I tested HeyGen's Photo Avatar, Digital Twin, Avatar V, and Instant Avatar using identical scripts. Avatar V preserved likeness most closely. See the results, costs and recommendations.
An estate agent I coach asked me a simple question last month: should she create a Photo Avatar, a Digital Twin, an Avatar V avatar or an Instant Avatar? She had seen all four names in HeyGen tutorials and assumed they were four competing products.
They overlap more than their names suggest. Photo Avatar and Digital Twin are two ways to create an avatar of yourself.
Avatar V is the model that renders a Digital Twin. Instant Avatar is the earlier name for the video-avatar flow that became the Digital Twin.
So I ran the comparison she wanted.
Over two weeks, I created all four using the same person, then gave each the same market update script, a 10-minute training module and a translation test. This guide explains each avatar type, compares their features and covers three head-to-head tests, credit costs and which option best suits your needs.
At a glance
Winner: a Digital Twin rendered on Avatar V. It kept her face, teeth and gestures consistent across every test.
A Digital Twin on Avatar IV also learns real gestures, but becomes less consistent over longer scripts. Photo Avatar starts with a single photo but generates its own movement. Instant Avatar is the older workflow that now runs as a Digital Twin.
Key Features Compared
Avatar V
I recorded the agent once: 15 seconds on a laptop webcam, a live consent clip, and a two-minute expressive take as the motion reference. Avatar V became the default rendering engine as soon as her Digital Twin finished training.
Her habit of tilting her head before a key point carried into the first render, and her teeth looked identical at second 10 and minute 9. That detail matters because general video models redraw a face slightly differently every few seconds.
Avatar V makes preserving identity a core design constraint and trains voice alongside facial movement so mouth shapes follow the audio. It prioritises audio, followed by the expression in the input image, then the text prompt, making it well suited to an AI video avatar created from expressive footage.
Main strengths
- Identity Lock: Avatar V achieved a face similarity score of 0.840 compared with Veo 3.1's 0.714, and human assessors preferred it in 68.9% to 85.7% of direct comparisons with Kling, Seedance, Veo and OmniHuman.
- Gaze Control: The model maintains the gaze direction from the opening frame, with two presets for front-facing and side-angle setups for an AI talking head shot.
- Custom Motion: You can guide gaze, gestures, posture and energy in plain English when the audio alone does not convey the delivery you want.
- Stability for longer videos: A 30-minute talking video renders in a single pass with a consistent facial appearance, making it well suited to a complete course builder workflow.
- Cinematic scenes: A verified Digital Twin can feature in Avatar Shots, placing your face in Seedance 2.0 footage for clips lasting up to 15 seconds.
- Developer Access: The API provides access to Avatar V at $0.05 per second, so product teams can generate presenter footage within their own apps.
Where There’s Room for Improvement
- Avatar V does not work with photo-based looks, so you need at least one video look before you can use it.
- Prompts have the lowest priority with this model, so flat or monotonous audio will produce a static avatar, regardless of what you type.
Verified customer outcomes
Educator Anton Voroniuk saves 15.5 hours a week creating courses with his avatar and has reached more than 1 million students. At enterprise level, Advantive cut content creation time by 50% and reduced voice-over production from days to 2–3 hours for a team supporting over 600 employees.
Pricing
Avatar V renders cost 20 credits per minute on web plans, starting with Creator at $29 per month or $24 per month billed annually, with 600 monthly credits. API usage costs $0.05 per second, charged to a separate USD wallet.
This engine comes out ahead in the comparison as the only option that learns your natural movements and preserves your exact likeness, whether in a 30-second clip or a 30-minute course.
Digital Twin (on Avatar IV)
Before Avatar V launched in April 2026, the Digital Twin ran on Avatar IV, which remains available as an option for each video. I re-rendered the agent's twin on Avatar IV using the same script and appearance to isolate the difference.
Avatar IV maps audio to facial expressions and captures body language, eyebrow raises and pauses from your footage. Her hand gestures looked natural, and full-body framing worked well for an introduction delivered standing up.
