History / Process note001
Harold Cohen kept the rules and let AARON draw the arrangements
Museum records describe AARON as Cohen's decades-long research tool, not a stand-alone artist.
1973 record
3 min readRead note ↗
Retrospective process notes / 100 notes
The process wall
Model releases, tool launches, rulings, guild terms, studies and the older precedents behind creative AI, each read from the maker's bench: what the tool does, where the human decisions sit, and what survives.
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.
100 notes
History / Process note002
Vera Molnar programmed a plotter to draw with 1% disorder
Pompidou and MoMA records describe a quantified disorder Molnar set herself, not open-ended randomness.
1968 record
3 min readRead note ↗
History / From the wall003
Two designers built a sketchbook language instead of a prompt box
Processing's own records describe a teaching tool for traceable visual code, not one-click images.
Event: 9 August 2001
3 min readRead note ↗
History / From the wall004
A 2014 paper set two networks against each other to make images
The GAN paper and its NeurIPS record describe adversarial training, not memorised training images.
Event: 10 June 2014
3 min readRead note ↗
History / From the wall005
A classifier debugging technique became an aesthetic overnight
Google's 2015 post describes DeepDream as diagnostic gradient ascent, not a built creative tool.
Event: 18 June 2015
3 min readRead note ↗
History / From the wall006
A 2015 paper separated an image's content from its style
Gatys, Ecker and Bethge's paper and its CVPR version describe optimisation, not brushwork transfer.
Event: 26 August 2015
3 min readRead note ↗
History / From the wall007
Christie's named an algorithm's loss function as the signature
The lot record, Christie's own story and contemporaneous reporting describe a contested code lineage.
Event: 25 October 2018
3 min readRead note ↗
History / From the wall008
StyleGAN's own dataset record limits its licence to non-commercial use
NVIDIA's paper and repositories describe a style-based generator and a Flickr-sourced dataset's terms.
Event: 12 December 2018
3 min readRead note ↗
History / From the wall009
MoMA and Anadol's studio name different scopes for one dataset
The museum names publicly available data; the studio names a specific processed image count.
Event: 19 November 2022
3 min readRead note ↗
History / From the wall010
OpenAI documented CLIP's prompt sensitivity from the start
The 2021 announcement and paper describe a scoring model and its own stated limits on counting and wording.
Event: 5 January 2021
3 min readRead note ↗
Image tools / From the wall011
DALL-E 2 turned inpainting into a waitlist product
OpenAI's 2022 announcement and beta record inpainting and variations, and stated safety limits, not the training data.
Event: 1 April 2022
3 min readRead note ↗
Image tools / From the wall012
Midjourney ran its first models entirely inside a chat window
Midjourney's own version history dates the July 2022 open beta and shows why a Discord bot changed iteration.
Event: 1 July 2022
3 min readRead note ↗
Image tools / From the wall013
Stable Diffusion shipped weights, a licence and named limits
The August 2022 release and its model card name the LAION training data, the OpenRAIL licence and stated biases.
Event: 22 August 2022
3 min readRead note ↗
Limits & evidence / From the wall014
A compression trick made image diffusion run on a laptop
The 2021 latent diffusion paper explains the mechanism behind Stable Diffusion's speed and its text artefacts.
Event: 20 December 2021
3 min readRead note ↗
Process & craft / From the wall015
ControlNet gave diffusion models a structural anchor
A frozen-copy training method let makers steer generation with edges, depth and pose, the 2023 paper shows.
Event: 10 February 2023
3 min readRead note ↗
Image tools / From the wall016
Adobe named its generator's training sources at launch
Firefly's 2023 announcement and FAQ tie a commercially safe claim to Stock, licensed and public domain images.
Event: 21 March 2023
3 min readRead note ↗
Image tools / From the wall017
Photoshop put generated fills in their own reversible layer
The 2023 Generative Fill release and help guide document a non-destructive layer, not a pixel edit.
