HunyuanVideo Fresca

FreSca wraps a technique from the WikiChao/FreSca research project (linked directly in this node's description) for reshaping classifier-free guidance in frequency space instead of just scaling it uniformly. The idea, in plain terms: instead of one guidance strength applied to everything, you split the guidance signal into a low-frequency band (broad structure and composition) and a high-frequency band (fine detail, texture, edges) at a chosen cutoff, and scale each independently. That lets you push detail sharpness up without also over-cranking the overall composition strength, which is usually what happens when you just raise a single guidance scale too far.

It's a settings node - connect the output to HyVideoSampler's optional fresca_args input.

The inputs that matter

• fresca_scale_low (default 1) - how strongly the low-frequency (structural) component of guidance is scaled. Left at 1, structure follows your prompt's guidance normally.
• fresca_scale_high (default 1.25) - the high-frequency (detail) scaling. Set above 1 by default, meaning this node ships already nudging toward sharper detail out of the box.
• fresca_freq_cutoff (default 20) - where the split between "low" and "high" frequency happens. Lower values push more of the signal into the "high" bucket that gets the stronger scale; higher values keep more of the signal treated as structural.

Output: fresca_args.

How to install it

Ships with the pack - no separate download for the node itself.

• ComfyUI Manager - search ComfyUI-HunyuanVideoWrapper, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-HunyuanVideoWrapper, then pip install -r ComfyUI-HunyuanVideoWrapper/requirements.txt, restart.

No model files tied to this node - it's a pure sampling-time technique, nothing to download beyond the pack.

Common issues & troubleshooting

Connected it and see no visible difference. Confirm fresca_args is actually wired into HyVideoSampler - like the other settings nodes in this pack, a disconnected node silently does nothing rather than erroring.

Output looks oversharpened, gritty, or artifact-heavy. fresca_scale_high defaults above 1 already; if you've raised it further, that's the likely cause. Bring it back toward 1-1.25 and reassess.

Not sure whether this is worth using over just raising embedded_guidance_scale on the sampler. This is a genuinely different lever, not a replacement - the sampler's guidance scale is a single global strength, while FreSca lets you push detail and structure separately. If you've already found the sampler's guidance scale is fighting you (too much detail costs composition, or vice versa), this is the node built for exactly that tension. If you haven't hit that specific problem yet, there's less reason to reach for it.