The difference became apparent over time. By minute six of the training script, her smile appeared slightly wider than in the original footage. The Help Centre recommends tackling this issue in the following order: switch to Avatar V first, then Avatar IV, then re-record.
Key strengths
- Responsiveness to prompts: Custom prompts such as "talks excitedly" have more influence on gestures in Avatar IV than in Avatar V, which is useful for one-off creative edits.
- Full-body motion: The twin supports full-body shots with simple or dynamic backgrounds, rather than just head-and-shoulders framing.
- One Asset, Two Engines: You keep the same twin and choose an engine for each video, so you can try both without recording again.
- Speed or Quality Modes: Avatar IV offers a faster creation mode when turnaround matters more than polish.
Limitations
- Facial detail varies more over longer scripts than with Avatar V, most noticeably in the teeth and the width of the smile.
- Avatar IV renders cost 20 credits per minute, the same as Avatar V, so you pay the same rate for less consistent results.
- Correcting a shifting smile often means re-rendering, and each A-roll re-render uses more credits.
Pricing
One Digital Twin slot is included at no extra cost, with up to 500 looks. Avatar IV rendering costs 20 credits per minute on paid plans.
The Avatar IV Digital Twin suits creative experimentation where prompt control matters, but falls short of Avatar V in maintaining the consistent likeness that makes a digital twin worth creating.
Photo Avatar
The agent’s first choice was to use a photo, as she hates being filmed. I uploaded her best headshot, and a talking version was ready before she’d finished choosing a voice.
Photo Avatar brings a still image to life with generated expressions, head movement, and gestures. Avatar IV powers it by default, and it handles real people, illustrations, 3D renders, and non-human characters. That range makes it the most flexible option here for brand mascots and stylised presenters.
The trade-off is that the movement comes from the model, not from you. Her natural head tilt never appeared, and a broad smile in the original photo made her mouth look too large on certain words. The Help Centre recommends a photo with a subtle smile for precisely that reason.
Key strengths
- No camera needed: One photo gets you started, and uploading more photos gives the identity more reference material for future looks.
- Character Range: Avatar IV animates illustrations, 3D characters and non-human designs that video-based avatars cannot represent.
- Prompt Control: Motion prompts have more influence on photo looks than on Avatar V, so you can specify the energy you want.
- Plenty of slots: Free accounts get three photo avatar slots, while paid plans offer unlimited slots with 500 looks each.
Limitations
- Avatar V does not support photo-based looks, so you need a video look to access the most consistent engine.
- The model generates movement rather than learning yours, so close colleagues will notice that your mannerisms are missing.
- Each photo avatar training session costs 60 credits on paid plans, and each generated look costs 1 credit.
Pricing
Photo Avatar renders cost 20 credits per minute on Avatar IV or 3 credits per minute on Avatar III, with training at 60 credits per avatar on paid plans.
Using a photo works well for mascots, illustrated presenters and quick tests, but it cannot match a twin trained on video when it comes to looking and moving like you.
Instant Avatar
The agent had an Instant Avatar from 2024, created from around two minutes of footage. Many long-standing HeyGen users have one, and the name still appears in older tutorials and in some API error codes.
Instant Avatar was originally the quickest way to create a custom video avatar on the platform. Now, the same "Clone a Real Person" process creates a Digital Twin instead, using as little as 15 seconds of webcam footage plus a live consent clip.
Her old avatar still rendered and lip-synced accurately. However, the footage had a cut in the middle and uneven lighting, both of which are rejected by the current upload checks because they affect the avatar's stability.
The more significant limitation is access to the engine. Avatar V requires clean, continuous video footage, so an avatar trained on older footage often cannot use the model that delivered the best results in every test below.
Main strengths
- Existing Assets Keep Working: Older Instant Avatars still render, so any video library built using them continues to work as before.
- Low-cost testing: Rendering with Avatar III loops the footage and applies lip sync for 3 credits per minute, making it suitable for testing scripts and timing.
- Familiar workflow: Turning footage into an avatar works just as it does with the Digital Twin, so you only need one new recording to upgrade.
Limitations
- Some older custom studio avatars do not support Avatar V, so they cannot use the most consistent engine.
- The Help Centre's own guidance says recreating an older avatar with high-quality footage and the new Look Packs noticeably improves the likeness.
- Looped footage on Avatar III repeats the same gesture cycle, which viewers spot in videos longer than a minute or two.
Pricing
Existing Instant Avatars render at 3 to 20 credits per minute depending on the engine, and new users create Digital Twins instead.
Instant Avatar has earned its place in HeyGen’s history, but if you still use one, take 15 seconds to record a Digital Twin and render it on Avatar V.
Side by Side: Two-Minute Market Update
The same two-minute script about local mortgage rates, the same cloned voice and the same office backdrop for all four avatars. I assessed likeness, lip sync and whether the estate agent's own clients could tell it was AI.
Avatar V Assessment
Avatar V reproduced her head tilt before each statistic and kept her eyeline on the camera for the full two minutes. Lip shapes on "refinance" and "basis points" matched the cloned voice closely. Two of the three clients I showed it to assumed she had filmed it.
Digital Twin (on Avatar IV) Result
The Avatar IV twin captured her hand gestures well and looked convincing for the first minute. Around 1:20, her smile widened slightly on emphasised words. One client said it looked like her "on a different day".
Photo Avatar Outcome
The photo version produced accurate lip sync and natural-looking generated movement. Without her own mannerisms, every client recognised her face, but none recognised her delivery. It still worked as a social media clip with captions from the subtitle generator.
Instant Avatar Outcome
Her 2024 Instant Avatar delivered the script without errors using Avatar III. The looped source footage repeated a hand movement twice in two minutes, which one client noticed.
Winner: Avatar V. It was the only version clients took to be a genuine recording.
Side-by-Side Comparison: 10-Minute Training Module
The same 10-minute induction script for new agents at her brokerage. This test measured how consistently each avatar maintained its identity over time — the clearest point of difference between avatar types.
Avatar V outcome
Teeth, jawline and skin texture at minute nine matched minute one when I paused the frames side by side. The render maintained a continuous take with no visible joins. For a training video her brokerage plans to reuse for a year, that consistency matters more than any single frame.
Digital Twin (on Avatar IV): the result
Avatar IV completed the full 10 minutes in one go and remained watchable. Comparing frames revealed slight changes in the width of her smile and the shape of her cheeks after the sixth minute. Most viewers would not notice, but she did.
Photo Avatar Result
The photo avatar remained consistently framed, but its generated gestures became repetitive by the fourth minute. In a longer lesson, the absence of her natural expressions made the content feel more like narration than teaching.
Results with Instant Avatar
The legacy avatar rendered the module, but the looped footage made its gesture cycle obvious over 10 minutes. Re-recording as a Digital Twin was the clear next step.
Winner: Avatar V. It was the only avatar to show no measurable drift over 10 minutes.
Side by Side: Spanish and Japanese Versions
The same two-minute market update, translated into Spanish and Japanese using her cloned voice. I checked how well the lip movements matched unfamiliar phonemes and whether she still sounded like herself.
Avatar V result
The Japanese version handled rapid sequences of syllables without the mouth movements falling behind the audio. Her cloned voice retained its warmth in both languages, and a bilingual client said the Spanish version sounded as though she were speaking naturally in a second language, rather than being dubbed.
Digital Twin results with Avatar IV
Spanish looked excellent on Avatar IV. In Japanese, the mouth movements for a few rapid syllables lagged a frame or two behind the audio.
Photo Avatar Results
Lip sync remained accurate in both languages, an impressive result from a single photo. The generated movement did not adapt to the rhythm of Japanese speech, so the delivery felt slightly mechanical. For quick localised social posts, it is still more effective than AI dubbing over a static image.
Instant Avatar Output
The legacy avatar worked well in Spanish. In Japanese, the looped source footage made the mismatch between gesture timing and speech rhythm more noticeable.
Winner: Avatar V. It kept her voice, face and timing aligned in a language she does not speak.
Compare pricing
Avatar V and Avatar IV both cost 20 credits per minute, so choosing Avatar IV offers no saving. Photo Avatar on Avatar III is the most affordable way to test scripts. Even so, the Creator plan's 600 credits cover around 30 minutes of Avatar V video featuring your real likeness.
Which Option Is Right for You?
Choose Avatar V if you want viewers to see you, rather than an approximation of you. That includes estate agents sharing weekly market updates, founders sending investor messages, and creators posting daily on TikTok without filming. It also suits L&D teams that need a consistent presenter across a year of modules, with SCORM export on the Business plan and SSO plus audit logs on the Enterprise plan. Pair it with AI voice cloning for the closest match.
Choose a Digital Twin on Avatar IV if you already have a twin and want to use prompts to create variations for a single campaign. Switching to avatar iv for each video lets you experiment with the same asset.
Choose Photo Avatar if your presenter is a mascot, an illustration or a 3D character with no video to record. It is also the quickest way to test an AI spokesperson concept before anyone sits in front of a camera.
Choose Instant Avatar if you already have a library built around one and need to finish a project before re-recording. Plan to switch to a Digital Twin soon afterwards.
Our verdict
A Digital Twin rendered on Avatar V is the best fit for most people: it learns how you move and keeps your facial features consistent, whatever the video length. Avatar IV adds prompt control, Photo Avatar lets you create characters you cannot film, and Instant Avatar keeps older libraries running. Record 15 seconds on your webcam with HeyGen's Free plan and compare the result with your current avatar.
Frequently asked questions
Can I convert my Photo Avatar into a Digital Twin later on?
Yes, but it means recording video, because a Digital Twin trains on footage of you speaking. You can keep your ai photo avatar and add a video look to the same identity. Once a video look exists, Avatar V becomes available for that avatar.
I have an older Instant Avatar. Will I lose my looks if I re-record?
No, your existing avatar and its videos remain in your account. A new Digital Twin creates a separate avatar with its own capacity for 500 looks. HeyGen's Look Packs can then generate outfits and settings tailored to the new twin's persona in just a few taps.
Does a Digital Twin need my voice too?
No, but a cloned voice gives the most natural match, and HeyGen's own guidance recommends one for twins. Public voices suit general business presenters, while designed voices fit stylised characters. Expressive source audio also makes Avatar V gestures livelier.
Can I use my avatar for live sales calls or customer support chats?
Yes, through LiveAvatar, which provides real-time conversational avatars for support, sales and education. LiveAvatar is a separate platform with its own plans, so your HeyGen avatar slots do not transfer automatically. Your Digital Twin stays in your HeyGen account for pre-recorded video.
Which avatar type costs the least to get started?
You can get started with Photo Avatar at no cost on the Free plan, which includes three slots and one photo training session per month in the mobile app. You can also create a Digital Twin at no cost on the Free plan, with one slot included. On paid plans, Avatar III uses 3 credits per minute to render, offering an affordable way to test scripts before a final Avatar V render.
Can I use any of these to create a vertical video for Instagram or TikTok?
Yes, all four render in portrait format for short-form platforms. Avatar Shots adds cinematic Seedance 2.0 opening shots for verified Digital Twins, and you can switch to a presenter shot for the main message. HeyGen's YouTube Shorts workflow handles the 9:16 framing and captions.
Can my avatar deliver the same course in five languages for my global team?
Yes, and your Digital Twin retains your face and cloned voice in every version. Workday produce each training video in 10 to 15 languages and have doubled capacity without increasing staff numbers. The AI video translator also exports subtitles and an editable transcript for review.
Is my likeness protected when I create a Digital Twin?
Yes, creating a Digital Twin requires a live consent recording that matches the training footage, as well as human moderation. The platform is SOC 2 Type II, GDPR and CCPA compliant, and enterprise data is never used to train models. Screen-recorded or AI-generated consent clips will not pass verification.
Greetings! My name is Ayesha Shaheryar. My words have helped millions over the past two years. As a HeyGen expert and a writer, I am here to introduce tips and tricks to edit your next video in no time.








