.
Credit
Lawrence Butler; Petro Vlahos; Ub Iwerks; Wadsworth E. Pohl
1977
Camera moves a computer repeats
For Star Wars, George Lucas's new Industrial Light & Magic (ILM), founded in 1975, built the Dykstraflex: a computer-controlled VistaVision camera on a seven-axis dolly and head, on its own track, that could repeat a move exactly. Crews filmed each model ship in its own pass with the same move and combined the passes, so the ships could bank and dive with real motion blur. ILM used the rig and its upgrades until about 1992.
Credit
John Dykstra, Alvah J. Miller and Jerry Jeffress, who received an Academy Scientific and Technical Award in 1978; Star Wars won the visual effects Oscar
1993
Computer animals in live action
For Jurassic Park, the ILM artists Steve Williams and Mark Dippé showed Steven Spielberg a test of digital dinosaurs, and ILM dropped stop motion for the creatures during production. ILM made fewer than 60 shots of fully computer-generated dinosaurs, by its own count, using camera tracking to sit the animals in the live footage, and Phil Tippett's studio helped build a Dinosaur Input Device that let stop-motion animators drive the computer models. Tippett later said of the switch, "If you didn't do [CGI], you'd just be out of the picture."
Credit
ILM (Dennis Muren); Tippett Studio (Phil Tippett); Steve Williams; Mark Dippé; the visual effects Oscar went to Muren, Stan Winston, Tippett and Michael Lantieri
2002
An actor drives a creature
For Gollum in The Lord of the Rings: The Two Towers, Weta Digital in New Zealand filmed Andy Serkis on set with the other actors, then had him perform again on a motion-capture stage, where cameras follow markers on a suit and turn movement into data. That data drove only the body, and 18 animators fixed, changed or replaced it; the face had no capture at all and was animated by hand with 875 sculpted shapes.
LessonWho authors a captured performance stays disputed, and Cook wrote that Gollum "was a synthesis, a collaborative performance delivered by both Andy and a team of highly-skilled animation artists.".
Credit
Andy Serkis; Weta Digital, including the animation director Randall William Cook; the visual effects Oscar went to Jim Rygiel, Joe Letteri, Cook and Alex Funke
2025
A global VFX group collapses
In February 2025 Technicolor Group, owner of the effects companies MPC and The Mill and of Mikros Animation, failed to find new investors and filed for a court "recovery procedure" in France; its UK business went into administration, and most of its roughly 440 UK staff lost their jobs. Estimates of the jobs hit worldwide range from 4,500, an industry consultant's estimate reported by CG Channel, to "more than 10,000," PetaPixel's figure citing India's News9, including over 2,000 staff in India facing unpaid salaries.
LessonOne parent company's failure put artists in the UK, US, France, Canada and India out of work within days.
Credit
Technicolor Group, whose chief executive Caroline Parot blamed the slow post-COVID recovery, a costly separation from its former parent group and fewer orders after the 2023 writers' strike
2025
Cheap AI effects, few tickets
Run to the West, a South Korean film released on 15 October 2025 and co-directed by Kang Yun-sung and the AI filmmaker Kwon Han-seul, used generative AI for chase scenes, car explosions and collapsing buildings, and was finished in five months for an estimated 1.5 billion won, by the Korea Times' estimate. It drew 27,996 admissions against a break-even point of at least 200,000.
LessonA year later Kang said the AI imagery already looked dated at release and that "even the creatures didn't blend well with the footage we'd shot.".
Credit
Kang Yun-sung and Kwon Han-seul
Watch for
Relightable splat crowds
A Gaussian splat is a 3D recording built from millions of small, soft, colored shapes, and a 4D splat records a moving performance. Union VFX and Clear Angle Studios say they capture performers in Clear Angle's 40-camera half-dome rig, relight the splats inside computer-made scenes with ordinary production lights, and can fill stadium seats with "20,000 or 30,000" of them for an unnamed feature now in production; Union's chief executive Tim Caplan said, "As far as we know, we're the only studio currently offering truly relightable Gaussian Splats in production."
Achieved whenThe feature is released and an outside party, such as a trade-press breakdown, confirms relit splat crowds in final shots and gives a shot count.
Claimed by
Union VFX and Clear Angle Studios, June 2026
Generative AI de-aging
Netflix says that in Happy Gilmore 2 "filmmakers used GenAI coupled with ML and Eyeline's proprietary volumetric capture technologies to de-age characters during the opening flashback scene"; Eyeline is Netflix's own effects company. By Disney Research's account, standard de-aging relies on frame-by-frame painting that can take skilled artists days.
Achieved whenSomeone outside Netflix, such as a trade breakdown, a conference paper or another studio's credited use, documents which shots used the method and how its time and quality compare with frame-by-frame work.
Claimed by
Netflix, October 2025 (third-quarter shareholder letter)
AI cuts an action budget
At Netflix's Creative Asia panel during the 2026 Busan festival, the director Kang Yun-sung said his action film The Gulf of Aden, made entirely with AI using the Chinese software Kling AI, cut financing costs by "around 40 percent," from replacing computer graphics and effects and from a shorter shoot. The film is in production.
Achieved whenThe film is released and someone outside the production, such as the Korean Film Council, a financier or a trade audit, publishes its final cost beside a comparable conventional budget.
