Image to Video

Talking Photo AI [2026]: 6 Tools, HeyGen $0 to $29

Talking Photo AI [2026]: 6 Tools, HeyGen $0 to $29

Talking photo AI takes one still image plus an audio track or a script and returns a video where the face speaks. In November 2026 the category splits in two. Avatar tools such as HeyGen and Hedra build around a face and a voice. General video models such as Gemini Omni, Kling 4.0 and Seedance 2.5 treat lip sync as one feature among many. HeyGen Video, HeyGen’s newer general video model, generates 5 to 15-second clips from text, images or references with native sound, according to HeyGen’s help centre.

The entry prices are low, and the limits are where the real differences sit. HeyGen’s Free plan costs $0 and allows three watermarked videos a month, each up to one minute. Creator runs $29 a month, or $24 a month billed annually. Hedra has no free API tier, and a new wallet must be funded before it generates anything. Gemini Omni clips top out at 15 seconds. Kling’s own page still lists 1080p and 4K HDR as “coming soon” as of November 13.

This comparison covers six options: HeyGen, Hedra, Kling 4.0, Gemini Omni 1.2 Flash (with Veo 3.1 as the older Google line), OmniHuman 2.0 and Seedance 2.5. Prices and limits come from vendor pages and release notes read on November 18, 2026, and any number that rests on a single source is hedged in the text. If you only need gentle motion and no speech, the open-weight photo-to-video routes are a better starting point than anything below.

What talking photo AI means in November 2026

A talking photo tool needs three inputs: a portrait, a voice source and a way to drive the mouth. The voice source is either an audio file you upload or a script the tool reads aloud with a synthetic or cloned voice. The model then animates the lips, jaw, eyebrows and head so they match the sound. Backgrounds and camera moves are secondary for this job.

Avatar tools put that pipeline front and centre. HeyGen’s Free plan lists Avatar IV, one custom avatar, one voice clone and up to three photo avatars. Hedra started as a talking-character studio and now sells access to its own avatar model next to third-party models. These tools charge by plan or by credit, and their interface assumes you want a speaking face.

General video models work the other way round. Gemini Omni generates and edits video with audio from text, images and video. Kling describes its 4.0 generation as having improved lip sync, but the sources checked do not show a separate talking-photo product. Seedance 2.5 accepts video and audio references and produces synchronized audio. You get cinematic motion and sound in one pass, yet nothing on the vendor pages promises the face will hold a likeness through a minute of speech.

The name “Omni” causes confusion. Google’s Gemini Omni is a video model. OmniHuman 2.0 is a separate portrait-plus-audio model listed on HiAPI’s changelog. A third product called Omni is an enterprise analytics company that has nothing to do with video. This article uses Gemini Omni and OmniHuman by their full names throughout.

Pick the family first, then the tool. If your clip is a presenter reading a script for several minutes, avatar tools fit. If it is a 10 to 15-second cutaway where a photographed grandparent or a mascot says one line, a general model can do it with richer lighting and camera motion. The image-to-video category on this site covers the non-speaking side of the same technology.

One practical point decides many projects before any price comparison. Avatar tools let you reuse the same face across many videos, which matters for a channel that posts weekly. General models generate each clip fresh from the source photo, so consistency depends on you feeding the same photo and a similar prompt every time.

Industry data on creator adoption

Adoption data helps explain why so many vendors chase this niche. Deloitte’s 2024 State of Generative AI in the Enterprise research reported that “31% of surveyed organizations reported using generative AI for video applications, including autonomously generating marketing videos and simulating scenarios for safety training” (Deloitte, 2024). That figure is two years old, so treat it as a floor and not a current reading.

On the creator side, Deloitte’s 2024 Media and Entertainment Outlook said “Around three in five creators reported using generative AI in some capacity, primarily for content ideas, workflow management, and writing captions” (Deloitte, 2024). Its 2024 Digital Media Trends survey found that “18% of Gen Z and millennial respondents reported using generative AI to create images” (Deloitte, 2024). Image generation was a common entry point, and animating the result is a natural next step.

