Everyone complains about triggering guardrails when creating AI videos with human faces, but I still get a lot of good results. I don’t know for sure, but I think many websites promising “Seedance” are actually offering some “flavor” of Seedance 2.0.
The Chinese endpoint that I’m using still gives me solid results.
One photo. Four shots. No video model
A new LoRA just turned Qwen-Image-2.1 into a film director: the [Next-Scene LoRA](https://t.co/OaJXsWvNTY).
You give it a frame and tell it what the camera does, and it generates the **next scene**. The world, the characters, and the style all stay the same. Then you feed that output back in, and a single image grows into a full storyboard.
Their example starts from an Apollo 11 photo. The camera pulls back to a wide shot of the Moon, pans to Earth rising over the horizon, cuts to a low angle on the flag, then pushes in on the astronaut's visor. All of it is chained edits.
How to use it in ComfyUI:
- Any Qwen-Image-2.1 workflow with a LoRA node, at strength **0.7 to 0.9**
- Your image goes in as the edit/reference input
- The prompt always starts with `Next Scene:`, with **the camera move first**
`Next Scene: The camera pulls back to a wide shot, revealing the city at dawn.`
Tips from the people who tested it:
- Use the `next_scene_step2500` checkpoint (the recommended one)
- For long chains, stay at **0.7** and add a light 0.5px blur to each input, or the grain turns into color noise by scene 3 or 4
- Landscapes and cinematic establishing shots work best. Static portraits aren't its thing
Não consigo entender o tanto que a Higgsfield compra de influencer, são piores que as empresas de bet, abro o Instagram e a mina babando ovo deles como se fossem o auge da inovação e preço 😂 A empresa mais scam que existe em inteligência artificial