The films in this post
In short. You can make 3D animation with AI on one PC without a 3D team, if you pick the right method for each shot. We used four: clay films from storyboard sheets animated by LTX-2.5, physics scenes in Blender rendered on one GPU, layout-to-render where Blender blocks the camera and LTX-2.5 renders it, and a Pixar-style podcast from stills. Each has a different thing you control and a different way of failing, and the films above show all of them.
What is 3D animation with AI, in our setup?
Three tools, each doing one job. Claude Code is the director: it writes and runs the scripts. Codex makes the storyboard images with ChatGPT Images. LTX-2.5, an open video model, animates them on our RTX 5070 Ti (16 GB). Blender, which is free, does real 3D. Claude drives Blender, and Unreal Engine, through their MCP connections: you describe the scene, the light and the camera, and Claude builds and renders it.
We are not 3D artists. We are two people who describe what we want and check what comes back.
How do you make a claymation film with AI?
The Last Light (a clay moth and a lighthouse, no words, 30 seconds), The Last Bus, The Pickup and Agnayya in clay all follow one recipe.
- Write the story and set the clock, with voice first if there is a voice.
- Ask Codex for one storyboard sheet per sequence, in a clay look: fingerprints in the clay, and one material that catches the light.
- Cut the sheet into panels so every shot shares the same cast and light.
- Animate each panel with LTX-2.5, in shots of about four to seven seconds.
- Over-generate, trim before the point where the clay starts to drift, add sound, and let a critic agent review the cut.
Clay hides a lot of AI faults, because clay is supposed to be a little uneven. It also punishes faces: a face that drifts between shots looks like a different puppet. The fix is a camera lock we call hold, and cutting the shot if hold does not work.
Can Blender with Claude do physics a clay film cannot?
Yes. 1 cut vs 25,000 cuts (53 seconds) counts cuts up at a fruit cart in Chennai. We liked the format in another reel and rebuilt it ourselves. Blender 5.2 simulated the physics, and Cycles with OptiX rendered about 1,800 frames on the RTX 5070 Ti. Two Codex plates form the stall. All the sound was synthesised in code: no music, no voice.
Here the video model is not involved. Everything you see is a real simulation, so the objects obey gravity, and you can re-render any frame. The cost is time: you build the scene and wait for frames.
What is layout-to-render?
It is for shots where the camera path matters. In Stashed: Out of Office (14 seconds), the camera path is blocked in Blender with grey boxes. LTX-2.5 then renders that layout as a photoreal shot.
You control the camera exactly, and the model decides what the surfaces look like. It fails when a prompt asks for text: logos and prices garble, so we add them in the edit, and use a real photograph of the product for the packshot.
What about a Pixar-style 3D podcast?
On 11 November we made a 22-second podcast of the two of us in a 3D look. The pieces: 3D-style stills from Codex, lip-sync from 2x2 mouth sheets drawn from the voice, and real website clips cut in. It is in the films library. It is the cheapest of the four methods to run, because no video model renders the faces. The trade-off: the mouths change with the voice, but the bodies do not move like a rigged 3D model.
Which method should you pick?
| Method | What you control | GPU time | Best for |
|---|---|---|---|
| Clay from storyboard sheets, LTX-2.5 | Look, cast and story through the sheet | LTX-2.5 benchmark on this PC: four named jobs averaged 61 s cold and 66 s warm; roughly three to four renders per kept shot | Short emotional films, characters |
| Blender physics, Cycles with OptiX | Everything, frame by frame | About 1,800 frames for the 53 s film; no per-frame number measured | Physics, counting, objects |
| Layout-to-render | Exact camera path | The Blender layout, then LTX-2.5 render time per shot | Product shots, moving cameras |
| 3D-style stills with lip-sync sheets | Poses and mouth shapes | Little or none | Talk and explainer formats |
Across Trishul episodes 3 to 5 and Kolar, an LTX-2.5 render averaged 63.8 seconds. Peak in the logs was 73 °C and 277 W on the GPU. We have no per-frame Blender figure, so we did not invent one.
What does it still cost, and where does it fail?
The PC costs about ₹1.5 to 1.6 lakh once. Monthly, you pay for a Claude plan, the ChatGPT plan that makes the images, and electricity. On the Plus plan we hit the image limit at 75 images in one usage window while making NETA. Read the current LTX-2.5 licence before you sell work: when we checked, the community licence allowed commercial use for companies under US$10 million a year.
The failures repeat. Faces drift, hands melt, objects morph, text garbles. Our fixes are boring and they work: hold, a plain instruction, cutting the shot, and a critic before anything ships. More in LTX-2.5 offline AI films on an RTX 5070 Ti and in our weekly log, what we made in eight days.
Learn it
The 2BFT Lab teaches this: ₹10,000 once for the live course (live every Wednesday and Sunday, 7 PM IST, with every recording) or ₹1,000 for the recordings alone. The AI films session is 7 December, the short-series session 14 December and 3D on 28 December. You need your own Claude plan (Pro is about ₹1,800 a month, paid to Anthropic, not included). See how this works on your own PC and pricing.
