This ComfyUI workflow creates continuous, coherent scene sequences from a single initial image using Qwen Image Edit, a next scene LoRA and automatic prompt augmentation. The system features intelligent prompt generation, anti-repetition mechanisms, and seamless video output capabilities.

## ✨ Key Features

- **🔄 Single Button Operation**: Generate infinite scene sequences with one click

- **🧠 Intelligent Prompting**: Qwen VL understands cinematic progression and scene continuity

- **🔄 Cache Queue System**: Rolling buffer maintains reference frames for coherent generation

- **⚡ Anti-Repetition**: Built-in mechanisms prevent camera movements that revert to previous frames (WIP)

## How to use:

-**Pre-configure (IMPORTANT)**: This workflow uses an external llama.cpp or cloud API to provide prompt augmentation. Install custom nodes and follow the instruction at  https://github.com/asaddi/ComfyUI-YALLM-node to configure API for your local or cloud VLM API (Qwen 3 VL is preferred). You may also swap it out for the Qwen 3 VL node if you have enough VRAM to avoid constant reloading. I have tested with Qwen 3 VL 30B A3B Q4_K_M with decent results.

- Use group muter to activate the **init** group and turn off the **run** group. Load the initial frame on the top left and run the workflow.

- Turn off **init** group and turn on **run** group. Now hit "run" as many times as you want!

- Eventually Qwen Image Edit might drift too far from the theme or degrade the image too much, so this is not entirely infinite. But generally, images generated by Qwen Image itself will likely work better.

- You can also use Qwen-Image-Edit-Rapid-AIO checkpoint or lightning LoRA, but they do degrade quality and diversity. The default option of 20 steps at CFG=2.5 is a good starting point.

## How to go from here:

Use WAN 2.2 FLF! You can get very nice results even with 4 step LoRA for WAN!