Event: 23 May 2023
3 min readRead note ↗
Image tools / From the wall018
A chatbot started rewriting prompts before DALL-E 3 saw them
OpenAI's 2023 announcement and system card describe prompt rewriting, artist refusals and a training opt-out.
Event: 20 September 2023
3 min readRead note ↗
Image tools / From the wall019
One model family shipped under three different licences
Black Forest Labs' 2024 launch and model cards tie FLUX.1's tiers to commercial, non-commercial and open terms.
Event: 1 August 2024
3 min readRead note ↗
Image tools / From the wall020
OpenAI listed 4o image generation's failures at launch
The 2025 announcement and system card name text, graphing and cropping limits alongside the capability claims.
Event: 25 March 2025
3 min readRead note ↗
Video & motion / From the wall021
A tools company shipped video-to-video, then text-to-video
Runway's own research pages describe Gen-1's structure-conditioned editing and Gen-2's move to text-to-video.
Event: 1 February 2023
3 min readRead note ↗
Video & motion / From the wall022
A minute-long demo became a bounded product, then vanished
OpenAI's technical report and release post describe Sora narrowing from a research demo to a capped product, later withdrawn.
Event: 15 February 2024
3 min readRead note ↗
Video & motion / From the wall023
A video model paired synchronised sound with the picture
Google's model card and technical report describe Veo 3's joint audio-video diffusion and its stated limits.
Event: 23 May 2025
3 min readRead note ↗
Video & motion / Process note024
Two video generators price a shot by the rendered second
Kling's and Hailuo's own pricing and terms describe metered compute and a broad reserved licence, not a finished cut.
Process note
3 min readRead note ↗
Video & motion / From the wall025
An open motion module animated other people's fine-tunes
The AnimateDiff paper and repository describe adding motion to personalised diffusion models without retraining them.
Event: 10 July 2023
3 min readRead note ↗
Design & 3D / From the wall026
A neural field learned a scene from photographs alone
The NeRF paper describes optimising a volumetric scene representation from posed images to synthesise new views.
Event: 19 March 2020
3 min readRead note ↗
Design & 3D / From the wall027
A splatting method rendered captured scenes in real time
The SIGGRAPH paper describes representing captures as 3D Gaussians for real-time rendering, under a research-only licence.
Event: 8 August 2023
3 min readRead note ↗
Design & 3D / From the wall028
Text-to-3D borrowed a 2D diffusion model to sculpt shapes
The DreamFusion paper describes distilling a 2D diffusion prior into a NeRF, with named artefacts and no 3D training data.
Event: 29 September 2022
3 min readRead note ↗
Design & 3D / From the wall029
A free tool assembled rigged digital humans in minutes
Epic's early-access announcement and current documentation describe MetaHuman Creator's pipeline and Unreal Engine dependency.
Event: 14 April 2021
3 min readRead note ↗
Design & 3D / From the wall030
An LED wall let a crew shoot finished backgrounds live
Epic's and ILM's own accounts describe how real-time rendering on set replaced location shoots for The Mandalorian.
Event: 20 February 2020
3 min readRead note ↗
Authorship & rights / From the wall031
A registrar's letter split a comic's text from its generated images
A 2023 Copyright Office letter on 'Zarya of the Dawn' kept the comic's text and arrangement but cancelled the Midjourney images' registration.
Event: 21 February 2023
3 min readRead note ↗
Authorship & rights / From the wall032
A federal notice told applicants to disclaim AI-made material
A March 2023 Copyright Office policy statement requires applicants to disclose AI-generated content and claim only their own human contribution.
Event: 16 March 2023
3 min readRead note ↗
Authorship & rights / From the wall033
An appeals court said only humans can author a copyrighted work
The D.C. Circuit affirmed that a machine cannot be a copyrighted work's author, but left open what a maker who directs a model can claim.
Event: 18 March 2025
3 min readRead note ↗
Authorship & rights / From the wall034
A federal report says prompts alone do not make an AI user an author
The Copyright Office's January 2025 report finds prompts alone insufficient for authorship but protects human selection, arrangement and edits.
