A museum of AI video from before it learned physics. Faces slid, forks fused with mouths, and a stock-footage watermark floated through every scene. The models that made it are kept here, run again, and put in order.
Scroll to move forward in time. Each model's clips run down its own column.
Between 2022 and early 2024, open video models could hold a scene together for about two seconds. Inside that window everything drifted. A cat became a second cat, a hand became the fork, a family swapped faces as the pizza went round. People called it slop, and made a great deal of it.
Newer models fixed the drift, and the look went with it. A 2026 model can't be prompted back into 2023, because the dream logic came from what those networks couldn't yet do. The weights still exist, spread across Hugging Face accounts that can disappear. In August 2024 Runway deleted its Stable Diffusion 1.5 repository, the base that nearly every animation tool of the era was built on. It survives because other people had copies.
So this site keeps copies. Every model here is mirrored with the code that ran it, each file checked against the hash Hugging Face published. Then each one was run again on the same six prompts with the same seed.
Google researchers turn an image classifier inside out and ask it to amplify whatever it half recognizes. Clouds grow eyes, buildings sprout dog heads. Run frame by frame over footage, it makes the first machine video that looks like a fever.
Source: Inceptionism, Google Research
A recurrent neural network trained on science fiction screenplays, which named itself Benjamin, writes a short film. Director Oscar Sharp and researcher Ross Goodwin shoot it as written, with Thomas Middleditch delivering lines that almost make sense. The pictures are human. The logic is not.
Source: Sunspring, Wikipedia
A Reddit user called deepfakes posts face-swapped videos made with an open-source autoencoder, and the username becomes the word. In April 2018 BuzzFeed and Jordan Peele publish a fake Barack Obama warning viewers not to trust videos like it. The faces swim at the edges, and for now that is how you spot one.
Sources: Know Your Meme, The New Daily
OpenAI releases CLIP, a model that scores how well an image fits a caption. Within weeks, notebooks like Big Sleep and VQGAN+CLIP use that score to steer an image generator toward a prompt. Animators feed each frame back in with a small zoom, and the camera falls forever into a landscape that keeps changing its mind.
Source: CLIP, OpenAI
The genealogy site MyHeritage, using technology licensed from D-ID, animates faces in old photographs: a turn of the head, a blink, a small smile. A million photos are animated in two days. People call it moving and creepy in the same sentence.
Sources: MyHeritage, PetaPixel
The Disco Diffusion notebook by Somnai and Gandamu lands on GitHub with 2D and 3D animation modes for CLIP-guided diffusion. It runs on borrowed Google Colab GPUs and gives the period its painted, swirling look.
Source: alembics/disco-diffusion. Code archived.
Stability AI publishes the weights of Stable Diffusion. The Deforum notebook builds animation on top: each frame is the previous one, nudged by a zoom, a turn or a depth warp, then redrawn by the model. Nothing holds still, and that becomes the point.
Sources: Stability AI, deforum-stable-diffusion. Code and Stable Diffusion 1.5 weights archived.
Meta's Make-A-Video, then Google's Imagen Video and Phenaki, turn text prompts into short clips. They publish papers and demo reels, not weights. For now, text-to-video is something you watch other people have.
Papers: Make-A-Video, Imagen Video, Phenaki
A Twitch channel starts streaming an endless, pixelated Seinfeld parody whose script is written live by GPT-3. The laugh track lands at random and the characters walk into walls. By February tens of thousands are watching, until a generated stand-up bit gets it suspended for two weeks.
Sources: Wikipedia, TechCrunch
Runway's Gen-1 takes an existing video and redraws it to match a reference image or a prompt, as a paid web app. The motion is borrowed from the source. Only the surface dreams.
Corridor Digital films actors on a green screen and redraws every frame with a Stable Diffusion model fine-tuned on one anime film. Seven minutes of flickering, hand-tuned style transfer get millions of views and a furious response from animators. It is the first AI film that a lot of people argue about.
Sources: Know Your Meme, Kotaku
A YouTuber called demonflyingfox runs the cast of Harry Potter through Midjourney, a cloned voice and a talking-head animator, and dresses them for a fashion show. Faces barely move, the stares are vacant, the bass is loud. It is copied with every franchise for months.
Sources: Know Your Meme, Highsnobiety
Alibaba's DAMO Academy puts a 1.7 billion parameter text-to-video model on Hugging Face. It makes two seconds of 256 × 256 video on a single consumer GPU. It learned from stock footage, and the Shutterstock watermark came along: it floats, half-formed, through most of what the model makes.
Source: ali-vilab/text-to-video-ms-1.7b. Weights archived, both the original release and the diffusers version.
Gen-2 makes video from a text prompt alone. Its clips are sharper than ModelScope's and just as unsure what a hand is.
Source: Runway Research
Picsart's researchers make a plain Stable Diffusion image model produce short clips by linking its frames' attention and nudging the noise along a path, with no video training at all. The result twitches rather than melts.
Sources: paper, Picsart-AI-Research/Text2Video-Zero. Code archived.
Reddit user chaindrop posts a ModelScope clip of Will Smith eating spaghetti to r/StableDiffusion. His face slides, the fork merges with his mouth, the pasta never gets smaller. It becomes the field's informal benchmark: every video model since has been asked to feed him.
Sources: the original post, Wikipedia, Know Your Meme. The clips on this site use an anonymous man instead.
A creator called Pizza Later posts a fake pizza commercial made with Midjourney stills, Runway Gen-2 motion and a synthetic voiceover. Diners chew with the wrong mouths and the cheese melts into their hands. It spreads everywhere. A month later, Pizza Hut Canada turns one of its Toronto restaurants into a real Pepperoni Hug Spot.
Sources: Know Your Meme, PC Gamer, Daily Hive
The opening titles of a Disney+ Marvel series are made with generative AI by Method Studios: green faces that shift and dissolve in the era's signature way. It airs in the middle of the Hollywood writers' strike, and artists, including people who worked on the show, are angry.
Sources: TechCrunch, Animation Magazine
cerspense fine-tunes ModelScope on 9,923 clips without watermarks, first square, then at a wider 576 × 320. A second model, XL, takes those clips and redraws them at 1024 × 576. Around it, camenduru releases Potat 1 at 1024 × 576, and cerspense a moodier dark variant. For a summer, this is open-source video.
Sources: zeroscope_v2_576w, zeroscope_v2_XL, potat1. All four Zeroscope models and Potat 1 archived, with the training code.
Yuwei Guo and colleagues release a motion module that plugs into any Stable Diffusion 1.5 checkpoint, so every anime and photoreal fine-tune can suddenly move. With prompt travel, the prompt changes every few frames and the scene melts from one subject into the next. A second motion module follows in September.
Sources: paper, guoyww/AnimateDiff. Weights and code archived.
Hotshot-XL brings AnimateDiff-style motion to Stable Diffusion XL for one-second GIFs. Tencent's VideoCrafter1 opens a 512 × 320 text-to-video model.
Sources: Hotshot-XL, VideoCrafter. Weights and code archived.
Stability AI releases an image-to-video model trained on a large, filtered set of clips. Shots hold together for longer, and the drift starts to read as camera movement instead of melting.
Source: stable-video-diffusion-img2vid. Weights archived.
Pika, a small startup that had been running a text-to-video bot on Discord, opens a web app and raises 55 million dollars. Generating a clip becomes a consumer product, with buttons.
Sources: VentureBeat, Tom's Guide
Alibaba's I2VGen-XL and Tencent's VideoCrafter2 arrive with sharper frames and steadier subjects, and AnimateDiff gets a third motion module. The melt is now a flaw to be trained away.
Sources: i2vgen-xl, VideoCrafter2, AnimateDiff v3. Weights and code archived.
OpenAI previews Sora: minute-long clips with people who keep their faces, reflections that behave and cameras that move on purpose. The two-second dream stops being a limit of the technology. From here on it is a style.
Source: OpenAI
Will Smith posts his own spaghetti video, eating it as frantically as his AI double did a year earlier.
Source: Benzinga
ByteDance distils AnimateDiff so a clip takes four steps instead of twenty-five. The melt now takes seconds to make, just as the frontier leaves it behind.
Sources: paper, ByteDance/AnimateDiff-Lightning. Weights archived.
The Toronto trio shy kids make a short about a man with a yellow balloon for a head, from Sora clips. It looks like film. They later explain the work behind it: hundreds of generations for each usable shot, then recolouring and cleanup by hand.
Sources: Notebookcheck, IMDb
Within two weeks, China's Kuaishou opens Kling, Luma releases Dream Machine to anyone with a browser, and Runway announces Gen-3 Alpha. Five-second clips with steady faces become something anyone can make. At Cannes Lions, Toys "R" Us shows a brand film made mostly with Sora.
Sources: Kling AI, YourStory on Dream Machine, CNN on Toys "R" Us
Runway removes runwayml/stable-diffusion-v1-5 from Hugging Face. Scripts across the ecosystem fail with 404s until they are pointed at mirrors. The community copy most tools use now was created on 30 August 2024.
Sources: Comfy-Org archive, a broken script, the community mirror. Archived from that mirror.
