TMochiLearn Post-Submission Update
OpenAI Build Week
Hi everyone, just wanted to share a few updates that were made to backend and frontend of the TMochiLearn application- Create interactive ed-tech videos from simple text prompts. TMochiLearn now uses a focused model set, smoother path playback, and scene-based render pricing.
While all supported inference models work great for the singular text-to-video pipeline, interactive, branched videos is the next frontier. To ensure that Samsar pipeline works predictably and reliably in interactive media generation pipelines we constrained the inference, image and video models for this pipeline. See the updated model matrix and changes below-
Constrained Model Pipeline
Inference Models
| Model | Effort | Request value |
|---|---|---|
| GPT-5.6 Sol | High — default | gpt-5.6-sol + high |
| GPT-5.6 Sol | Extra High | gpt-5.6-sol + xhigh |
Both options use the same GPT-5.6 Sol model with different reasoning effort.
Image Models
| Model | Request value |
|---|---|
| GPT Image 2 | GPTIMAGE2 |
| Nano Banana Pro | NANOBANANAPRO |
Video Models
| Model | Request value | Production rate |
|---|---|---|
| Nvidia Cosmos 3 | COSMOS3SUPERI2V |
20 credits/second |
| Veo 3.1 | VEO3.1I2V |
60 credits/second |
| Veo 3.1 Fast | VEO3.1I2VFAST |
36 credits/second |
| Seedance 2.0 | SEEDANCE2.0I2V |
40 credits/second |
Available options depend on the connected deployment.
Runtime Model Validation
- Creator Studio loads model availability from the server.
- Only models marked compatible with branching are displayed.
- Every selection is validated again before generation.
- Generation is disabled when a required model stage is unavailable.
- Saved selections are restored only while they remain supported.
- Models are submitted explicitly without silent substitution.
Improved Path Playback
- The selected path is mounted and preloaded immediately, including on mobile.
- Only the chosen next-path video is buffered after a route is selected.
- Playback continues from the correct decision point.
- Volume, mute, and playback-speed settings remain unchanged across transitions.
- Path-map changes resume from the latest point where the routes diverge.
Scene-Based Render Pricing
Final billing is based on the distinct scenes generated, not the number of final branches.
render credits = total unique rendered scene-seconds × video-model rate
- Shared scenes, transitions, and outros are counted once.
- Each unique branch-specific scene is counted once.
- Narrative generation is included in the interactive-video rate.
- Only completed stages are charged.
- Duplicate stage charges are prevented.
If two 20-second paths share their first 10 seconds, TMochiLearn bills 30 unique scene-seconds instead of 40 path-seconds.
Creator Studio shows a conservative Estimated up to value before generation, followed by Charged so far and the final Credits charged.
Workflow Improvements
- Model and reasoning settings persist across resumable sessions.
- Stable request identifiers reduce duplicate submissions.
- Newly catalogued models remain compatible with older SDK versions.
- Real API errors are returned without automatic retries.
- Standalone deployments use provider billing without production credit redirects.
Result
TMochiLearn now delivers more consistent model selection, smoother interactive playback, and pricing based on the media actually generated.
Log in or sign up for Devpost to join the conversation.