OmniHuman-1.5

OmniHuman-1.5 is the follow-up to ByteDance's audio-driven human animation model, and it's the one I'd reach for of the two. Same core trick - one image of a person plus an audio clip becomes a talking, moving video - but 1.5 adds two genuinely useful things: a mask input that tells the model what is allowed to move, and a fast mode that trades a little quality for much quicker iteration. Still fully closed-source and cloud-hosted, so it's a pay-per-run node, but for iterating on a talking-head shot it earns its keep.

How it works

Like v1, the audio clip is the puppeteer: it conditions lip sync, expression, and motion on the audio signal. The request goes up to Runware's cloud, ByteDance's model renders, and a native VIDEO comes back. The new mask input is the interesting addition - feed it a mask and the model restricts animation to that region, which is exactly what you want when you have a still background and only the person should move, or when you want to keep part of the frame frozen.

The inputs that matter

• image (IMAGE, required) - the person. A clear front-facing shot still does the heavy lifting; this is your identity anchor.
• audio (STRING, required) - the driving audio as a UUID or URL. Clean clip, one voice.
• mask (IMAGE, optional) - the v1.5 upgrade. Restrict animation to the masked region. Great for keeping backgrounds rock-steady. Leave it empty to animate everything.
• seed - the v1.5 upgrade number two. v1 didn't expose it; 1.5 does, so you can lock a take and reproduce it.
• providerSettings.bytedance.fastMode (BOOLEAN) - the accelerated pipeline. Use it for exploration, switch it off for the final render.
• numberResults (1–4) - seed-based variations, the cheap hedge on which take nails the expression.
• outputFormat / outputQuality - delivery format (MP4/WEBM/MOV) and compression. Same as the rest of the video catalog.
• safety - the usual toggle gating safety.checkContent and safety.mode (none/fast/full). With human animation, leaving a content check on for shared workflows is cheap insurance; the node reports the result in its title bar.

Wiring it up

Output is video (native VIDEO). Source the image from anywhere in your graph - a generation node, a loaded photo - and the audio from a URL or uploaded UUID. If you're using the mask, it's an IMAGE too, so you can build it locally (a loaded mask, a black-white composite) and let the model do the rest.

Install and setup

The pack installs once and every node comes along: ComfyUI Manager → "Runware" → install → restart, or clone and pip install -r ComfyUI-Runware/requirements.txt. Dependencies are just runware-sdk, pillow, and soundfile. API key via Settings → Runware API key, RUNWARE_API_KEY, or runware auth login, plus funded credits - the community's standard gripe is the ~$20 minimum top-up with a card on file, and video runs burn credits faster than images, so watch the per-run cost on the title bar.

Troubleshooting

• Video appears as a path, not a VIDEO socket - older ComfyUI lacks VideoFromFile, and the pack falls back to saving to the output folder. The file's there; grab it.
• Fast mode output is rougher than expected - that's the trade, not a bug. Iterate in fast mode, flip it off for the final.
• Everything animates when you wanted a locked background - you skipped the mask. Feed one.
• Reproducibility - if a run gave you exactly the take you want, note the seed and reuse it; that's what it's for.

1.5 is the version I'd install for - the mask and the seed make it controllable where v1 felt like a lottery ticket. Still a cloud model, still per-run, but it's the most polished "photo speaks" node in the catalog.