STAR Video Super-Resolution

Upscale video frames using STAR diffusion-based super-resolution.

By ethanfel·Created 6 months ago·Updated 6 months ago· 0

◄upscale4►

◄steps15►

◄guide_scale7.5►

◄prompt►

◄solver_modefast►

◄max_chunk_len32►

◄seed0►

◄denoise0.90►

◄color_fixadain►

◄segment_size0►

CategorySTAR

Inputs (12)

Name	Type	Default	Description
star_model	STAR_MODEL		Connect from STAR Model Loader.
images	IMAGE		Input video frames (IMAGE batch). Can come from LoadImage, VHS LoadVideo, etc.
upscale	INT	41–8	Upscale factor applied to the input resolution. 4x is the default. Higher values need more VRAM.
steps	INT	151–100	Number of denoising steps. Ignored in 'fast' solver mode (hardcoded 15). More steps = better quality but slower.
guide_scale	FLOAT	7.51–20	Classifier-free guidance scale. Higher values follow the prompt more strongly. 7.5 is a good default.
prompt	STRING		Optional text prompt. STAR's built-in quality prompt is always appended. Leave empty to use only the quality prompt.
solver_mode	COMBO	fast	fast: optimized 15-step schedule (4 coarse + 11 fine). normal: uniform schedule using the steps parameter.
max_chunk_len	INT	324–128	Max frames processed at once. Lower values reduce VRAM usage for long videos. Chunks overlap by 50%.
seed	INT	00–18446744073709550000	Random seed for reproducible results.
denoise	FLOAT	0.900.1–1	How much noise to add before denoising. 0.9 = STAR default (preserves input structure). 1.0 = full denoise (starts from pure noise). Lower values preserve more of the original.
color_fix	COMBO	adain	Post-processing color correction. adain: match color stats from input. wavelet: preserve input low-frequency color. none: no correction.
segment_size	INT	00–256	Process video in segments of this many frames to reduce RAM usage. 0 = process all at once. Recommended: 16-32 for long videos.

Name	Type	Description
images	IMAGE	—