MiniMax H3 RAVEN LoRA: Real-Time Streaming Video in ComfyUI

RAVEN (Real-time Autoregressive Video Extrapolation with Consistency-model GRPO), a research project from Imperial College London, releases a LoRA adapter that turns MiniMax H3 into a causal streaming generator: instead of denoising one bidirectional clip, the model generates video chunk by chunk, extrapolating each new chunk from previously generated content. The released preview adapter generates 192 frames at 1376×768 / 24 fps with just 4 sampling steps (NFE), and the project ships official ComfyUI nodes (ComfyUI-MiniMax-H3-RAVEN-Streaming) plus an end-to-end example workflow.

How it works

RAVEN keeps MiniMax-H3's architecture and adds a streaming head trained with consistency-model GRPO: each chunk is generated from the previous context with a small, fixed number of function evaluations. This makes generation progressive — content appears chunk by chunk rather than all at once — which is the key building block for real-time interaction. The 4-NFE preview adapter targets the video and audio grids jointly, so text-to-video-and-audio (T2VA) streams come out of the same causal pass.

The released weight is a LoRA adapter (r=128, lora_alpha=128) loaded on top of MiniMax-H3; it is an initial preview release, and the authors note that texture detail is still being improved.

RAVEN turns bidirectional denoising into causal, chunk-by-chunk streaming extrapolation (figure from the project page)

Demo

Streaming generation reel from the RAVEN project page (video from yanzuo.lu)

Human-preference comparison from the RAVEN paper

Model details

ComfyUI usage

The official nodes were released on August 20, 2026 and are published on the Comfy Registry as comfyui-minimax-h3-raven-streaming. Install by searching MiniMax H3 RAVEN Streaming in ComfyUI-Manager, or comfy node install comfyui-minimax-h3-raven-streaming, then restart ComfyUI.

1. Install the custom nodes (ComfyUI-Manager or git clone into custom_nodes/).
2. Download minimax_h3_raven_streaming_lora_4nfe_preview.safetensors into ComfyUI/models/loras/ and the standard MiniMax H3 model files into their usual folders.
3. Load the bundled example workflow (Workflow → Browse Templates → Extensions, or drag the JSON below onto the canvas) and run.

The node pack implements the chunk-major streaming sampler, a streaming preview widget, and the mandatory RAVEN LoRA wiring on top of ComfyUI's built-in MiniMax H3 implementation. The authors validated 1376×768 / 192-frame generation inside a simulated 24 GiB VRAM envelope.

Download RAVEN streaming workflow

This is not native ComfyUI support: the streaming nodes depend on the third-party ComfyUI-MiniMax-H3-RAVEN-Streaming custom node pack.

Availability

• Weights: mvp-lab/MiniMax-H3-RAVEN-Streaming-LoRA (MiniMax H3 community license)
• Project page: yanzuo.lu/raven
• Paper: arXiv:2605.15190
• Code: mvp-ai-lab/RAVEN
• ComfyUI nodes: YanzuoLu/ComfyUI-MiniMax-H3-RAVEN-Streaming · Comfy Registry