Pilot 01
A shape with one side
Walk a strip through a half-twist to reveal the Mobius strip. Immediate visual payoff, accessible to students, and easy to turn into a classroom pause point.
Creative systems · 19 August 2026
Make the strange understandable.
A practical plan for educational mathematics, deliberately degraded fictional footage, and 3D story assets on a 12 GB GPU.
The short answer
The Manim idea is strong. It has a clearer audience, a reusable production system, and a better educational moat than generic AI footage. Do not fork Manim. Keep the upstream project intact and put scenes, scripts, visual language, and publishing workflow in a separate project.
Run the math series alongside AtlasLab. Two AtlasLab episodes are already finished, so replacing them would throw away completed work. Make AtlasLab the umbrella and launch a math season inside it. The washing-machine episode can remain an earlier experiment instead of defining the whole brand.
Recommendation: ship one math pilot and one archive-style pilot before creating a new brand. The math format is the stronger long-term bet. The archive format is the faster format test.
A. Manim educational maths
Manim is a programmatic animation engine for explanatory mathematics. The advantage is not just attractive motion. It is the ability to show the exact reasoning, pause on a definition, and reuse a visual grammar across lessons.
Pilot 02
How circles draw waves
Use rotating vectors to build a Fourier series, then show why more terms sharpen a square wave. This is the strongest first spectacle.
Pilot 03
The shortest path
Introduce geodesics on curved surfaces and connect the idea to maps, travel, and physics without turning the lesson into a formula dump.
Episode shape
Target 60 to 90 seconds for the first pilot, with a longer classroom cut later. Give every episode one claim, one visual mechanism, one worked example, and one teacher-ready recap. Keep a reusable scene library for title cards, axes, vectors, labels, theorem cards, worked examples, and recaps.
What it costs
The current creative machine has an RTX 3060 with 12 GB of VRAM and Manim Community 0.19.1 installed. A production budget for a 60 to 90 second, vector-heavy episode is 15 to 30 minutes of render time at 1080p, plus 2 to 5 minutes for a low-quality preview. A simple 720p pilot should fit in roughly 5 to 15 minutes. A 3D-heavy OpenGL episode can exceed that range.
Measured pilot
The Mobius strip pilot rendered as a 10-second, 1280x720, 30 fps MP4 in 24.594 seconds, with exit code 0. The render used Manim's OpenGL renderer and the existing FFmpeg toolchain. GPU samples ranged from 0% to 6% utilization with about 1.791 GB of VRAM in use. This simple scene is CPU-bound, so the 15 to 30 minute episode budget remains a conservative planning range for a longer, denser episode rather than a direct multiplication of this one test.
Do not start with 4K. Render a 15-second proof first, then the 60-second cut.
Manim supports Cairo and OpenGL renderers, and its documented quality levels make a preview-first workflow practical. Read the upstream project and its renderer and quality guide.
B. Archive-style AI content
This idea is worth testing because it turns generation errors into a consistent visual language. The rule is that the footage must be clearly fictional. Never frame an invented clip as real evidence, a real emergency, or a real person.
Pipeline for a 12 GB card
- Write a 30 to 60 second micro-story with one mystery and one visual rule.
- Generate a small set of 4:3 still keyframes at 512p or 768p. Keep the camera simple and reuse the same reference image.
- Animate only short shots. Start with Wan2.1 T2V-1.3B at 480p, or use AnimateDiff when a still image and controlled motion matter more than free camera movement.
- Edit on the CPU: hard cuts, dropped frames, exposure jumps, timecode, scan lines, tape wobble, muffled audio, and restrained compression noise.
- Add a clear fictional title card or description. The degradation should support the story, not hide what the viewer is seeing.
Wan2.1's official repository lists 8.19 GB of VRAM for its 1.3B text-to-video model and recommends 480p for stability. On a 12 GB card, this is the realistic starting point. Plan on 5 to 15 minutes per five-second shot until a local benchmark says otherwise. Six to ten shots makes a first 45-second piece a one to three hour GPU budget, with editing time separate. These are planning ranges, not a measured local render.
Wan2.1 is the first candidate. AnimateDiff is the controlled-motion alternative.
First three pieces
Piece 01
Trail Cam 04
A fixed camera, a timestamp jump, and one impossible change in the background.
Piece 02
Channel 9 Emergency Test
A low-resolution public-access interruption that slowly reveals the message is addressed to the camera operator.
Piece 03
Tape 07: The Stairwell
A handheld descent where the floor count changes between cuts.
C. The 3D-object repo
The exact repository called “modly” is unverified. It was not among the 25 starred repositories visible on the ZionBoggan account, and the old account may contain it. Send the repository link before evaluating its license, checkpoints, or installation path. Do not guess the name.
Link needed: the repo name alone is not enough to select a safe model or promise that it fits the 12 GB card.
| Option | What it actually does | 12 GB verdict |
|---|---|---|
| Shap-E | Official text-to-3D and image-to-3D research release. Useful for fast concept meshes and intentionally strange props. | Best immediate text-to-3D experiment. Expect rough geometry and cleanup. |
| Hunyuan3D-2 / 2.1 | Strong image-to-shape pipeline with a separate texture stage. | Shape generation is listed at 6 GB; the full shape-plus-texture path is listed at 16 GB. Use a generated concept image, shape first, and an optimized or decoupled path on 12 GB. |
| Stable Fast 3D | Image-to-GLB reconstruction with UVs and optional remeshing. | Practical at about 6 GB for one image. Use it after a concept-image step, not as direct text-to-3D. |
| TripoSR | Fast single-image-to-3D reconstruction. | Useful for a controlled reference sheet. It is not the text-prompt solution by itself. |
| TRELLIS | Higher-end image-to-3D research pipeline. | Treat as a larger-GPU option, not the first 3060 target. |
The recommended story workflow is text prompt to concept sheet, concept sheet to mesh, mesh cleanup in Blender, then a small set of consistent renders. That keeps silhouette and camera under control even when the generated mesh is imperfect.
Recommendation
Start with one Mobius-strip or Fourier pilot and one 45-second fictional archive piece. Run them as two formats under AtlasLab rather than creating another brand. Keep the 3D repo evaluation blocked until the exact link arrives.