Event: 29 January 2025
3 min readRead note ↗
Authorship & rights / From the wall035
Artists' claims against image-model makers survive to discovery
Two federal orders in Andersen v Stability AI let artists' copyright claims proceed against four AI image companies while dismissing DMCA claims.
Event: 13 January 2023
3 min readRead note ↗
Authorship & rights / From the wall036
A UK court found narrow trade mark use but no infringing AI copy
The High Court found Stability AI infringed Getty's trade marks in limited, historic cases, but that Stable Diffusion is not itself an infringing copy.
Event: 4 November 2025
3 min readRead note ↗
Authorship & rights / From the wall037
EU rules now require AI model makers to publish training summaries
From 2 August 2025 the EU AI Act requires general-purpose model providers to summarise training content and honour artists' machine-readable opt-outs.
Event: 2 August 2025
3 min readRead note ↗
Authorship & rights / Process note038
A registry lets artists flag training images, if trainers cooperate
Spawning's Have I Been Trained and ai.txt let artists search datasets and signal an opt-out, a tool the company says depends on voluntary honour.
Process note
3 min readRead note ↗
Authorship & rights / From the wall039
Japan's law lets machines analyse works no person will ever enjoy
Japan's 2019 copyright amendment lets AI systems analyse works without a licence unless a person perceives them; 2024 agency guidance sets its limits.
Event: 1 January 2019
3 min readRead note ↗
Authorship & rights / From the wall040
Two tools cloak and poison images against unwanted style training
Researchers built Glaze to shift how a model learns an artist's style and Nightshade to corrupt a model's prompt-to-image associations.
Event: 8 February 2023
3 min readRead note ↗
Design & 3D / From the wall041
A design tool rolled back its layout generator after launch
Figma's own posts describe the commissioned design-system library behind Make Designs and why the feature was pulled back.
Event: 26 June 2024
3 min readRead note ↗
Design & 3D / From the wall042
A template platform bundled generation with a royalty fund
Canva's announcement and its Shield page describe what Magic Studio bundles and what its creator-payment and indemnity claims cover.
Event: 4 October 2023
3 min readRead note ↗
Authorship & rights / From the wall043
A software maker rewrote its terms after a licence backlash
Adobe's own posts and its current terms describe what changed after a re-acceptance prompt alarmed creators over content access.
Event: 6 June 2024
3 min readRead note ↗
Authorship & rights / From the wall044
A portfolio site met an AI protest with an opt-in meta tag
ArtStation's own FAQ pages describe the NoAI tag it built after community concern and what the tag actually does.
Event: 14 December 2022
3 min readRead note ↗
Process & craft / From the wall045
A drawing app publicly refused to add generative features
Procreate's own statement says why it will not build generative AI and describes the by-design privacy claim behind that refusal.
Event: 1 August 2024
3 min readRead note ↗
Process & craft / From the wall046
A manga software maker withdrew its image generator in days
Celsys's own notices describe the Stable Diffusion feature it announced and pulled after listing the community's objections.
Event: 2 December 2022
3 min readRead note ↗
Process & craft / Process note047
A canvas tool turned the prompt into a live brush
Krea's documentation describes how a live canvas drives a fast image-generation loop instead of a single render step.
Process note
3 min readRead note ↗
Process & craft / From the wall048
A node graph made a generation pipeline visible and shareable
ComfyUI's repository and documentation describe how a saved graph, or even a PNG file, carries the exact steps.
Event: 17 January 2023
3 min readRead note ↗
Process & craft / From the wall049
A low-rank method let makers fine-tune models cheaply
The LoRA paper and its image-community adoption describe a small file that redirects a frozen model toward a style.
Event: 17 June 2021
3 min readRead note ↗
Process & craft / From the wall050
A tuning method taught models a subject from a few photos
DreamBooth's paper and project page describe the method's identifier binding and its own stated misuse risk.