Specs table: six talking photo AI tools side by side

The table below lists only values that appear on vendor pages or in release notes checked on November 18, 2026. Where a vendor page did not state a figure, the row says so instead of guessing. Read the rows as separate facts, not as a like-for-like score, because each vendor publishes different limits.

ToolSpecValue on the vendor page
HeyGen Avatar 6Photo inputA single photo, optionally backed by a short identity-reference recording
HeyGen Avatar 6OutputUp to 2K, no stated length limit
HeyGen VideoClip length5 to 15 seconds, with native sound, up to 2K
HeyGen FreeLimits3 watermarked videos a month, each up to 1 minute
HeyGen CreatorCredits and length600 monthly credits, videos up to 30 minutes
HeyGen ProCredits and output1,000 credits at $49 a month up to 100,000 credits at $4,300, with 4K output
HeyGen BusinessSeats$149 a month for the first seat, $20 for each extra seat, 1,500 workspace credits
HedraModel catalogMore than 110 models through one API key
HedraBillingPrepaid USD wallet that starts at $0 and must be funded, no API free tier
Kling 4.0Lip syncImproved lip sync listed, but no dedicated talking-photo product
KlingResolutionNot listed
Gemini OmniClip lengthUp to 15 seconds, text, image and video in, video with audio out
OmniHuman 2.0Price$0.10 per output second, reported rate on HiAPI
Seedance 2.5Clip length and inputsUp to 30 seconds, video and audio references, synchronized audio

Two gaps stand out. None of the pages checked publishes a lip-sync accuracy score, so you cannot rank these tools on sync quality from vendor data alone. And only HeyGen and Hedra state how billing works in plain terms, while Kling’s resolution claims and Dreamina’s plan prices contradict each other across sources.

The spread in clip length is the number to anchor on. A 15-second cap forces stitching for anything longer than two or three sentences. A 30-second cap covers a short ad read. HeyGen’s 30-minute Creator ceiling and Avatar 6’s lack of a stated limit cover explainers and lessons, which is why the avatar family wins for talking heads and the general models win for cutaways.

HeyGen: Avatar 6, photo avatars and the Free plan

HeyGen is the most direct answer to “make this photo talk.” Its homepage describes Avatar 6 as its most capable avatar model, able to reproduce a person’s appearance, voice, movement, emotion, gestures and expressions. The avatar can be built from a single photo, with an optional short identity-reference recording, and the page says output can reach 2K with no stated length limit.

HeyGen scheduled a live introduction of Avatar 6 for November 18, 2026, the day this article went up. The event description mentions a workflow of about 30 seconds of video, several photos and one voice recording. One research result also reports that Creator’s Avatar IV and V listings were replaced by Avatar 6 Turbo and Avatar Lipsync, and that Pro gained Avatar 6 Pro. Treat that lineup change as reported, not confirmed, until HeyGen’s pricing page shows it.

The separate HeyGen Video model makes 5 to 15-second clips from text, images or references. HeyGen’s help centre lists image-to-video at $0.01 a second at 480p, $0.018 at 768p, $0.04 at 1080p and $0.05 at 2K. Those are launch-promotion rates that reportedly run through November 30, with a standard 768p rate of $0.03 a second listed from December 1. Outputs longer than 15 seconds cost more during the promotion, per the same help page.

HeyGen’s own material does not say whether HeyGen Video replaces Avatar 6 or sits beside it. For a talking photo, Avatar 6 is the product built for the job, while HeyGen Video suits a short cinematic cutaway around the face. Creator-plan users can reach HeyGen Video, Seedance 2.0 Mini, Veo, Wan, Gemini, Hailuo and LTX inside Creative Studio, while Seedance 2 and Seedance 2.5 stay locked on that plan.

HeyGen plan limits for talking photo work

The Free plan is a genuine trial. It costs $0, needs no credit card and includes three watermarked videos a month, each up to one minute, plus Avatar IV, one custom avatar, one voice clone and up to three photo avatars. That is enough to test whether your portrait survives the animation before you pay anything.

