Turn one (or two) reference photos plus speech audio into identity-preserving talking videos — monologue or side‑by‑side dialogue — with DreamID‑Omni, Wan 2.2 video backbone, and MMAudio‑driven lip motion.

Who it’s for: creators shipping avatar clips, personalized messages, or two‑character dialogue tests. Not for: quick experiments if you cannot install the DreamID‑Omni node stack and weights locally — use cloud first to validate the idea.

Try this workflow online before you download

See it run on RunComfy (browser)

Why RunComfy first
- Skip weight hunting — DreamID‑Omni + Wan 2.2 + MMAudio dependencies are heavy; RunComfy can host the runnable graph so you are not stuck on missing nodes mid‑queue.
- Sanity‑check lip‑sync fast — hear whether your audio drives the mouth convincingly before you tune local VRAM settings.
- Same graph intent — the published archive matches the RunComfy runnable workflow JSON (cloud is the quick path; local is the full‑control path).

When downloading for local ComfyUI makes sense — you want offline batches, custom paths, or to plug this into your own model folders and extensions; the zip is for that workflow.

How to use (local ComfyUI)
1. Load Load Reference Image (Person 1) and Load Reference Audio (Person 1) for single‑speaker; add Person 2 image/audio for dialogue.
2. Run ComfyUI DreamID-Omni Sampler on the single path (#21) or dual path (#22) with a short scene prompt (background, framing, tone).
3. Export from 💾 Save Video (Single Person) or 💾 Save Video (Two Persons); use distinct filenames so single vs dual outputs do not overwrite.

Expectations — First local run may download large weights (DreamID‑Omni / Wan). Use clean speech audio (minimal music/noise) for best lip‑sync; a free RunComfy account may be required to run in the browser.

Overview

DreamID‑Omni keeps the face from your photo while audio drives the mouth. Two parallel subgraphs cover one speaker or two speakers in one frame for conversational shots.

Features

• Single‑person and two‑person talking video in one graph

• Identity‑aware talking head with MMAudio conditioning

• Shared DreamID‑Omni loader for both sampler paths