Event: 25 August 2022
3 min readRead note ↗
Authorship & rights / From the wall051
The screenwriters' contract set limits on what AI can write
The 2023 WGA agreement keeps AI text out of credit and lets writers refuse the tools without being forced by a studio.
Event: 9 October 2023
3 min readRead note ↗
Authorship & rights / From the wall052
Performers' contract made consent the price of a digital replica
SAG-AFTRA's 2023 agreement defines two kinds of digital replica and requires advance notice, consent and pay for each.
Event: 5 December 2023
3 min readRead note ↗
Authorship & rights / From the wall053
Animation workers won notice and a veto on generative AI jobs
The Animation Guild's 2024 contract requires notice before AI is a job requirement and bars using crew likenesses to train it.
Event: 1 December 2024
3 min readRead note ↗
Authorship & rights / From the wall054
The film academy tied AI use to a human-authorship test
The Academy's 2025 Oscars rules say generative AI neither helps nor hurts a nomination, judged by human creative authorship.
Event: 21 April 2025
3 min readRead note ↗
Process & craft / From the wall055
A film's own account of an AI voice tool drew a narrow line
The Brutalist's editor and director described an AI tool used only to refine Hungarian vowels, not to alter performance or English dialogue.
Event: 21 January 2025
3 min readRead note ↗
Process & craft / From the wall056
A superhero series swapped storyboards for prompts on screen
Method Studios says a custom AI tool built Secret Invasion's title sequence and that no artists' jobs were replaced.
Event: 21 June 2023
3 min readRead note ↗
Video & motion / From the wall057
Netflix told investors a generative shot reached the screen
Netflix's co-CEO said its in-house VFX group used generative AI to finish a building-collapse shot in El Eternauta ten times faster.
Event: 17 July 2025
3 min readRead note ↗
Authorship & rights / From the wall058
A studio let a model company train on its entire catalogue
Lionsgate's 2024 deal let Runway build a model on its film library, and neither announcement addresses the people in that library.
Event: 18 September 2024
3 min readRead note ↗
Video & motion / From the wall059
A studio's own animation pipeline became public software
OpenToonz opened Studio Ghibli's long-customised Toonz software under a BSD licence, without touching the rights to any film made with it.
Event: 18 March 2016
3 min readRead note ↗
Video & motion / From the wall060
A brand film showed what a generate-and-reject loop still needs
Toys R Us and Native Foreign's Sora-made film relied on human rejection of bad iterations, but carried no on-screen AI disclosure.
Event: 1 June 2024
3 min readRead note ↗
Limits & evidence / From the wall061
A 2023 paper extracted training photographs from diffusion models
A paper reports exact counts of near-duplicate training images extracted from Stable Diffusion and Imagen, and compares the rate to GANs.
Event: 30 January 2023
3 min readRead note ↗
Limits & evidence / From the wall062
An audit counted unlawful images indexed inside a training dataset
A Stanford audit counted suspected unlawful images among LAION-5B's indexed links, and the dataset's own release notes state its intended use and limits.
Event: 20 December 2023
3 min readRead note ↗
Limits & evidence / Process note063
Papers named why generators drew bad text and broken hands
SDXL's own limitations section and a text-rendering paper name different causes for bad hands and illegible text, and what each requires to fix.
Process note
3 min readRead note ↗
Limits & evidence / Process note064
Two studies measured how generators stereotype people and jobs
Two 2022-2023 papers measured how generators represent occupations, gender and ethnicity against labour data, with stated methods and findings.
Process note
3 min readRead note ↗
Limits & evidence / Process note065
A distance score ranks generators without judging one image
FID compares feature statistics across image sets; ImageReward and Pick-a-Pic instead learn from human comparisons, each bounded by its own rater pool.
Process note
3 min readRead note ↗
Limits & evidence / Process note066
A provenance standard records edits, not what happened before
The C2PA specification defines a chained, signable record of edits; its own documents state adoption is opt-in and its metadata can be stripped.
