AIAny
AI Video2025
Icon for item

ViMax: Agentic Video Generation

Turns a raw idea, novel, or screenplay into a complete multi-shot video through a multi-agent pipeline that scripts, storyboards, and renders shots while a vision model checks character and scene consistency across the whole story.

Introduction

The hard part of AI video was never generating one good clip — it was keeping the same character's face, the same lighting, and a coherent plot across the dozens of clips stitched into minutes of footage. ViMax reframes long-form video as an orchestration problem rather than a single model call, wiring a crew of role-specialized agents (director, screenwriter, producer) around whatever underlying image and video generators you plug in.

What Sets It Apart
  • Four entry points, one pipeline. Idea2Video, Novel2Video, Script2Video, and AutoCameo all funnel into the same scripting → storyboard → shot-generation flow, so you can start from a one-line concept or a finished screenplay and reach the same downstream consistency machinery.
  • A consistency check that actually runs. A multimodal model (MLLM/VLM) inspects generated frames and flags drift in characters and scenes before shots are assembled — the step most DIY text-to-video chains skip, which is why their outputs visibly wander.
  • RAG-grounded screenwriting. Script generation is retrieval-augmented rather than one big prompt, keeping a long story's continuity intact and letting shots be planned and rendered in parallel instead of in one slow chain.
Who It's For

Great fit if you want minutes-long, multi-scene videos — episodic adaptations of a novel, a narrated idea, or cameo clips built from your own photos — and care more about character and scene continuity than per-clip polish. Look elsewhere if you need a single hero shot at maximum fidelity, real-time generation, or a turnkey hosted product: this is a research-grade, self-hosted Python system that orchestrates external generators, so final quality and cost ride on the models you supply.

Information

  • Websitegithub.com
  • OrganizationsData Intelligence Lab @ HKU
  • AuthorsHKUDS
  • Published date2025/03/30

Categories

More Items

Hugging Face
AI Video2026

Provides Parallel Decoding Distillation (PDD) LoRA adapters that accelerate MiniMax-H3 video generation into few inference steps. Includes official 8-step Acc LoRAs for FL2VA and Ref2VA (rank=64, network_alpha=64, BF16), demo comparison videos, and example scripts using Diffusers' MiniMax-H3 ModularPipeline.

Hugging Face
AI Video2026

Conditions a MiniMax‑H3 video generator with a single ControlNet‑Union checkpoint to accept Canny, Depth, HED, MLSD or Pose control videos and run video inpainting. Guidance‑distilled for one‑pass inference; requires the base MiniMax‑H3 weights and specific control-branch config.

Hugging Face
AI Video2026

Upscales Minimax H3 24-channel VAE latents in-place to increase spatial resolution while preserving the time dimension. Replaces the decode→pixel-upscale→encode round-trip with a learned 2D/3D latent upscaler to save compute and avoid interpolation ghosting; supports 1.0–4.0× scaling.