Creator costs $29 a month, or $24 a month billed annually, with 600 monthly credits and limited Avatar V access. Pro starts at $49 a month, or $41 billed annually, for 1,000 credits and adds 4K output and third-party video models. Business costs $149 a month, or $119 billed annually, for the first seat and $20 for each additional seat. HeyGen phased out its Team plan for new purchases in January 2026. Its API runs separately on pay-as-you-go credits with no subscription required.

HeyGen also made its Voice Turbo API mode free for a limited promotional period on shared capacity. After that, the listed prices are $4 per 1 million characters for Turbo and $15 per 1 million characters for standard HeyGen Voice. If your talking photo reads a long script, the voice line item matters more than the video line item.

Hedra: a talking-avatar studio turned model marketplace

Hedra changed shape in 2026. On August 4 it announced that its models were available through an API, native SDKs, a CLI and MCP. On October 6 it published the Hedra Developer API as one offer combining models, SDKs, MCP, CLI and a console. Hedra says the API exposes more than 110 models through one API key.

For talking photo work, that means Hedra Avatar sits in a catalog next to Veo 3.1, Kling V3, Seedance 2.5, MiniMax H3 and Wan 3.0. The sources checked do not confirm a dedicated talking photo model beyond Hedra Avatar, so check Hedra’s docs for the current model ID before building on it.

Billing works unlike HeyGen’s. API and CLI use draws from a prepaid USD wallet, not a subscription, and Hedra’s FAQ states that a new wallet starts at $0.00 and refuses generations until it is funded. Estimates and file uploads are free. Hedra’s quickstart shows a 6-second MiniMax H3 clip at 768p costing $0.36, which works out to $0.06 a second. That example uses a third-party model, so it shows the wallet mechanics and not the price of an avatar clip.

The API version matters if you already have scripts. Hedra moved from the legacy v2 structure toward v3 in October and November 2026, with model-specific endpoints, a separate prepaid wallet and public model IDs. The migration guide is dated October 23 and was updated November 14, and the legacy v2 interface had no announced shutdown date at that point. Hedra also described a Model Factory effort on November 12 and 13, which the pages checked present as internal research and engineering and not as a product you can buy.

Hedra fits developers and small studios that want to swap models without rewriting integrations. It fits casual users less well, since the pages checked lead with the API, the CLI and the console. Anyone who wants to drag a photo into a browser and hit render will likely find HeyGen faster.

Kling 4.0: better lip sync, no talking-photo product

Kling’s newest model is Kling 4.0, and its official site lists Kling 4.0 Flash and Kling 3.0 Omni beside it. The sources checked confirm improved lip sync as a Kling 4.0 capability. They do not show a dedicated talking-photo feature or a separately named product comparable to HeyGen’s avatars. You feed Kling a still through its image-to-video workflow, describe the speech in the prompt or attach audio, then judge the result.

The launch timeline is unclear. The official results checked do not confirm the exact public-launch date or whether the full rollout completed. Kling 3.0 now counts as the predecessor, not the newest model.

According to Kling’s announcement, the 4.0 generation brings native 30-second generation, 10-bit HDR, Omni Reference, multiple keyframes and upgraded motion. That single-source list matters for talking photos because 30 seconds is twice the Gemini Omni cap. Kling VIDEO 3.0 is described as supporting image and video creation, native audio, multi-shot storytelling, flexible duration and text rendering.

Pricing is the weak spot. The sources checked do not give a complete, confirmed Kling 4.0 price list, and no plan, billing period or credit allowance is stated. Confirm what resolution your plan actually exports before you commit. Developers can start from Kling’s open platform documentation.

Kling suits creators who want the face to move inside a scene, such as a character delivering a line while the camera pushes in. It suits long monologues less well, because the sources checked show no avatar-style identity lock across a long script.

Gemini Omni 1.2 Flash and Veo 3.1

Google’s Gemini Omni generates and edits video with audio from text, images and video, with clips up to 15 seconds. It is available through the Gemini API and Google Cloud, and reports add Flow, Google Vids and Google Ads. Google said at launch that it was rolling out to the Gemini app, Flow and YouTube Shorts. Google’s own Gemini Omni experts roundtable answers common questions about the model.