Process note
3 min readRead note ↗
Limits & evidence / From the wall067
A lab watermarked pixels instead of stamping a visible logo
DeepMind's SynthID embeds an imperceptible signal in generated pixels; its own documents state what it survives and what it cannot cover.
Event: 29 August 2023
3 min readRead note ↗
Limits & evidence / Process note068
A photo contest drew its line at added or removed information
World Press Photo's current rules bar synthetic images outright but permit tools that neither add nor remove camera-captured information.
Process note
3 min readRead note ↗
Limits & evidence / From the wall069
An artist declined a photography prize for an AI image
Boris Eldagsen won and declined the Sony World Photography Awards' Creative category in 2023 after disclosing AI use, as both statements record.
Event: 13 April 2023
3 min readRead note ↗
Limits & evidence / From the wall070
Researchers named a decay from training on model output
Shumailov et al. define 'model collapse' in language models, autoencoders and Gaussian mixtures, and state its implications for training data.
Event: 24 July 2024
3 min readRead note ↗
Process & craft / Process note071
A fiction tool bundles drafting into named, tracked steps
Sudowrite's own terms keep content ownership with the writer while licensing broad use of the file, as retrieved on 16 September 2026.
Process note
3 min readRead note ↗
Authorship & rights / From the wall072
Authors and their guild allege OpenAI trained on their books
A federal docket and the Guild's own statement record the claims against OpenAI and the case's status as of retrieval.
Event: 19 September 2023
3 min readRead note ↗
Authorship & rights / From the wall073
A court split lawful book training from pirated acquisition
The Guild's own account of the order and settlement records what was ruled fair use, what was not, and what followed.
Event: 23 June 2025
3 min readRead note ↗
Authorship & rights / Process note074
A self-publisher tests authorship by who wrote the first draft
KDP's own guidelines define AI-generated versus AI-assisted content and set the disclosure duty on origin, not on edits.
Process note
3 min readRead note ↗
Authorship & rights / From the wall075
A magazine closed submissions under a wave of generated stories
The editor's own tracking records a submissions spike and the closure it prompted within days.
Event: 20 February 2023
3 min readRead note ↗
Authorship & rights / From the wall076
A writing nonprofit's own statements resist one closure cause
The organisation's posts on a 2024 AI dispute and a 2025 closure report describe separate, only partly linked pressures.
Event: 1 March 2025
3 min readRead note ↗
Process & craft / Process note077
A writing assistant now records how a draft actually formed
Grammarly's own documentation describes a process record, what it captures, and where it states its own tracking can fail.
Process note
3 min readRead note ↗
Authorship & rights / Process note078
A writers' guild spells out what an AI clause should say
Model contract clauses and disclosure guidance describe what a publishing deal should promise about AI use, in the Guild's own language.
Process note
3 min readRead note ↗
Limits & evidence / From the wall079
Readers rated AI poems well and could not spot them reliably
A Scientific Reports study of 1,634 readers found below-chance detection and higher ratings for machine poems styled on ten poets.
Event: 14 November 2024
3 min readRead note ↗
Limits & evidence / From the wall080
Shared AI prompts lifted stories but narrowed them together
A controlled writing experiment found AI ideas raised individual story ratings while making assisted stories more alike.
Event: 12 July 2024
3 min readRead note ↗
Authorship & rights / From the wall081
A stock agency priced indemnity into its own generator
Getty's launch and generation page show a licensed-only claim paid for by contributor royalties and tiered legal cover.
Event: 25 September 2023
3 min readRead note ↗
Authorship & rights / From the wall082
A stock library paid contributors before it sold their trained model
Shutterstock's 2022 release and current FAQ separate training-data royalties from a hard refusal of AI submissions.
Event: 25 October 2022
3 min readRead note ↗
Authorship & rights / Process note083
Three stock libraries drew three different lines around AI images
Adobe labels and accepts generated submissions, Shutterstock refuses them, and Getty keeps its own generator's output out of the library.
