Eleven AI lip-sync generators are worth your attention in October 2026, and they split into three unrelated jobs. Hedra, D-ID and Captions animate a still photo; Sync Labs, VEED and Akool retime real footage to new audio; HeyGen, Synthesia and Tavus run digital humans at scale; LatentSync and MuseTalk are the two open-source models you can actually ship. Buying the wrong category is the most expensive mistake in this market.
Why lip-sync got interesting again in 2026
Three things changed. The open-source models crossed a quality threshold: ByteDance’s LatentSync 1.6 trains at 512x512, and Tencent Music’s MuseTalk 1.5 holds 30fps on a single Tesla V100, which makes self-hosting an option rather than a hobby. Pricing transparency split apart: one vendor now publishes per-model, per-second rates so you can estimate a job before running it, while most publish credits and hide the conversion.
And least discussed: the EU AI Act’s transparency obligations are now live.
Article 50(4) of the EU AI Act, in force since 2 August 2026: deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake “shall disclose that the content has been artificially generated or manipulated.”
That is not a future obligation. Providers whose systems were on the market before 2 August 2026 must meet the machine-readable marking duty in Article 50(2) by 2 December 2026: seven weeks from publication.
Three different kinds of lip-sync tool
Prices and specifications are from vendor documentation, model repositories and licence files checked on 16 October 2026. Prices are in USD unless stated; Akool’s plans are billed annually, and VEED displays regional pricing. Quality claims come from published sources rather than hands-on tests.
- Talking-photo avatars take one still image plus an audio file. There is no original video to preserve.
- Real-video lip-sync retiming takes footage you already have and redraws the mouth region. Performance, lighting and camera work are kept.
- Digital-human platforms take a script, brand kit and avatar library and generate the whole video.
Choose the category that matches your input: a still photo, existing footage or a script.
The 11 tools at a glance
| Tool | Entry price (16 Oct 2026) | Max clip | Resolution | Languages | API | Licence note |
|---|---|---|---|---|---|---|
| Hedra | $20/mo (2,000 cr.) | Not published | Not published | None listed | Yes | Commercial from $20/mo |
| D-ID | $4.70/mo Lite | 30 min; 5 min translated | Not published | 30+ | All tiers | Personal only on Lite; commercial from Pro ($16/mo) |
| Captions | Free; Max $24.99/mo | Not published | Not published | 100+ captions | None found | Credit-to-second rate not published |
| Sync Labs | $5/mo + $0.05/sec | 1–30 min by tier | 512×512; 4K on sync-3 | 29 dubbed; any language if you bring audio | Python + TS SDKs | No watermark from Creator up |
| VEED | Region-priced | Not published | Not published | 50+ | Lip Sync API listed | No watermark from Creator |
| Akool | $12/seat/mo annual | 15–30 min | 1080p → 4K | 155+ | Yes | Personal licence on Starter and Pro |
| HeyGen | Free; Creator $29/mo | 1–60 min; no max on Ent. | 1080p free, 4K paid | 30+ free, 175 paid | Lip Sync + Translation | Photo avatars capped at 3 on free |
| Synthesia | Free; Starter $29/mo | 3 min per video | 1920×1080 | 130+; 135 locales on Ent. | Dubbing API | Lip sync unavailable on Basic |
| Tavus | Free (25 min); Starter $59/mo | 10–100 min video gen | 1080p, 24kHz | 30+ | Yes | White-labelled replica consent |
| LatentSync 1.6 | Free, self-hosted | No API limit | 512×512; 18GB VRAM | Multilingual | Self-host | Verify repo licence first |
| MuseTalk 1.5 | