In mid-December, many small channels started using Midjourney and Stable Diffusion to create slideshows of fantasy images. Umbrahammer2’s now-unlisted “Dark Souls as an 80’s Dark Fantasy Film” is widely credited with starting the trend.
They were clearly inspired by earlier imgur galleries (as spotted by KYM), but adding a dreamy ambient track and a video format expanded the audience. The images used were made from Midjourney prompts like “DVD screengrab from Excalibur 1981” paired with an image from the relevant IP (Zelda, Star Wars, etc.) or a detailed scene description.
The trend expanded to ‘80s sci-fi, “live-action thing as anime,” “animated thing as live-action,” etc. The YouTube sidebar quickly filled up with suggestions for an AI slideshow binge. People recreated things that already existed, like a Soviet Lord of the Rings. Some of the biggest hits appeared on already-established channels about ancient Egypt, animation, and of course Dark Souls. A few channels used AI slideshows to promote their own music, which makes sense. There were also a flood of brand new channels with names like Deus Vult AI or AI Knight pumping out the same stuff plus tutorials.
For a minute the platform was overrun with huge wax faces, ambient music, and surreal organic armor. Plenty of commenters were dazzled:
Human preferences weren’t driving the trend, though. Creators were lining up things Midjourney could do pretty well (screenshots of fantasy/SF movies with elaborate costumes) with things the YouTube algorithm already likes (big game and movie IPs). If they persuaded the two robots to shake hands, they could collect a million views on weird disposable novelty content.
But that wasn’t true for long. That Egypt channel got 1.5M views for live-action Adventure Time in January but only 19K for live-action Urusei Yatsura the next month. A few channels even claim to have been demonetized (maybe for programmatically generated content?). The creator of that trendsetting Dark Souls montage delisted it and went back to Elden Ring PVP.
The lull ended in mid-March, when this video went viral and prompted a swarm of copycats:
The video’s creator, demonflyingfox, used three different AI tools to create it: Midjourney, ElevenLabs speech, and D-ID animation. (See this Vox how-to.) Again the concept was not something any human was asking for but an alignment of Midjourney’s well-known strengths (fashion photography/fantasy costumes) with subjects that have a lot of search juice.
As a human I have to mention that I dislike this video and its endlessly tilting heads. It feels like a “high fashion Harry Potter” parody that lost its way and became an ad instead of arriving at a punchline. But demonflyingfox succeeded in branding AI content in a way other channels haven’t. Only Excalibur’s John Boorman and the fathomless mind of Midjourney V4 own the dark ‘80s fantasy, but demonflyingfox clearly owns the Balenciaga meme. Their watermark is still on the rips of the Harry Potter video that spread on Discord and Tiktok. Their Balenciaga follow-ups did bigger numbers than the many imitators. Their channel now has 117K subs and more than 300 Patreon supporters, which looks to me like the greatest success story for any AI slideshow creator.
The upshot is that it’s totally possible to build an audience with this kind of protean AI content, even though it doesn’t resemble most other stuff on YouTube, it all relies on existing IP, and a lot of channels in the first gold rush couldn’t make it work. Demonflyingfox just needs to keep locating new concepts that exploit the brain-tingling/scraping sensation of viewing AI artwork, which they now have a financial incentive to do.
Pivoting to video
Slideshows are an odd format to choose for YouTube. The only non-AI slideshows I’ve seen recently were music-driven mood boards and playlists made to evoke earlier eras of the internet. And yet Midjourney slideshows amassed millions of views and countless dumbstruck reactions from a crowd that gathered just to leaf through a folder of concept art.
Back in January, YouTubers were making slideshows because AI video wasn’t ready for prime time. But now Runway Gen-2 is out, and it’s possible (if not exactly easy) to string together a few minutes of full AI video. Two interesting things about the first wave of these videos: 1) they don’t rely on name-dropping existing IP like all the big slideshows did and 2) they don’t reach the view counts that peak ‘80s fantasy/Balenciaga hit.
Pizza Later’s “Pepperoni Hug Spot” and the agency-created “Synthetic Summer” ad are funnier than most AI stuff because they’re traditional comedy sketches: they recreate a familiar hokey scene and then disrupt it by adding something violently “wrong” (body-warping AI chaos vision).
They also don’t look like something that Midjourney spits out after you type “funny ad –ar 16:9”. They look like work done by people with video production experience. Another recent AI hit, “Star Wars by Wes Anderson,” is a similarly polished product that appears to use heavily retouched and curated Midjourney images; it’s being used to sell “AI filmmaking” lessons.
Runway joke ads haven’t blown up on the scale of the dark 80s or Balenciaga slideshow booms. They’re not just harder to make — they’re also harder to watch. Slideshows gave you snapshots of a fictional world and let your imagination fill in the blanks. But when AI tries to fill the gaps itself to create real video, the result is fuzzy and gnarly and hard to wrestle into coherence. The traffic for slideshows came from viewers feverishly excited about AI expanding their favorite franchises; fewer people are amped about the Tim & Eric-core state of AI filmmaking right now.
That’s not to say there’s no audience. The channel AI Lost Media is the only one I’ve seen that survived each of 2023’s different AI content metas and actually gained ground with video. (Many channels that had a huge slideshow, like Lyrical Realms’ 5.1M-view Family Guy upload, have been dormant for a month or more.) Now the creator seems to be leaning into hell-dimension AI sicko content, uploading surreal ads with ChatGPT-written voiceovers and lyrics that he sings himself:
The creative effort of editing and adding sound, music, and graphics to AI video gives the final result more human personality than slideshows had. (I’m not sure when we’ll see a full-length “AI movie,” but I bet it’ll be a collaboration between enough people to fill out the credits on an indie film.) I’d guess that means there’s a lower viewership ceiling than there was for slideshows (smaller “the future is here!!” effect), but more potential for audiences to get attached to particular creators.
You can see hints of that in AI Lost Media’s comments:
What is this thing?
Every day on YouTube, thousands of channels play new and baffling video games. These are a few of them.
An incredibly long line of people has formed to kick the developer Arkane in the balls after the release of their vampire looter-shooter Redfall (May 2023) – a game that has forever disgraced its studio, the Xbox Brand, and the profession of vampire hunting itself. Redfall’s broken state and impoverished imagination have been widely contrasted with Arkane classics like Dishonored and Prey, which were prime video essay fodder on YT for years. Something went terribly wrong, and everyone’s looking at the dev’s new (since 2021) corporate parent Microsoft.
The procedurally generated detective sim Shadows of Doubt (April 2023) promises players infinite random murders to solve. This makes for a lot of odd situations and powerful video titles, so it was all over YouTube this week.
When someone tells you to play a game without reading more about it, the game is always a creepypasta. That said, Myhouse (March 2023) is a really ambitious Doom II map that you should play without investigating further. More YTers are discovering and explaining it each week.
World S (March 2023) is a huge total conversion mod for the original Resident Evil 4 that feels nightmarish outside of its hilarious cutscenes. RE4 Remake has dominated gaming discussions for the last few weeks and got people talking about older projects like this.
Eating Simulator: Physics Food (Dec. 2022) is a mobile absurdist comedy game driven by visual gags. Within each MSPainted scene you pull and twist a sometimes-edible object through a digestive tract. It’s perfect for grossout thumbnails.
The Klout™ Zone
Everyone got mad at Adin Ross for abusing YouTube’s content ID system and bragging about it. This was resolved in creator Internet Anarchist’s favor, but it seems like channels’ only defense against false copy strikes is kicking up such a fuss that YouTube’s Twitter account takes notice.
Creators posted a bunch of takedowns of the non-copyright variety: Pinely made “How Sunny V2 Ruined Video Essays,” Maggie Mae Fish took on off-grid giant Ghost Town Living, CGY accused Funko-core essay factory Nerdstalgic of plagiarism (again), and Twitter user @geologisms pointed out that many urban planning channels are just remaking the same video.
Different YouTube channels continue to get seriously hacked. This happened to Linus Tech Tips in March. If it happens to you, it seems again that your only defense is to get the attention of YouTube’s Twitter account.
Other Notes
I didn’t talk about AI voice-only channels above; there’s an entire history of text-to-speech channels on YouTube that goes back years before the current generative AI explosion. Video game life coaching is new, though.
The track “Skybridge” on this Sierra adventure game album is really good. I would have guessed the record was made just to attract game OST listeners in the YT sidebar, but it was apparently on Bandcamp for two years before migrating.
After the leak of the new Zelda game last week, a lot of people got interested in emulation for some reason. YouTube hosts a ton of easy-to-follow guides on setting up emulators like Cemu and Yuzu. YouTube seems to be playing whack-a-mole with tutorials that supply too many forbidden links, but it’s not hard to find some that have been up for months.
Thanks for reading
That’s it for this week. See you at the Pepperoni Hug Spot.