LIVE WIRE Ubuntu vs Windows 11 for AI Image Generation in 2026: A Careful ComfyUI Comparison
[ SYSTEM DISPATCH // V.4.2 ]
AI INTELLIGENCE & GENERATIVE SYSTEMS
[ FILED UNDER // AI CINEMA ]

How to Make Long AI Story Videos Locally: Unlimited Scenes on an RTX 3060

AIwithWalt October 18, 2026 11 min read

Making a three-minute AI cartoon on your own PC sounds impossible until you see it done. In his latest build, creator The_Frizzy1 shows a complete children’s story about Pip the fox — every image, every spoken line, even the song — generated locally in ComfyUI on a single RTX 3060 12GB, with nothing ever leaving the computer.

No daily limits. No tokens. No cloud subscription. Just one big workflow, one click, and a lot of patience.

The finished story, rendered end-to-end on one PC: Pip the fox and Clover overlook the frozen lake.

The finished story, rendered end-to-end on one PC: Pip the fox and Clover overlook the frozen lake.

That is the promise: actually unlimited long-form AI video, for free after hardware and electricity. It is not pretty, it is not fast, and it broke twice during development. But it works, and the approach is cleverer than just chaining clips together.

What These Long AI Videos Really Are

A long AI story video is not really one long video at all.

A 10-minute story is about 116 clips - the real work is gluing them together, not generating them.

A 10-minute story is about 116 clips – the real work is gluing them together, not generating them.

It is a list of short clips, each about five seconds long. A 10-minute story is more than 100 of them. In the cloud, every one of those means a prompt, a picture, a clip render, and a download.

The creator’s point is blunt: the real work isn’t the AI, it’s the gluing. Anyone can generate five seconds. Making 100 five-second clips feel like the same story, with the same character, the same voices, continuous music, and clean joins, is the hard part.

His workflow is built around one idea to solve that:

Make one picture of your character, and make every scene an edit of that same picture.

That single reference image goes straight into image-to-video, one line at a time, in a loop. At the end, all the finished clips are joined back into one story. Every intermediate scene is also saved as soon as it finishes, so even if you stop halfway, you keep everything that is done.

In other words, consistency is not left to chance or to re-prompting. It is enforced structurally.

The Script Format: One Line Per Scene

Everything starts with the script, and the format is strict.

WORLD, camera, dialogue, narrator

The story is written as a scene list, one line per scene. The first line can be a WORLD: line that sets the look of the whole story — where it happens, art style, how everyone looks. Every scene inherits it.

Each scene line then follows the same pattern:

  • [camera] in brackets, like [wide shot] or [close up], followed by what we see
  • Then || — two straight lines as a separator
  • Then what we hear

If what we hear starts with NAME:, like Clover:, the character says it on screen. Anything else is treated as narrator.

There is one more command: [continue] in brackets turns the cut into a transition into the next scene, instead of a hard cut.

The story is a script: a WORLD line sets the look, then one line per scene with [camera], dialogue and narrator parts.

The story is a script: a WORLD line sets the look, then one line per scene with [camera], dialogue and narrator parts.

Written by ChatGPT or Claude, not the built-in writer

Honestly, the creator recommends letting a big chatbot like ChatGPT or Claude write the script, using the 58-line structure prompt available on his GitHub.

There is a small 8B writer still inside the workflow, but it is switched off by default. The reason is simple: a detailed script written by a frontier model just looks way better than what the tiny local writer produces.

You load the workflow, paste your scene list, and go. The script is the directing.

One Character, Every Scene

Character consistency is where most long AI videos fall apart. This build attacks it with image editing, not re-generation.

First, Qwen Image 2.1 Turbo makes your character once — from a one-sentence description, or you can load your own picture.

Then every scene still is made as an edit of that same character image, at 8 steps each. Not a new character from text, but the same starting image warped into a new pose, shot, and background.

For the three-minute demo story, that was 36 pictures in 14m48s on the 12GB RTX 3060.