The version history is short. Google announced Gemini Omni Flash on May 19, 2026, and said it was available to Google AI Plus, Pro and Ultra subscribers through Gemini and Flow, and at no cost through YouTube Shorts and YouTube Create. One research result reports that Gemini Omni 1.2 Flash became generally available on November 5, 2026. The same result reports a 720p video price near $0.07 a second, about 30% lower than Omni 1.1 Flash. Those figures come from a single source, so confirm them on Google’s pricing page.

At that reported rate, a full 15-second clip at 720p is cheap, though that is my own estimate and not a Google quote. The clip cap is the real constraint. A photo that says two sentences fits. A photo that reads a one-minute script has to be stitched from four clips, and the face may drift between them.

Veo is the older line. Google DeepMind’s Veo page still lists Veo 3.1 as the current Veo model. No source checked shows a newer Veo version. Hedra and HeyGen both still list Veo 3.1 in their catalogs, so the model remains reachable through third parties either way.

Gemini Omni fits creators already inside Google’s tools, especially Shorts makers, since access through YouTube Shorts and YouTube Create was described as free at launch. It fits worse for long-form talking heads, where its 15-second cap forces stitching.

OmniHuman 2.0 and Seedance 2.5

OmniHuman 2.0 is the closest thing to a pure talking-photo model among the six. HiAPI’s changelog lists it with a reported rate of $0.10 per output second. A single third-party listing dated November 12, 2026 says it takes one portrait and 4 to 300 seconds of audio and returns a lip-synced video up to five minutes long, with lip, facial-expression and motion synchronization. By that rate, a 60-second clip costs about $6.00, again simple arithmetic.

The caveat is provenance. The pages checked include no official statement from a model maker, and HiAPI is a reseller. Run one short clip before you budget a series, and keep the source audio so you can switch tools if the price or availability changes.

Seedance 2.5 comes from ByteDance and is reported to have launched July 31, 2026 as the successor to Seedance 2.0. It supports clips up to 30 seconds and generation that uses video and audio references. Seedance 2.0 was announced on February 12, 2026, according to a dated third-party developer guide, and is described as accepting text, image, video and audio inputs with outputs up to 15 seconds and resolutions including 4K. A cheaper, faster Seedance 2.5 Flash variant limited to 720p was reported in mid-November, but treat that as a rumor until ByteDance confirms it.

Runway Dev publishes API pricing for the family. Its Seedance 2.0 page lists $0.14 a second at 480p, $0.30 at 720p, $0.40 at 1080p and $1.50 at 4K. Runway also runs a Seedance 2.5 API page with access, pricing and code, and one research result reports $0.30 a second at 720p, $0.68 at 1080p and $1.10 at 4K with a $0.80 minimum per generation. Dreamina’s own consumer prices conflict across sources, from a free tier and about $8.80 a month to $15 a month for the paid entry plan, so check Dreamina directly.

If the goal is only a gentle photo-to-video effect with music and no speech, a photo video maker with a music bed costs far less than a per-second model. Seedance earns its price when you need sound, motion and a second person in the frame.

Pricing table: what each route costs

The table lists only prices that two or more sources confirmed. Consumer plans bill monthly, while API and per-second prices bill by output. Because the units differ, compare them through the worked examples after the table.

