← Back to blog
AiAbout 5 min read

A 67-Second Seinfeld Episode, Made at Home on a Chinese Video Model

Published Nov 15, 2026
A 67-Second Seinfeld Episode, Made at Home on a Chinese Video Model

A 67-second episode of a 1990s sitcom appeared on Reddit this month, starring two characters who look and sound like Jerry and George, set in a rebuilt apartment that never existed. It is labelled as AI in the title. Nobody in the top comments argued about whether it was generated. They argued about the joke.

The production note is the part worth reading. The creator, posting as MandelDuck, said he used MiniMax H3 inside a local ComfyUI setup, split the episode into shots of five to six seconds, pinned the first and last frame of each shot to hold the characters together, prompted the model for the likenesses, and baked the audio into the video rather than dubbing it afterwards.

That is the whole method. No fine-tune, no character LoRA, no studio pipeline. A node graph, a consumer GPU, and a lot of patience.

Consistency was always the wall

Anyone who has tried to make more than one shot of the same person knows where these projects die. Shot one looks right. Shot two has a slightly different nose, or the jacket changes shade, or the hairline drifts two centimetres. By shot four you have a stranger wearing the same clothes.

The usual fixes are expensive. Train a character LoRA, and you get consistency inside a narrow style but lose flexibility on lighting and angle. Feed a reference image every time, and the model treats it as a suggestion rather than an instruction. Generate long and cut around the drift, and you spend compute on frames you throw away.

Pinning the first and last frame is the cheap version of the same idea, and it works because of how these models are conditioned. Give the model both ends of a shot and the interpolation has somewhere to land. The character has to arrive at a specific face at a specific moment, so the drift has less room to accumulate. Chain those shots and the errors do not compound the way they do when each shot starts from a text prompt alone.

The cost is control at the seams. Cuts have to be designed around the pinned frames, so you end up writing to the technique rather than the other way around. A 67-second piece tolerates that. A seven-minute piece probably would not, at least not without a lot more shots.

Audio in the same pass changes the schedule

Baking the audio into the generation, instead of generating silent video and dubbing later, removes a whole stage. It also removes a whole class of mismatch: lip sync becomes the model's problem rather than the editor's.

The trade is that you cannot fix a line without re-rendering the shot. For a parody built on one joke, that is fine. For anything with a script that gets revised, it is a trap.

What it does do is collapse the distance between an idea and a finished clip. The thread's most-upvoted comment is about exactly that: memes of this quality now arrive almost in real time. A week ago the same joke would have needed a writer, two voice actors, an editor and a week.

The cost of scaling this up

Run the numbers on a season and the picture changes. Sixty-seven seconds took an afternoon of iteration and a collection of five-second shots. A twenty-two minute episode at the same shot length is roughly 260 shots, and the failure rate per shot does not fall with practice, because each new camera angle resets the consistency problem.

At an afternoon per fifteen shots, that is a couple of weeks of full-time work for one episode, before writing, before sound design beyond what the model bakes in, before anyone checks whether the jokes land. A season is a different kind of project rather than a bigger version of the same one, and the technique that made the parody possible is the thing that caps it.

That is where the current generation of tools sits: good enough that a single scene is a solved problem, and far from the point where a series is.

The audience did the quality control

The criticism in the thread is specific, which is a good sign. George's voice needs work while Jerry's is close. The hair is distracting. One commenter calls the accent a little cringe and still gives it eight out of ten. Two people argue about whether a particular line mocks trans women or mocks George's expectations.

That is what a normal television episode gets. Not a debate about whether the thing exists, a debate about whether it is any good. The bar has moved from "can a machine do this" to "is this funny", which is a much harder bar and a much less interesting one to demo.

Two smaller details are worth noting for anyone tracking cost. The thread includes a gag about the water and RAM the render consumed, which is the kind of joke people make when the compute bill is real but small enough to laugh at. And there is an AMD punchline, which suggests the creator was not running on the most obvious hardware.

What this does to the people who make sitcoms

Not much, in the short term. A 67-second parody is not a season. The writing is the hard part, and the model did not write the joke.

In the medium term, the interesting shift is that the technical floor for a convincing multi-shot scene has dropped to a single person with a mid-range card. That floor used to be a crew. Everything above the floor, taste, timing, casting, a reason to watch, is still human labour, and it is now the only thing separating the output.

The uncomfortable question the thread raises is not about likeness or copyright, both of which are live issues this parody happens to sidestep by being obviously a joke. It is about volume. When a decent fake episode takes an afternoon, the number of decent fake episodes goes up by orders of magnitude, and the scarce resource becomes attention rather than production capacity.

Sixty-seven seconds is a small thing. It is also the first time the recipe has been written down in public, in a Reddit comment, by someone with no particular interest in teaching it. That is usually the moment a technique stops being a demo and starts being a default.

Related articles

Google replaced the model behind Search, and the number that moved was a shell benchmark

Gemini 4 Hydrogen became the default on 12 November, taking over from Gemini 3.8 Flash across Search, the Gemini app and the developer API. It runs about 2.5 times faster, scores 57% on Terminal-Bench 4.1 where its predecessor scored 18.2%, and costs $1.50 in and $7.50 out per million tokens.

Three ways a coding agent wastes money, measured on 30 repositories

TokenEconomizer watches what an agent did in a repository you work in often, then stops it from rediscovering the same build command, loading tools it never uses and looping on one error. Across five agent and model combinations it cut mean cost by 18.4% to 30.9%, cheaper in all 30 repository cells, with Pass@1 up in four of five.

earshot gives coding agents 922 sounds, and every one is rendered from code

An agent can write a game, an interface or a film and never hear any of it. earshot answers requests like "three variations of a footstep on gravel" from a catalogue of 922 designs built from public-domain recordings, with no audio files in the repository at all. It runs on CPU, as a CLI, a Python package, an MCP server and a Claude Code plugin.

The agents stealing API keys were not sophisticated. The gateways were open.

Amazon's decoy servers spent October posing as inference APIs and logged agents built on ARTEX pulling provider keys out of LiteLLM, Dify, Langflow and Open WebUI. The way in was a default master key and a half-finished install. ARTEX itself ships with budgets, scope gates and evidence for every finding.