Home
Creators
Search Recent Random
Posts
Search DMs Popular Hash Lookup Tags Random
Importer
Import FAQ
Register Favorites Keys Review DMs Logout
Community
Telegram ThePornDude
Home Creators Posts Import Register Favorites Logout

MiniMax H3 and Krea 2 - I Created This AI App for Idea-To-Video - Calliope (Patreon)

Published:
2026-08-19 13:00:17
Edited:
2026-08-20 23:54:52
Imported:
2026-09

Content

https://youtu.be/szJF38GzvWk

Hey everyone.

The video is up. I built Calliope, an open-source, local-first story studio that runs your ComfyUI as the engine, and tonight's walkthrough shows the whole pipeline end to end.

You start with a story idea. An LLM drafts the beats into a storyline, you edit or accept, and then the app walks the pipeline the way a studio would. Characters and environments become reference images through Krea 2. Each scene becomes a shot through MiniMax H3 reference-to-video, with the app formatting your prompts into the H3 prompt structure so you can paste the script straight in. Scene durations, the timeline, and the final stitch of every clip are handled on export. The LLM sits in an agentic loop the whole way, and nothing renders until you approve it.

The part I keep coming back to: the ComfyUI workflow is just a tool call now. You design them once, import the API-format JSONs, and the app auto-detects their inputs. Any OpenAI-compatible endpoint works for the language model.

It's free and open source, with example workflows included:

github.com/benjiyaya/Calliope

For paid supporters — the demo story pack is up in a separate paid post: the complete 10-scene MiniMax H3 Ref2VA prompt set from the demo story (every scene with character speechlines and reference-slot wiring), the four character sheet images, and the demo video files, plus the gotchas I hit while building the demo project.

Thanks for supporting the channel.

— Benji

P.S. If you build something with it this week, drop a link in the comments. I want to see what stories you come up with.

Files

MiniMax H3 and Krea 2 - I Created This AI App for Idea-To-Video

Calliope, a local-first story-to-video studio app he built on top of ComfyUI. The video covers the full pipeline: creating a project from a story idea, AI-drafting the storyline beats, generating character and environment images (including character sheets), AI-writing the per-scene script, then generating each scene's video clip with MiniMax H3 reference-to-video workflows — and finally stitching everything into one exported video. It also covers the ComfyUI API-format workflow import system (role-tag based auto-detection), the MiniMax H3 prompt auto-formatting, and the Omni-style Playground for freeform image/video testing. AI video creators who are tired of juggling ten browser tabs between a story doc, ComfyUI, and a file explorer. Also developers interested in building their own AI apps on top of ComfyUI's API, and anyone following the Krea 2 / MiniMax H3 ecosystem who wants to turn story ideas into finished videos without leaving one interface. Most AI video tools are single-shot generators. Real story content needs a pipeline: story, characters, environments, script, shots, timing, export. Calliope connects those stages with language models in an agentic loop while keeping a human in the approval loop — and it runs entirely on your own machine with your own ComfyUI install and any OpenAI-compatible LLM endpoint. No cloud subscription, no vendor lock-in, and the ComfyUI workflow becomes a tool call inside a larger system. Calliope https://github.com/benjiyaya/Calliope Local Workstation GPU : https://amzn.to/3XfXsAO -------------------------------------------------------------------------------------------------------------------------------- If You Like tutorial like this, You Can Support Our Work In Patreon: https://www.patreon.com/c/aifuturetech Timeline: 00:00 - Introduction: Benji introduces the new app built on ComfyUI, focusing on Minimax H3 for video and Krea 2 for images. 00:24 - App Concept & "Studio" Philosophy: Explanation of the local-first, story-driven studio design and the "agentic loop" with human approval. 01:00 - UI Rebuild & Backend Setup: Why the app was rewritten for a simpler artist-friendly interface and how to connect OpenAI-compatible APIs. 01:32 - Importing ComfyUI Workflows: Demonstrating how to drag and drop JSON API files and how the app auto-detects parameters. 01:58 - Project Creation & Story Drafting: Starting a new project (Sci-Fi movie example) and using AI as a co-pilot to generate story beats. 03:13 - Asset Creation (Characters & Environments): Reviewing and generating consistent character sheets and environment images from text prompts. 05:36 - Script Generation: Using the AI to transform story beats into detailed scenes with durations and character assignments. 06:38 - Video Generation Timeline: The core "fun part" showing how scenes are pre-aligned and how to run Minimax H3 reference-to-video (R2V) workflows. 07:45 - Prompt Formatting & Automation: How the app handles specific Minimax H3 prompt formatting automatically. 09:32 - Advanced Workflow Logic: Discussing reference requirements and deep-diving into the JSON detection for input/output titles. 10:48 - Final Video Stitched Result: A demonstration of the completed AI-generated story clips stitched together. 12:15 - Customization & Camera Movement: Tips on emphasizing camera movement in prompts for better Minimax H3 results. 13:24 - The Playground & Grid View: Testing individual prompts/images in the playground and using the grid view to manage generations. 14:45 - Cinematic Showcase: Final full-screen reveal of the AI-generated Sci-Fi narrative. #comfyui #minimaxaivideo #minimaxh3 #aiagents

Comments

Lee

Been waiting in anticipation for this post. Can't wait to try this out, thank you for your hard work Benji!

R D

awesome stuff. It appears (i could be incorrect) that the krea2_T2I_CharSheetV1_API.json file is missing connections. I had AI fix the connections for me and now working great. I can generate the story and assets (characters and environments) but the generate script fails also. Love the idea.

Innovate Futures @ Benji

yes script I have change the code. The GitHub current version was very basic loop for text gen. It will over load number of characters in message for return if running a long length storyline. I switch to agent loop in the new one, will do some more in depth integration and push the new version to Git soon.

Andreas Vogler

Great, work. I have had some issue, so i used ki to korrect it i have put it online for you maybe you could use somthing of it https://github.com/tyl0re/Calliope/tree/feat/portable-krea-workflow-controls 1.My problem was that there was no negative Prompt for the image heneration 2. The Video Prompt can't be changed afterward, for correcting errors 3. I wanted to use an Uncensored (Lora) for the Image generation, since sometime i had the feeling he ignores prompt text 4. Refresh of promts in Storys hast worked for me And also some smaller changes.

Innovate Futures @ Benji

thanks, but the upcoming version, I will have a big change on the inferencing part. Keep this version on your repo at this moment, I will check later ver 1.4 release and see if any features from your repo can merge. thanks

Frank Hunley

Nice. It has an issue with overwriting files and indexing confusion when you make changes. Not sure how you are storing the data for the runs, but it definitely holds some pathing and naming that completely replaces output videos.