OmniHuman 2.0 vs 1.5: Which One to Run for Your Clip
Five minutes in one pass or 1080p with a turbo switch? Here's OmniHuman 2.0 vs 1.5 compared input by input, with a verdict for every kind of clip.
By Eftal Yurtseven · Nov 20, 2026 · 9 min read

You have one photo, one audio file and a deadline. Two versions of the same ByteDance avatar model sit next to each other in the catalog, both promise a talking, gesturing person, and the model pages don't tell you which one fits the clip you're actually making. Picking wrong doesn't break anything. It just costs you an afternoon of stitching, or a render that looks softer than the brief needed.
So here's the OmniHuman 2.0 vs 1.5 decision, made on the facts ByteDance, BytePlus and the each::labs model pages publish. OmniHuman 2.0 and OmniHuman 1.5 are both live on each::labs, they share a family resemblance, and they split cleanly on one thing more than any other: how long the performance has to run. Everything else follows from that.
If you want the deep dive on 2.0 alone, our OmniHuman 2.0 spotlight covers it. This post is narrower. It's the side-by-side, and a verdict per job.
What OmniHuman 1.5 Actually Is
OmniHuman 1.5 comes out of ByteDance's Intelligent Creation team, and its research paper, "OmniHuman-1.5: Instilling an Active Mind in Avatars via Cognitive Simulation", landed on arXiv on 26 August 2025. The idea in that title is the whole pitch. Earlier avatar models mostly synced to low-level cues like the rhythm of the audio. 1.5 tries to understand what's being said.
It does that with a two-part design borrowed from the "System 1 and System 2" theory of thinking. A multimodal large language model plays the slow, deliberate planner: it reads the audio's meaning and emotional subtext and turns it into structured guidance. A Diffusion Transformer plays the fast, intuitive half and renders the motion. The practical result, per the project page, is gesture that matches words and intent instead of a head bobbing politely to the beat.
The each::labs model page describes 1.5 as generating "expressive, full-body video performances from a single image and up to 30 seconds of audio." It handles portraits, full-body shots and cartoon-style avatars, and it accepts speech, dialogue and singing. It's the successor to the original OmniHuman, which still runs on each::labs with nothing but an image and an audio file as inputs. Our older OmniHuman-1 explainer covers where the family started.

What OmniHuman 2.0 Changes
OmniHuman 2.0 is offered through BytePlus, ByteDance's enterprise cloud, and the headline on the BytePlus OmniHuman page is length: "30s to 300s Video Output." Five minutes from one image and one voice track, in a single generation.
The rest of the 2.0 page reads like a list of things 1.5 hinted at, promoted to headline features. BytePlus says 2.0 "Enables rhythmic, emotional, and multi-person performances," lists a dedicated Multi-Person Scene feature, and claims precise gesture control and camera movement through text prompts. It creates digital singers and handles high-pitched performances. It also widens the cast: animals such as cats and dogs, plus more non-humanoid images and styles. And if something isn't in your reference image, BytePlus says it can be added with prompt words, so a prop or a background element doesn't have to be painted in beforehand.
Output is native 720p, per BytePlus. That's a real trade, and it matters later in this post.

OmniHuman 2.0 vs 1.5, Input by Input
The request schemas on each::labs tell the clearest story, because they show what you can actually set.
OmniHuman 2.0 takes three fields. image_urls is required and is an array, with at least one image. The schema accepts JPEG, PNG, WebP, BMP, TIFF, GIF, HEIC and HEIF, each under 30 MB, between 300 and 6,000 pixels on a side, with a width to height ratio between 0.4 and 2.5. That ratio range is generous: tall phone portraits and wide landscape frames both fit. audio_url is required too, WAV or MP3, between 4 and 300 seconds, under 15 MB. The optional prompt carries "scene, movement or camera instructions" and supports Chinese, English, Japanese, Korean, Mexican Spanish and Indonesian.
OmniHuman 1.5 takes five. image_url is a single portrait image, and audio_url is the voice or song. The optional prompt supports the same six languages. Then come the two knobs 2.0 doesn't have: resolution, which accepts 720p or 1080p and defaults to 1080p, and turbo_mode, off by default, which the schema says "speeds up generation by sacrificing some quality."
Read those side by side and the shape is obvious. 2.0 gives you more runway and fewer dials. 1.5 gives you a shorter clip with control over sharpness and speed. Neither schema exposes a duration field: in both cases the audio sets the length of the video.
Length Is the Real Split
This is where most decisions get made, so be blunt about it. If your audio runs longer than half a minute, the comparison is mostly over.
With 1.5, the each::labs page frames clips at up to 30 seconds of audio and recommends splitting long scripts into shorter pieces, because long clips can drift in pose and expression. That's fine for an ad cut or a social hook. It's painful for a three-minute product walkthrough, where every split is a seam: a hand that jumps, an expression that resets, a camera that starts over. You can hide seams with B-roll. You can't make them disappear.
2.0 accepts 4 to 300 seconds of audio in one call. A five-minute explainer becomes one continuous performance with one arc of gestures, which is exactly what a viewer reads as "a person talking" rather than "a sequence of clips of a person talking." Treat that as the default reason to reach for 2.0.
The flip side is also true. Under 30 seconds, length stops being an argument, and the other differences decide it.

Motion, Music and More Than One Face
Both versions are built to perform, not just lip sync. The 1.5 research page talks about analyzing the emotional subtext of audio without any text prompt, producing gestures that fit the meaning of the words, and handling music: one image plus one song becomes a singer, with natural pauses and breaks. 2.0's BytePlus page pushes the same direction further, describing "soulful digital singers with expressive motion" and lip movements, emotions and gestures synced to audio.
On more than one person, both sources make claims. The 1.5 project page shows complex multi-character interactions, and the each::labs 1.5 FAQ says the model handles multi-character scenes where you can specify who speaks and how background characters react. BytePlus lists Multi-Person Scene as a 2.0 feature. On each::labs, each version takes one audio file per request, so a two-person exchange is one mixed track, and the prompt is where you say who's talking.
Non-human subjects lean toward 2.0. The 1.5 paper says its approach extends to non-human subjects, and the project page shows animals and stylized cartoons, but the each::labs 1.5 page steers you toward human-centric content and front-facing or three-quarter faces. BytePlus explicitly markets 2.0 for cats, dogs and non-humanoid styles. If your mascot has whiskers, start with 2.0.
Prompt Control and Output
The prompt is optional on both, and on both it's worth writing. The 1.5 FAQ on each::labs lists camera angle, gesture emphasis and emotional tone as things a prompt can guide. BytePlus says 2.0 responds to prompts for object generation, camera movement and specific actions, and adds the ability to bring in elements missing from the reference image.
Keep prompts short and physical. Describe what the body and the camera do, not what the speech means: the audio already carries the meaning, and both models were built to read it. Something like this works as a starting point for either version:
Medium shot, she speaks directly to camera with open hand gestures on the key points, slow push in during the last sentence, warm window light from the left.
Output is an MP4 on both model pages, returned as a hosted video URL. The visible difference is resolution. 1.5 defaults to 1080p and lets you drop to 720p, while BytePlus states native 720p for 2.0. If the clip is headed for a large screen or will be cropped hard in the edit, that extra resolution on 1.5 is a reason to keep it in the rotation.

When to Pick Which
Here's the verdict by job, with the reason behind each call. Where a job splits in two, you get both halves.
Talking-head explainer: OmniHuman 2.0
Explainers, onboarding videos, course modules and product walkthroughs run long, and continuity is the whole effect. 2.0's 300-second ceiling lets you render the full script in one pass, with no seams to cover. Write the voiceover in a model from the text-to-voice catalog, such as MiniMax Speech 2.8 HD, then hand the file straight to 2.0. The exception: a 20-second intro for a big-screen presentation, where 1.5 at 1080p is the better fit.
Music performance: OmniHuman 2.0 for the full song, 1.5 for the hook
Both versions sing. A full track almost always runs past 30 seconds, and BytePlus names AI music videos as a 2.0 use case, so a complete performance belongs on 2.0. For a short chorus clip meant for a vertical feed, 1.5 at 1080p gives you a sharper loop.
UGC ad: OmniHuman 1.5
UGC-style ads live in the 10 to 30 second window, and you'll make many versions of each one. That's 1.5's home turf. Turbo mode lets you rough out hooks fast, then you switch it off and render the keeper at 1080p. Same image, same audio, two settings. If the ad features a pet or a cartoon spokes-creature, jump to 2.0 instead.
Multi-character scene: OmniHuman 2.0
Two people in one frame, trading lines, is where scene length and performance both stretch. BytePlus calls out Multi-Person Scene and natural interaction with the environment as 2.0 features, and a conversation rarely fits in half a minute. Put both characters in the reference image, mix the dialogue into one track, and use the prompt to say who speaks and how the listener reacts.
Running OmniHuman on each::labs
Both versions sit in the OmniHuman family on each::labs behind the same API key and the same prediction call. Switching from one to the other is a change of model slug plus the input names: image_urls as an array for byteplus-omnihuman-2-0, image_url as a single string for bytedance-omnihuman-v1-5. That symmetry makes a split workflow easy: draft short hooks on 1.5, render long cuts on 2.0, and keep one integration.
A clean pipeline looks like this. Start with a well-lit, front-facing reference image; if you need to create one, Seedream 5.5 keeps the ByteDance pipeline in one family, and image-to-video models are one click away if you want B-roll around the avatar. Generate or record the audio and check its length first, because that number picks the model for you. Add a short physical prompt. Then chain the steps in an each::labs flow so voice, avatar and edit run as one job instead of three manual hops. If you're wiring up more of the ByteDance lineup, our ByteDance API guide shows how Seedance, Seedream and OmniHuman share one key.
For a multi-character prompt on 2.0, keep the same physical register:
Two friends at a café table, the woman on the left speaks first and leans in, the man on the right nods and laughs, then answers with a shrug, static camera at eye level.
Frequently Asked Questions
Is OmniHuman 2.0 better than 1.5?
Longer, yes: ten times the audio runway on each::labs, and BytePlus markets 2.0 for multi-person scenes and animal subjects. Sharper, no: 1.5 keeps a 1080p option and a turbo toggle that 2.0's schema doesn't expose. For anything over 30 seconds, 2.0 wins; for short, sharp clips you're iterating on, 1.5 is the stronger pick.
What is the maximum length in OmniHuman 2.0 vs 1.5?
The each::labs schema for 2.0 accepts audio from 4 to 300 seconds, which matches BytePlus's "30s to 300s Video Output." The each::labs 1.5 page describes clips driven by up to 30 seconds of audio and recommends splitting longer scripts.
Can OmniHuman 2.0 animate animals or cartoons?
BytePlus says so directly: 2.0 supports more styles and image subjects, including cats and dogs, plus more non-humanoid images. The 1.5 research shows stylized and animal characters too, but its each::labs page recommends human faces for the most reliable results.
Do I need a prompt for either version?
No. Both models read emotion and meaning from the audio on their own, and the prompt field is optional in both schemas. Still, a short line about camera and gesture is the fastest way to steer the performance, and both accept it in six languages.



