STAR Video Super-Resolution

Upscaling video is a different job from upscaling an image, and this node is built for the hard version. Feed it an IMAGE batch of frames and it runs the STAR diffusion pipeline - 4x super-resolution with temporal coherence, the thing a good image upscaler applied frame-by-frame can't give you. The classic failure: a per-frame upscaler that handles a patterned wallpaper perfectly in stills and makes it shimmer like a broken GIF in motion, because nothing forces consecutive frames to agree. STAR is temporally aware by design - a temporal VAE, ControlNet conditioning, and a diffusion pass that sees whole chunks of frames at once, built on the I2VGen-XL / VEnhancer lineage.

Wire star_model from the STAR Model Loader, and images from anything that outputs frames - LoadImage for a single photo, VHS LoadVideo, or a video loader. It emits images back out as a batch you save with VHS or SaveAnimatedWEBP. Single-image use works fine, but the model was born for footage: genuinely degraded old clips are where it shines, not clean renders that an ESRGAN pass handles faster.

The inputs that matter

• upscale (1–8, default 4): the factor. 4x is the sweet spot; higher eats VRAM fast.
• solver_mode: fast uses STAR's optimized schedule - 4 coarse + 11 fine steps, hardcoded, and the steps input is ignored. normal uses a uniform schedule and respects steps. Leave it on fast unless you're deliberately experimenting; it's the difference between "slow" and "bring a book".
• denoise (default 0.9): how much of the input structure survives. Lower = more faithful to the original, less invented detail. 0.9 is the STAR default and a good place to sit.
• guide_scale (default 7.5) and prompt: STAR always appends its built-in quality prompt, so leaving the prompt blank is not a mistake - that's the intended use. Raise guide_scale if you want the model's interpretation to dominate.
• max_chunk_len (default 32) and segment_size (default 0): your VRAM and RAM levers. Lower max_chunk_len to shrink the frames-per-chunk and fit longer videos in less VRAM (chunks overlap 50% to keep motion coherent). Set segment_size to 16–32 for long videos to bound system RAM; segments crossfade at 25% overlap so the seams stay invisible.
• color_fix (default adain): post-pass color matching so the upscaled footage doesn't drift from the source. wavelet preserves low-frequency color instead; none if you want to grade later yourself.

Install and first run

Same as the loader - ComfyUI Manager, search ComfyUI-STAR, or:

cd ComfyUI/custom_nodes
git clone --recursive https://github.com/ethanfel/Comfyui-STAR.git
cd Comfyui-STAR
pip install -r requirements.txt

First run downloads the checkpoint into ComfyUI/models/star/ (choose light_deg.pt vs heavy_deg.pt in the loader), plus the OpenCLIP encoder and SVD temporal VAE. Restart ComfyUI after installing.

Where people get burned

It's heavy. This is diffusion, not a filter - expect minutes per clip, and real footage costs real GPU. One community run needed a rented pod and ~90 minutes for a single 4K clip, and a 299-frame scene at 430p wouldn't fit even in 80GB VRAM until it was downscaled first. On a consumer card, set the loader's offload to model or aggressive and drop max_chunk_len - or skip ComfyUI entirely for very long videos and run the pack's standalone inference.py, which streams frames straight to ffmpeg and copies the audio:

python inference.py input.mp4 -o output_4x.mp4 --offload model --segment-size 8

Heavily compressed sources will make the model invent structure (it has to). Clean the footage first if you can, and give it footage, not a slideshow - temporal coherence only pays off when there's motion to be coherent about.