RoutePriceWhat it buys
HeyGen Free$03 watermarked videos a month, up to 1 minute each
HeyGen Creator$29 monthly or $24 billed annually600 credits a month, limited Avatar V access
HeyGen Pro$49 monthly or $41 billed annually1,000 credits, 4K output, third-party video models
HeyGen Business$149 monthly or $119 billed annually, plus $20 a seat1,500 workspace credits, collaboration, SSO
HeyGen APIPay as you goCredits with no subscription required
Hedra APIPrepaid wallet from $0No free tier, wallet must be funded first
Seedance 2.0 on Runway Dev, 480p$0.14 a secondClips up to 15 seconds
Seedance 2.0 on Runway Dev, 720p$0.30 a secondClips up to 15 seconds
Seedance 2.0 on Runway Dev, 1080p$0.40 a secondClips up to 15 seconds
Seedance 2.0 on Runway Dev, 4K$1.50 a secondClips up to 15 seconds
OmniHuman 2.0 on HiAPI$0.10 per output second (reported)Portrait plus audio, lip-synced output

Worked examples, using the rates above. A 15-second Seedance 2.0 clip on Runway Dev costs $2.10 at 480p, $4.50 at 720p and $6.00 at 1080p, with 4K priced at $1.50 a second. For a smaller example, a 6-second MiniMax H3 clip on Hedra at 768p costs $0.36, though that is a different model from Hedra Avatar.

For a creator who posts four one-minute talking photos a month, the arithmetic favors HeyGen Creator at $29 over any per-second route, since four clips at $6 each would already cost $24 on OmniHuman before retries. Retries are the hidden cost. A per-second model bills you every time the face drifts, while a plan with credits at least caps the monthly bill.

Kling’s yuan promotions and Dreamina’s conflicting consumer prices are left out of the table on purpose. Their sources disagreed on tier, billing period or region, and the conservative reading is that no confirmed US consumer price list exists for those two on the date of checking.

Benchmarks and failure modes: what the published numbers show

No independent lip-sync benchmark appeared in the sources checked, so this section lists the published numbers that do exist and says what each one measures. That keeps the comparison honest about what is known and what is vendor marketing.

SourceNumberWhat it measures
Hedra research post, November 12, 20261,223 held-out clipsGlitch detection by a fine-tuned Qwen3-VL-8B model at a 10% false-positive rate
Hedra developer quickstart$0.36 for 6 seconds at 768pCost of one MiniMax H3 clip through the Hedra API
HeyGen help centre$0.01 to $0.05 a secondHeyGen Video image-to-video price from 480p to 2K, promotional
Runway Dev pricing$0.14 to $1.50 a secondSeedance 2.0 price from 480p to 4K
Google Gemini Omni materials15 secondsMaximum clip length per generation

The Hedra post is the most useful for creators because it names real failure modes. It was written by Harrison Wang and uses clips and annotations from the Physion-Eval dataset, with a detector fine-tuned on human-verified annotations of Hedra-generated videos. One annotation reads “At 4.2 s, the yellow sphere in the hand suddenly turns blue without any change in camera angle.” Another flags Sora 2 at 1.2 seconds with “The water suddenly starts swirling without any external force.” A third flags Veo 3.1 Fast at 3.8 seconds with “The nut suddenly deforms without any external force, then is completely flattened.”

Those are object glitches, not mouth glitches, but they show how generated video fails: things change shape or color without cause. In a talking photo, the equivalent is a necklace that vanishes, earrings that change, a mouth that stretches or a face that shifts toward a different person partway through. The post also shows a detector holding a 0.96 confidence score on one example even as an elephant’s trunk shortened inside an unchanged bounding box, which tells you automated checks can miss slow drift.

Use three checks on every clip before you publish. First, scrub the video frame by frame near the start, the middle and the final second, comparing teeth, jewelry and the hairline against the source photo. Second, play it muted, then with sound, and watch whether the lips close on “m”, “b” and “p” sounds. Third, compare the first and last frames side by side for identity drift. Fixing a bad clip costs one more generation, so check before the render goes into your edit.

One more number deserves caution. HeyGen’s own pages claim a 40% increase in video watch time and a 25% boost in course completion rates. Those are company claims on its marketing pages. The lines checked do not name the customers or the method, and they describe avatar video in general, not talking photos.

Real-world examples of talking and AI-generated video

Public case studies for talking photos specifically are thin, so the examples below mix showcase facts from the vendors’ own pages with costed scenarios. Each scenario is labeled as arithmetic, and none claims a result a vendor did not publish.

