Video catalogue · Verified local path

RAVEN local guide

Non-commercial causal autoregressive video model for streaming local text-to-video extrapolation on Wan 2.1 or MiniMax H3.

Choose an app

Start with the recommended app, or choose another supported option.

Compare video models
Open model filesRecommended verified starting pointOpen in UnslothLaunches Unsloth Desktop on this modelGet ComfyUI DesktopOfficial desktop app for this compatible modelOpen advanced setupPython runtime for experienced usersCheck my machineSee whether this model fits your saved hardware

Desktop app links require the app to be installed. If nothing opens, LocalClaw will show app-download and model-file fallbacks.

What it does

Text To VideoStreaming VideoLong Video GenerationText To Audio VideoAnimation

MVP Lab publishes public RAVEN Wan2.1-T2V-1.3B checkpoints under CC BY-NC 4.0, including the post-distillation backbone, CM-GRPO LoRA, merged backbone and archival full bundle, plus an official Linux/Python/CUDA training and validation repository. The same project now points to ComfyUI MiniMax H3 RAVEN Streaming nodes with a mandatory public RAVEN LoRA, stock ComfyUI model folders, Comfy Registry installation and measured 1376 x 768, 192-frame T2VA generation inside a simulated 24 GiB VRAM envelope. LocalClaw records 64 GB RAM and 24 GB NVIDIA VRAM as the conservative practical floor for that ComfyUI path, while the research repository itself targets Hopper/CUDA 12.8 and site-specific trial YAML paths.

Hardware figures are practical entry floors, not performance guarantees. Resolution, duration, precision, offloading and runtime versions can materially change memory use.

Strengths

  • Official RAVEN weights and primary code repository
  • Streaming causal extrapolation with four-NFE preview LoRA support
  • ComfyUI node path validates a 24 GiB VRAM MiniMax H3 workflow

Limits to know

  • Weights are non-commercial and inherit upstream Wan/MiniMax component terms
  • The MiniMax H3 preview LoRA is described as undertrained with limited texture detail
  • The research scripts target Hopper GPUs and require manual checkpoint path wiring