Process note
3 min readRead note ↗
Authorship & rights / Process note084
Midjourney made image ownership a paid subscription tier
The current terms tie asset ownership above a revenue threshold to the Pro or Mega plan, named in the pricing page.
Process note
3 min readRead note ↗
Authorship & rights / From the wall085
Two AI vendors offered legal cover with a price and a limit
OpenAI's Copyright Shield and Adobe's Firefly indemnity both attach conditions, exclusions and, for Adobe, a stated dollar cap.
Event: 6 November 2023
3 min readRead note ↗
Process & craft / Process note086
A model-sharing site polices faces closely and styles barely at all
Civitai's rules ban real-person likeness proactively but leave a living artist's style protected only by a reactive takedown notice.
Process note
3 min readRead note ↗
Process & craft / From the wall087
A model card discloses; an OpenRAIL licence restricts
Hugging Face's documentation and the 2022 OpenRAIL post show disclosure and enforceable use-limits are two separate layers a maker must check.
Event: 31 August 2022
3 min readRead note ↗
Process & craft / Process note088
A model's parameter count sets the local hardware bill
FLUX's model card, ComfyUI's requirements and a flagship card's memory spec show VRAM, not clock speed, is the local-generation limit.
Process note
3 min readRead note ↗
Process & craft / Process note089
Three creative tools sell credits and GPU time, not images
Midjourney, Runway and Firefly pricing pages show proprietary usage units, tiered feature gates and discounted first-year rates, as retrieved.
Process note
3 min readRead note ↗
Authorship & rights / From the wall090
Adobe paid stock contributors then let its generator compete
The 2023 bonus announcement gives its calculation basis while opening Adobe Stock to the labelled Firefly images it also describes.
Event: 13 September 2023
3 min readRead note ↗
Process & craft / Process note091
Generative fill lands in a new layer, not a finished image
Photoshop's own documentation and ComfyUI's docs describe generation as staged output a maker must select, composite and mostly discard.
Process note
3 min readRead note ↗
Authorship & rights / Process note092
The registrar asks applicants to name and exclude the AI parts
The Office's 2023 guidance and the Zarya of the Dawn decision show what disclosure and exclusion actually require.
Process note
3 min readRead note ↗
Limits & evidence / Process note093
Two model cards answer the licence question very differently
FLUX.1 dev and Stable Diffusion 3.5 Large show what a model card does and does not say about training data and use.
Process note
3 min readRead note ↗
Authorship & rights / Process note094
Midjourney says its style reference copies mood, not an artist
The tool's own documentation, the registrar's guidance, and Glaze's technical page draw three different boundaries around style.
Process note
3 min readRead note ↗
Authorship & rights / Process note095
A photo contest bans AI images; a platform just labels them
World Press Photo, Content Credentials and YouTube's policy each set a different bar for disclosing generated work.
Process note
3 min readRead note ↗
Authorship & rights / Process note096
Tennessee's ELVIS Act reaches the tool, not just the output
A 2024 statute, a union agreement and a platform policy each require consent before generating a real person's likeness.
Process note
3 min readRead note ↗
Video & motion / Process note097
Sora's own limit is twenty seconds at 1080p, per its card
Sora, Runway Gen-3 Alpha and Veo 3's own documents set different duration, resolution and control limits for a generated plate.
Process note
3 min readRead note ↗
Process & craft / Process note098
Diffusion models can memorise a training image outright
Papers on memorisation and model collapse show what iterating on generated data preserves and what it erodes.
Process note
3 min readRead note ↗
Authorship & rights / Process note099
A rights reservation only counts if it's machine-readable
EU law, OpenAI's own help pages and a code of practice each define opt-out differently, at different depths.
Process note
3 min readRead note ↗
Process & craft / Process note100
OpenAI's indemnity for generated output has named exceptions
Firefly's FAQ, OpenAI's service terms and the registrar's guidance define what a client briefing actually needs to cover.
Process note
3 min readRead note ↗
No matching notes
Try another category or clear your search.