Claimed by
Kang Yun-sung, October 2026
Open
01
Grade AI wire removal blind
A public set of 20 short high-resolution plates (footage shot for effects work), each showing safe rigging in frame: a toy hung on fishing line, a light stand, a rope tied to a prop. Each plate gets a second clean pass of the same camera move with the rig taken out, shot on a motorized slider that repeats its move, so the correct answer exists. Run AI removal tools on every plate, then have at least three working paint artists, who paint rigs out of shots by hand, grade each result blind as "would final" or "would send back."
The briefCredit · unclaimed02
Count VFX names left off
A public spreadsheet of at least 30 films released from 2024 to 2026 that sets the number of VFX names in each film's end credits beside the effects company's own published crew list, and reports the share missing per film, per company and, where the company lists it, per country.
The briefCredit · unclaimed03
Check AI camera solves
Matchmove means working out the real camera's path and lens from footage so that computer-made objects can be locked to it. Film 20 short handheld and dolly shots of a room or yard on a camera that records lens focal length, with 30 or more points in the space whose 3D positions are measured by tape and laser distance meter. Run AI camera solvers and a standard tracker on every shot, and report how far, in pixels, measured points that the solver never saw land from where the solved camera puts them.
The briefCredit · unclaimed
Open 01 · the brief
Grade AI wire removal blind
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of 20 short high-resolution plates (footage shot for effects work), each showing safe rigging in frame: a toy hung on fishing line, a light stand, a rope tied to a prop. Each plate gets a second clean pass of the same camera move with the rig taken out, shot on a motorized slider that repeats its move, so the correct answer exists. Run AI removal tools on every plate, then have at least three working paint artists, who paint rigs out of shots by hand, grade each result blind as "would final" or "would send back."
Why it is still open
The largest academic wire-removal set, WRV2, has 4,232 clips, but its authors resize them to 432 × 240 pixels for training and testing, far below film resolution, and describe manual wire removal as "labor-intensive and time-consuming." Netflix's open-source VOID model (April 2026) and Runway Aleph (July 2025) remove objects from video, and each was tested by its makers. This research found no test graded by artists at film resolution with a known clean answer.
Built with what exists
Netflix VOID (Apache 2.0 licence; needs a GPU with 40 GB or more of memory, such as a rented A100); Runway Aleph (paid); a motorized slider with repeatable moves; Gemini Decision 1 as an automatic second grader, to see whether it agrees with the artists (see the Note "A thousand questions, one pass"). Video models such as VOID and Aleph are what make the test worth running now.
First step
Shoot five plates on a slider, each a toy on fishing line followed by the same move without it, post them under a Creative Commons licence, and recruit three paint artists.
Done looks like
20 plates, each run through at least three tools and graded by at least three artists, with a published table of each tool's share of "would final" grades and its PSNR, a standard score of how far two images differ, against the clean pass.
Open 02 · the brief
Count VFX names left off
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public spreadsheet of at least 30 films released from 2024 to 2026 that sets the number of VFX names in each film's end credits beside the effects company's own published crew list, and reports the share missing per film, per company and, where the company lists it, per country.
Why it is still open
The largest count found is Stephen Follows' study of 4,111 US releases from 2006 to 2015, which used IMDb's "uncredited" tag, catching only people who later asked IMDb to add them; it found 9% of VFX credits uncredited, and insiders told Follows that 40 names per effects company is a common cap. Single cases look worse: Oppenheimer's credits list 26 VFX crew plus the supervisor Andrew Jackson, while DNEG's own site listed over 160 people, by Cartoon Brew's count. In 2017 the UK union Bectu said that often less than 50% of artists are credited. This research found no newer film-by-film comparison.
Built with what exists
Each film's end crawl, typed in by a person from a legally bought or streamed copy; the effects companies' published crew pages; Muse Ember to pull names from long pages and match the two lists (see the Note "The jobs agents repeat"); a shared spreadsheet.
First step
List ten films from 2024 to 2026 whose main effects company posts a crew list, and count the first one.
Done looks like
30 films, each with two counts, links to both sources and a written method, and a second person who recounts five films and gets counts within 2% of the first.
Open 03 · the brief
Check AI camera solves
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Matchmove means working out the real camera's path and lens from footage so that computer-made objects can be locked to it. Film 20 short handheld and dolly shots of a room or yard on a camera that records lens focal length, with 30 or more points in the space whose 3D positions are measured by tape and laser distance meter. Run AI camera solvers and a standard tracker on every shot, and report how far, in pixels, measured points that the solver never saw land from where the solved camera puts them.
Why it is still open
New AI solvers report angle-error scores on research benchmarks such as Sintel and ETH3D, and VGGT-Ω (CVPR 2026) reports a 77% gain on Sintel, while matchmove software judges a solve by reprojection error in pixels, as Foundry's Nuke manual describes. InFlux++ (July 2026) has real high-resolution footage with lens data for every frame but camera-pose ground truth only for part of its synthetic set. This research found no public test of AI solvers against measured points on real plates.
Built with what exists
Meta's VGGT, which estimates cameras from video (a checkpoint licensed for commercial use is available); NVIDIA's ViPE (Apache 2.0 licence); Blender's free camera tracker as the baseline, with OpenCV grid calibration as Blender's manual recommends; Claude Fable 5.2 to write the scoring scripts (see the Note "Three working devices").
First step
Print a calibration grid, measure 30 points in one room, shoot five 10-second camera moves, and run VGGT and Blender on them.
Done looks like
20 shots published with the footage, the measured points and every solve, and a table of the median reprojection error in pixels on held-out points per tool and per shot, plus the number of shots under 1 pixel.