  1. Virtual idol MV on Kling. Kling’s showcase page credits Kling AI, Chaocreate Wu Zhiqi and Warner Music China with creating the virtual idol Wu Aihua and her debut MV, with all visuals generated by Kling AI, and reports over 11.2 million views worldwide. It shows a generated character carrying a full music video.
  2. Viral short film on Kling. The same page lists an AI short film that went viral during the 2026 Qingming Festival, created entirely with Kling AI, with over 100 million views across platforms. An all-AI short can clearly reach a wide audience.
  3. Cannes Lions-recognized ads. Kling’s page says an ad by Argentine studio Purga Films, made entirely with Kling AI, won Bronze in the Classic Film category at the 2026 Cannes Lions festival. It also lists an April Fools’ Day ad for The RealReal, by director Sebastian Strasser and his studio Lipstick, that won one Silver and one Bronze at the same festival.
  4. Feature-length output. Kling lists a feature co-produced by studio Fountain 0 and Kling AI, directed by Ash Koosha, whose 135 minutes of visuals were generated entirely with Kling AI. It proves long projects are possible, though it says nothing about talking-photo accuracy.
  5. HeyGen training video. HeyGen’s marketing claims a 40% increase in video watch time and a 25% boost in course completion rates for customers using its avatar video. A course creator animating an instructor’s portrait is the closest match to a talking photo here, but the customer is unnamed.
  6. Costed scenario: a one-minute memorial message. At OmniHuman 2.0’s reported $0.10 per output second, a family portrait speaking a 60-second script costs about $6.00, and three takes cost about $18.00. The same project on HeyGen Free is $0 for up to three watermarked videos of one minute each, so the Free plan covers a no-cost trial of the same idea.

The pattern across these examples: the largest audiences and awards came from teams that treated the model as one tool in a produced piece, with writing, music and editing around it. A talking photo works the same way. The model animates the face, and the script, the audio and the cut do the persuading.

Use-case recommendations: which tool for which job

Pick by job, not by brand. These six matches follow from the limits in the tables above.

  • A presenter for a course or explainer (1 to 30 minutes): HeyGen Creator at $29 a month. Its videos run up to 30 minutes and Avatar 6 has no stated length limit.
  • A free test of whether your photo works: HeyGen Free at $0, with three watermarked clips of up to one minute.
  • A developer building a talking-photo feature: Hedra, for 110+ models behind one key and a prepaid wallet. Budget for the v3 API and fund the wallet first.
  • A one-line cutaway inside a short film: Gemini Omni 1.2 Flash for up to 15 seconds with audio, or Kling 4.0 if you want the face inside a moving camera shot.
  • A long speech from a single portrait on a per-second budget: OmniHuman 2.0, reported at $0.10 a second, after a one-clip trial.
  • A cinematic scene with a second subject and sound: Seedance 2.5 at up to 30 seconds, priced through a host such as Runway Dev.

Prepare the inputs before you spend credits. For the photo, use a front-facing portrait with even light and a relaxed, closed mouth, and crop so the face fills at least a third of the frame. Skip images with glare on glasses, a hand near the chin or a busy hairline, since those are the details a model tends to invent or lose.

For the audio, record in a quiet room and keep sentences short. Short sentences give the model clean pauses to reset the mouth and reduce the chance of drift. Normalize loudness in your editor first, because a very quiet file can read as silence and a clipped file can cause jittery lip motion. If a tool offers a built-in voice, still render your own track once, so you can reuse it elsewhere.

If your channel needs a branded open before the talking photo starts, a short animated logo intro keeps the first seconds consistent while the AI clip handles the face. One place to look is an animated logo intro theme.

Render settings

Vendors publish few frame-rate figures, so match the output to your edit timeline instead of trusting a default. The table gives starting values drawn from the limits above.