Zhipu AI and Tsinghua open CogVideoX-5B, six seconds of 720 × 480 video that keeps its subjects. It is the first open model that doesn't look like 2023.
Source: THUDM/CogVideo
Genmo releases Mochi 1, ten billion parameters under the Apache license. A month later Lightricks releases LTX-Video, small enough to generate faster than the clip plays back.
Sources: genmo/mochi-1-preview, Lightricks/LTX-Video
Coca-Cola remakes its 1995 Christmas truck advert with generative video. Viewers call it uncanny and soulless; Coca-Cola calls it a collaboration between people and AI, and makes another one the next year.
Sources: Adweek, Campaign Asia
Tencent releases HunyuanVideo, at thirteen billion parameters the largest open video model so far. Six days later Sora opens to paying ChatGPT users, and a week after that Google answers with Veo 2.
Sources: tencent/HunyuanVideo, Sora, TechCrunch on Veo 2
A shark in Nike trainers, Tralalero Tralala, is posted to TikTok with a rhyming Italian text-to-speech voice. Within weeks there is a cast of generated creatures, a bomber crocodile and a cappuccino ballerina among them, and children everywhere know their names. Slop has stopped being a look and become a genre.
Sources: Wikipedia, Know Your Meme. The first video, by TikTok user @eZburger401, is gone: the account was banned in early 2025.
Alibaba opens Wan 2.1, a fourteen billion parameter model and a 1.3 billion one that runs on a gaming card, and it goes to the top of the public benchmarks. The best open video model now comes from the company whose lab made ModelScope two years earlier.
Sources: Wan-AI/Wan2.1-T2V-14B, PetaPixel
Google releases Veo 3, which generates sound with the picture: dialogue in sync, footsteps, room tone. Fake street interviews and news reports with talking people flood every feed within days, and so does a new Will Smith eating spaghetti.
Sources: Gigazine, Android Authority
Alibaba splits the model into two experts, one for the rough layout and one for detail, still with open weights.
Source: Wan-AI/Wan2.2-T2V-A14B
OpenAI releases Sora 2 with sound, inside an app that looks like TikTok and contains only generated clips. Users scan their own face as a cameo and let friends put them in scenes. The app tops the App Store, and the feed fills with dead celebrities and trademarked characters.
Sources: TechCrunch, NBC News
Merriam-Webster picks slop and defines it as digital content of low quality that is produced usually in quantity by means of artificial intelligence. The programmer Simon Willison had argued for the word in May 2024, as the spam of generated content. It now names everything, not only the melting kind this site keeps.
Sources: AP via ABC7, Simon Willison
Lightricks publishes the full weights and code of LTX-2, which makes picture and sound together. What Veo 3 did in May now runs on a workstation.
Sources: Lightricks/LTX-2, Wikipedia
A day after ByteDance releases Seedance 2.0, a filmmaker posts fifteen seconds of Tom Cruise and Brad Pitt fighting on a rooftop, made from a two-line prompt. The Motion Picture Association demands ByteDance stop. A Deadpool screenwriter writes that it is likely over for us.
Sources: The Hollywood Reporter, Film Stories
Less than six months after its app launched, OpenAI discontinues Sora. The app and website close on 26 April and the API in September, and users' generations are deleted after a final export window. Hosted models disappear with their services; only open weights can be kept.
Sources: OpenAI Help Center, The Hill
The open models in this timeline are copied to storage with their code, every file checked against the hash Hugging Face published, and run once more on the same six prompts. The closed ones can only be written about.
Same prompt, same seed, each model at its own resolution and settings. Nothing is upscaled or cleaned up, except by Zeroscope XL, whose job is upscaling. Select a clip to see it at its real pixel size.
Each model is stored at the exact revision below, every file checked against Hugging Face's published hash and listed in a checksum file for later checks. Each code repository is stored as a git bundle with its full history.
| Repository | Revision | Size | License |
|---|
| Repository | Commit | Size |
|---|
Every clip was generated in November 2026 on an NVIDIA RTX PRO 6000 with diffusers 0.42.1 and PyTorch 2.14, from the archived weights. Each model runs at the resolution, frame count and step count from its own model card, with seed 2023 throughout. Zeroscope XL redraws the 576w clips at strength 0.6, as its card describes. AnimateDiff runs on plain Stable Diffusion 1.5 with the DDIM settings from its card. The long morph near the start is AnimateDiff v2 with prompt travel: the prompt changes every 16 frames, through all six.
The prompts are deliberately the ones the era kept trying. Where the famous originals used real people, these use anonymous ones.
Assembled by cobanov, 2026. The clips are new; the models, and their licenses, belong to their makers.