Check out the latest model drops and powerful integrations.
In Stream DiffusionV1, feeding in a noise texture can produce immediate results. Scope works differently — basic noise alone will look flat. You need to deliberately tune four key parameters.
VACE is Scope's equivalent of ControlNet. It extracts structural information from your input and uses it to guide generation — giving you structure plus stylization.
In my example, I use a Scribble preprocessor that generates an edge map from the input. Without VACE, output is too unconstrained. When you activate VACE, Scope will use both the video input and the pre-processor input as information to generate the video.

When turning on VACE, also turn on 'Use Input Video' or else VACE will not work as intended
Scope defaults to 2 denoise steps, which typically produces weak results. Use 4 steps instead. The slight FPS decrease is worth the improved depth.
A strong configuration:
If you only adjust one thing, adjust Step 2 for best change in visual output.

Scope models like LongLive need much more descriptive prompts than typical diffusion workflows. "Galaxy" won't cut it — think multi-sentence, even paragraph-length descriptions.
Quick workflow: Write a simple base idea → combine it with the model's official prompting guide → use an LLM to expand it into a detailed, model-aware prompt.
Controls how strongly VACE guidance influences output. Range is typically 0.5–1.5: lower values give looser, more abstract results; higher values enforce stronger structural adherence. The ideal setting varies by preprocessor, so experiment.

Use VACE with a preprocessor, increase denoise steps (tune Step 2 especially), write detailed prompts, and adjust Base Scale. Scope rewards intentional tuning — and delivers more controlled, expressive results than V1 when you put in the effort.