SettingStarting valueWhy
Resolution1080p, up to 2K where offeredHeyGen Avatar 6 and HeyGen Video both list output up to 2K
Frame rateMatch your timelineThe pages checked did not state a frame rate, so set it in your editor
Length per clip15 seconds for Gemini Omni, 30 seconds for Seedance 2.5Published caps per generation
Export formatMP4 with H.264 video and AAC audioPlays everywhere and re-edits cleanly
Source photoFront-facing, even light, mouth closedGives the model a clean neutral mouth to start from

From the Vipid archive: pair a talking photo with one of the original intro themes, which list resolution, length and frame rate so you can match your clip settings.

Migration guide: moving a talking photo workflow between tools

Tools in this niche change fast, so build the workflow around files you own. Steps one to three apply to any move, and steps four to six cover the changes that took effect in 2026.

  1. Keep the master assets. Store the original portrait, the script and the audio file in one folder, named by date. Every tool in this comparison accepts a photo plus audio or text, so the assets move even if the tool does not.
  2. Export audio at the source. If you used a tool’s built-in voice, render the speech to a WAV or MP3 once and reuse it. That fixes the timing across tools, so differences in the output come from the model.
  3. Run the same clip in both tools. Use a 10-second excerpt first. Compare mouth closure, identity and background stability before you move a whole series.
  4. HeyGen: if your plan listed Avatar IV or V, expect Avatar 6 options. One research result reports Avatar 6 Turbo and Avatar Lipsync on Creator and Avatar 6 Pro on Pro, so open the plan page before you assume a legacy avatar still exists. The Team plan has been closed to new buyers since January 2026.
  5. Hedra: if you built on v2, read the migration guide. It describes model-specific endpoints, a separate prepaid wallet and public model IDs. The v2 interface had no announced shutdown, so you can migrate on your schedule, but new work should target v3.
  6. Google: if you call a Gemini model in the API, pin your model version in code in case requests are later moved to a newer version at a new rate. Watch the invoice for the first week.

When you finally deliver, convert the output to a clean MP4 for editing. This standard FFmpeg command re-encodes to H.264 with AAC audio and moves the metadata to the front so previews load quickly.

ffmpeg -i talking-photo.mov -c:v libx264 -crf 18 -preset slow -c:a aac -b:a 192k -movflags +faststart talking-photo.mp4

Keep the original file too. If a vendor retires a plan or changes a model, the source files let you rebuild the clip elsewhere without paying twice. HeyGen’s own example: it phased out its Team plan for new purchases in January 2026, so any workflow that depended on buying one had to move.

Pros and cons of each talking photo AI

HeyGen. Pros: a real free tier, Avatar 6 built for single-photo avatars, clips up to 30 minutes on Creator, plans from $29 a month. Cons: Seedance 2.5 and Seedance 2 are locked on Creator, the Avatar 6 plan mapping is still settling, and the Free plan watermarks every video.

Hedra. Pros: more than 110 models through one key, SDKs, CLI and MCP, and a prepaid wallet with free estimates. Cons: no free API tier, a v2 to v3 migration, and a consumer story that now trails its developer story.

Kling 4.0. Pros: improved lip sync, a reported 30-second native generation length and a deep showcase of finished projects. Cons: no dedicated talking-photo product, contradictory resolution claims and prices that are hard to pin down.

Gemini Omni 1.2 Flash. Pros: audio built in, Google distribution through the Gemini API, Flow and YouTube tools, and a reported price cut. Cons: 15-second clips, stitched for anything longer.

OmniHuman 2.0. Pros: purpose-built for a portrait plus audio, reported clips up to five minutes and a flat $0.10 a second. Cons: sold through a reseller, single-source specs and per-second billing that charges for every retry.

Seedance 2.5. Pros: 30-second clips, video and audio references and wide availability through Dreamina, Runway Dev and others. Cons: conflicting consumer prices, a reported Flash variant that is not confirmed and no talking-photo focus.

Verdict: which talking photo AI to use

The takeaway: start with HeyGen. It is the only option here that pairs a purpose-built single-photo avatar model (Avatar 6), a $0 plan with three watermarked one-minute videos, and a $29 plan with 600 credits and 30-minute videos. You can test your portrait for free, then step up to Creator at $29 a month if the result holds.