Only the main character gets this treatment. Side characters do not have a locked reference picture, so they can change from scene to scene. The workflow tries to compensate: it first writes a short story world describing where it all happens and how everyone looks, and every scene gets that description. You can also describe side characters directly in your lines.

But the guarantee only applies to the lead. Pip stays Pip. Clover, the seals, the bear, the hedgehogs — they can drift.

Stills Become Speaking Clips With MiniMax H3

Once the stills exist, they go straight into video.

The video model is MiniMax H3, specifically the fast 8-step turbo build. It is image-to-video, so each Qwen still becomes a short animated shot.

Every scene is as long as its line needs: 3 to 9 seconds, rendered at 0.4 megapixels. Crucially, H3 also generates the speech. It does not just move lips — it actually speaks the line.

In the three-minute story with 36 lines, a Whisper check heard 34 of the 36 lines word for word. Two lines were misread, which the creator reports honestly.

A scene from the finished film: MiniMax H3 turns each locked character still into a speaking clip.

A scene from the finished film: MiniMax H3 turns each locked character still into a speaking clip.

That is a remarkably high hit rate for local lip-synced dialogue, and it explains why scene length is variable. The clip stretches to fit the spoken sentence, rather than cutting the sentence to fit the clip.

Music by YuE2, Looped to Stay in Time

Dialogue alone feels empty. Music glues it together.

The music model is YuE2. You give it a music style and some lyrics, it makes the song once, and then the workflow loops it under the whole story.

Importantly, it does not just lay one long track underneath. It cuts the looped music scene by scene so it stays in time with the variable-length clips. That prevents slow drift where audio and video fall out of sync over minutes.

And again: every scene gets saved as soon as it is done. If the run crashes at scene 80 of 100, you still have 79 finished scenes on disk.

The Narrator Fix: Who Actually Speaks?

Watching the second test story with sound on revealed a real bug: characters would speak as if they were the narrator. Every line came from whoever was on screen, even pure narration.

The fix splits dialogue into two paths.

Now the workflow knows who is talking:

  • A line with NAME: stays in the scene, and that character’s mouth moves.
  • A narrator line tells the scene that nobody talks — characters only laugh or gasp, they do not mouth the words.

The narrator voice is then made in its own tiny H3 pass at just 256×160, and only its audio is laid on top of the scene video. That extra pass took 31-36 seconds per narrator line.

In tests, Whisper heard all 6 out of 6 narrator lines word for word.

A second narrator bug was also fixed: before, every narrator line had a different voice. Now you can choose one narrator voice, and it stays the same throughout the video.

There is also a new switch that upscales the final video to 1080p. On shorter videos it just worked. On longer videos, you need more than 32GB of system RAM for the upscale to succeed.

Scenes vs One Shot Mode

There are two ways to assemble the story.

Scenes mode: every scene starts from its own Qwen still, with a cut in between. Maximum consistency, but visible cuts.

One Shot mode: every scene carries on from the last 22 frames of the one before. Smoother, more continuous motion, like one unbroken take.

For a one-minute test, Scenes mode took 43m38s and One Shot took 45m20s — very close in speed.

The difference is drift. In One Shot, only the first scene has a Qwen picture. Everything after that is video continuing video. The look can drift more over time. The creator leaves the judgment to the viewer which tradeoff you prefer.

Real Numbers on an RTX 3060 12GB

This is slow. Here are the real times, measured on the RTX 3060:

  • 1 minute of story: 43 minutes
  • 2 minutes: 1h23
  • 3 minutes: 2h05

It scales in a straight line, about 210 seconds of compute per scene. In the three-minute story, the stills took 14 minutes and the clips took 1h50.

VRAM stayed just under the card’s 12GB. System RAM peaked between 17 and 22GB. The 1080p upscale switch is the outlier — on longer videos it needs more than 32GB RAM.

The creator really tried for 10 minutes like the big channels. He tried an overnight run, could not get it to work, and settled for 5 minutes instead, which in his words is already a long, long time.

This is start-it-and-forget-it software. Two hours for three minutes means one click, then go do something else.

