Replicate nateraw/audio-super-resolution

This one does for audio what an upscaler does for images: feed it a low-quality or low-bitrate clip and it generates a higher-fidelity version, filling in detail a simple resampler can't invent. nateraw packaged it for Replicate; the node calls the hosted model rather than running anything locally. The ddim_steps and guidance_scale inputs are the giveaway for how it works - this is a diffusion model doing the upsampling, not a classic signal-processing filter, which is why it can genuinely reconstruct detail rather than just interpolating between existing samples.

Use it for cleaning up a compressed voice memo, breathing life into a low-bitrate export, or polishing generated audio (TTS, MusicGen output) that came out sounding thin or muffled. It's not a magic fix for badly clipped or heavily distorted audio - it's built for resolution, not repair of every kind of damage.

How it works

Diffusion audio super-resolution works the same way image diffusion upscaling does: the model is trained to reverse a degradation process, so given your lower-quality input as conditioning, it runs a denoising process (ddim_steps controls how many steps - more is slower but generally cleaner) guided toward output that matches the input's content while restoring higher-frequency detail. guidance_scale controls how strongly the output sticks to that conditioning versus how much freedom the model takes.

The inputs and outputs that matter

• input_file (required, AUDIO) - the clip to upscale.
• ddim_steps (default 50, range 10–500) - more steps generally means better quality at the cost of a longer, pricier run. 50 is a reasonable starting point; only push it higher if you can hear artifacts at the default.
• guidance_scale (default 3.5, range 1–20) - how strongly the output is guided toward the input. Push it if the result drifts too far from your source; pull it back if the output sounds over-processed.

seed and force_rerun round it out. Output is AUDIO - wire it into a save-audio node or further down an audio chain.

How to install it

One install covers every node in this pack.

cd ComfyUI/custom_nodes
git clone https://github.com/replicate/comfyui-replicate
cd comfyui-replicate
pip install -r requirements.txt

Restart, or use ComfyUI Manager: search ComfyUI-Replicate, install, restart. Set your token before launching:

export REPLICATE_API_TOKEN="r8_************"; python main.py

Token from replicate.com/account/api-tokens. Nothing downloads locally - this runs entirely on Replicate's servers.

Common issues & troubleshooting

Worth knowing up front: the pack's own README says it's built for models that return simple text or image output, and warns that audio outputs "will not work as expected" - audio support is explicitly still on the roadmap. This node returns AUDIO, so if you hit a rough edge (saving or previewing the clip misbehaving, an unexpected format), that's the documented limitation surfacing, not necessarily a mistake in your graph.

Node errors immediately. REPLICATE_API_TOKEN isn't set in the environment ComfyUI actually launched from - the most common first-run problem across this entire pack. On the portable Windows build, a token set as a global Windows environment variable specifically has been reported not to get picked up automatically; set it in the launch .bat if that happens.

Output sounds over-smoothed or "off." Diffusion-based reconstruction can invent detail that wasn't quite there - try lowering guidance_scale slightly, or reduce ddim_steps if it's overcooking short clips.

High ddim_steps runs are slow and cost more. Each is a billed Replicate prediction; iterate at a lower step count, then push it up for your final render.

Same output on identical input. Replicate's cache - change something or flip force_rerun.