README.md

metadata

title: GreatCyber AI Video
emoji: 🤖
colorFrom: purple
colorTo: blue
sdk: gradio
sdk_version: 6.25.0
app_file: app.py
python_version: '3.12'
startup_duration_timeout: 1h
short_description: GreatCyber AI video generation with H3
models:
  - MiniMaxAI/MiniMax-H3
  - Beidouqixing/minimax-h3-4step-lora-flashgen

GreatCyber AI Video

Create cinematic videos with synchronized native audio using MiniMax-H3 and FlashGen 4-Step technology.

Features

• Text-to-video generation
• Image-to-video with an optional first frame
• Optional last-frame guidance
• Native synchronized audio
• FlashGen 4-step generation
• Optional prompt enhancement through the Qwen3-VL conditioner
• Multiple aspect ratios and supported resolutions
• Adjustable duration
• Seed control for reproducible generation
• Premium GreatCyber-branded interface

How the Space works

GreatCyber AI Video uses a split MiniMax-H3 deployment:

1. The Qwen3-VL conditioner prepares prompt_embeds, text_token_tags, and resolved geometry.
2. The MiniMax-H3 generator receives those conditioning tensors.
3. The FlashGen LoRA is merged into the H3 transformer at startup.
4. The distilled FlashGen sigma schedule is pinned to the video and audio schedulers.
5. MiniMax-H3 performs four denoising/model evaluations.
6. The video and native audio VAEs decode the result and the Space writes an MP4 with synchronized audio.

The Qwen3-VL conditioner is used as a separate Hugging Face Space because the H3 transformer and Qwen3-VL conditioner are too large to reliably coexist on the same ZeroGPU worker in this deployment architecture.

Models

• Base model: MiniMaxAI/MiniMax-H3
• FlashGen LoRA: Beidouqixing/minimax-h3-4step-lora-flashgen
• Conditioner: multimodalart/qwen3vl-conditioner

FlashGen 4-Step schedule

The FlashGen deployment uses the distilled base sigma schedule:

[1.0, 0.7, 0.4, 0.15, 0.0]