Move to Hedra if you build software or want to switch among 110+ models without rewriting code. Move to OmniHuman 2.0 if you need a long speech from one portrait, with the $6.00 per minute reported rate and the provenance caveat in mind. Choose Gemini Omni 1.2 Flash, Kling 4.0 or Seedance 2.5 when the face is one element in a cinematic cutaway of 15 to 30 seconds.

By the numbers: HeyGen offers a $0 trial, Hedra needs a funded wallet, Gemini Omni caps at 15 seconds, Seedance 2.5 at 30 seconds, and Kling still marks 1080p and 4K HDR as “coming soon”. Plan limits changed through 2026, so reread the pricing page on the day you subscribe. Every model name in this article can change within weeks, since HeyGen, Kling and ByteDance each shipped a new model between July and November.

Questions creators ask

Is there a free talking photo AI without a watermark? Not confirmed among the six checked. HeyGen’s Free plan costs $0 but adds a watermark to its three monthly videos. Hedra has no free API tier, and the other tools bill per second or through paid plans. Watermark rules change by plan, so read the plan page before you pay.

How long can a talking photo video be? It depends on the tool. HeyGen Free allows up to one minute per video, Creator up to 30 minutes, Gemini Omni up to 15 seconds per clip and Seedance 2.5 up to 30 seconds. Longer scripts on the clip-capped tools need stitching.

What is the difference between Gemini Omni and OmniHuman? Gemini Omni is Google’s general video model that generates and edits video with audio. OmniHuman 2.0 is a separate model that takes one portrait plus audio and returns a lip-synced video, and HiAPI lists it at a reported $0.10 per output second. The names look alike, but the companies and uses differ.

Does Kling 4.0 have a talking photo feature? Kling 4.0 lists improved lip sync, but the sources checked show no dedicated talking-photo product. You animate a still through its image-to-video workflow and prompt the speech, so expect to run several takes.

Is Veo still Google’s newest video model? Google DeepMind’s Veo page lists Veo 3.1 as the current Veo model. Gemini Omni is the separate video model family Google now promotes for video with audio, and it accepts images, so it is the one to test for photo-driven clips.

What photo works best for a talking photo? Use a front-facing portrait with even light and a closed, relaxed mouth. Avoid heavy glare on glasses, hands near the face and busy hairlines. A clean source gives every model here the best chance of holding identity through the clip.

Can I use someone else’s photo? Only with permission, and only where the platform’s rules and local law allow it. Avatar services ask you to confirm rights, and video platforms have their own synthetic-media rules, so read them before you publish a family portrait or a public figure.

More from the Vipid blog

Make it with Vipid

Lena Marsh

AI Video Editor · Vipid Team

Lena Marsh is the AI Video Editor at Vipid, covering AI video generators and the platform and software news that changes how creators render, upload and stream. She judges every model by its output — clip length, resolution, frame rate, watermark and cost — rather than by announcements, and runs Vipid’s news radar across the official YouTube, Apple, Adobe, Blackmagic, OBS, OpenAI and Google sources.

All articles by Lena Marsh ›   How we check tools ›

Keep reading

Unlimited Image to Video AI [2026]: 0 of 5 Confirmed
Image to Video

Unlimited Image to Video AI [2026]: 0 of 5 Confirmed

Lena Marsh·13 Nov 2026
Best Image to Video AI 2026: 7 Tools, 30s Clips
Image to Video

Best Image to Video AI 2026: 7 Tools, 30s Clips

Lena Marsh·30 Oct 2026
Free AI Video Generator [2026]: 6 Tiers, 50 to 800 Credits
AI Video Generators

Free AI Video Generator [2026]: 6 Tiers, 50 to 800 Credits

Lena Marsh·16 Nov 2026

Blog

Intros & Templates
Titles & Motion
Streaming
Editing & Export
AI Video

Vipid

About Vipid
Themes
News archive 2011–2016
Vipid Team

Editorial

Editorial policy
How we check tools
Corrections

Legal

Privacy policy
Terms of service
Making video intros since 2010.
The Vipid blog covers the tools behind them.