What You Need to Run It

What you need: ComfyUI 0.39+, nine model files, about 63.5 GB - and zero custom nodes.

What you need: ComfyUI 0.39+, nine model files, about 63.5 GB – and zero custom nodes.

The stack is deliberately simple:

  • ComfyUI 0.39 or newer, because loop nodes are now core to ComfyUI. There are zero custom nodes required.
  • 9 model files: 3 for pictures, 5 for video, and 1 for music — about 63.5 GB together. Every file with its correct folder is listed in the downloads file next to the workflow.
  • The whole thing is one workflow on GitHub, link in the video description. You paste your scene list, describe your character in one sentence or load your own picture, pick Scenes or One Shot, music on or off, and press run.

Licenses matter here. Qwen Image and YuE2 are non-commercial, and MiniMax H3 has its own community license. If you want to make money with these videos, read them first.

That is why the creator says he is considering an LTX 2.5 version — one you could actually monetize. He says he will check its license properly first, and asks viewers to comment if they want that.

Honest Limitations

To his credit, nothing is hidden:

  • Side characters change from scene to scene.
  • One Shot mode drifts further because only the first scene has a locked still.
  • Now and then H3 misreads a line — 2 out of 36 in the three-minute story.
  • Loudness jumps from scene to scene, because H3 makes the sound fresh for every scene.
  • It is slow: two hours for three minutes on a 12GB card.

And RAM for 1080p upscale on long videos is a real wall if you only have 16 or 32GB.

The finale under the northern lights - every frame, voice and note of it generated locally.

The finale under the northern lights – every frame, voice and note of it generated locally.

Still, the finished Pip story plays as a coherent little film — compass, frozen lake, seals, storm, bear, hedgehog village, mountain top, northern lights — all from one local machine.

My considerations

What changes here is not quality, it is economics. Cloud long-form video charges you per second, which makes a 10-minute cartoon with 100+ clips punishingly expensive to iterate on. A local loop like this flips the cost to zero marginal cost: once the 63.5 GB of weights are downloaded, scene 101 costs the same as scene 1, which is electricity and time. That invites experimentation that credits never allow.

The electricity-vs-subscription trade is real, though. Two hours of a 3060 at full load for three minutes of video is not free, and an overnight 5-minute render is a space heater. For one film a month, a $20 cloud plan may still be cheaper than the power bill joke at the end of the video. Where local wins is volume, privacy, and control — unlimited retries, no content filter roulette, no queue, no files leaving the PC.

The image-edit trick is the real breakthrough, more than the video model itself. Re-prompting a character 36 times guarantees drift; editing the same Qwen Image 2.1 Turbo portrait 36 times anchors identity. It is a simple, almost boring solution, and that is why it works. The split narrator pass is the same kind of practical thinking: stop asking one model to do acting and narration at once.

Who should use this? Patient tinkerers with a 12GB card, 20GB+ of system RAM, and a taste for overnight renders — educators, hobby animators, parents making a bedtime story starring their kid’s drawing. Who should not? Anyone on deadline, anyone without VRAM headroom, anyone expecting Pixar lip-sync or stable side casts, and anyone planning to monetize without lawyering the non-commercial and community licenses first.

Where this points is clear: AI filmmaking is moving from prompting clips to directing systems. Script format, persistent cast image, per-scene audio, looped score, automatic assembly — that is a tiny studio pipeline, not a prompt box. When faster distilled video models and properly commercial weights drop in, this exact loop will feel obvious. Right now on a 3060 it takes all night. Soon it won’t.

Source

Video title: How to Make LONG AI Videos for FREE (Actually Unlimited, No Credits)
Channel: The_Frizzy1
Published: 2026-10-18
URL: https://www.youtube.com/watch?v=dSzM3ovKBko

Leave a Reply

Your email address will not be published. Required fields are marked *

>_ JOIN THE INTELLIGENCE NETWORK

One email when a new breakdown drops: the workflow, the settings, and what actually worked. No spam, unsubscribe anytime.

Join the Intelligence Network

New AI video & image breakdowns, straight to your inbox. Free, no spam.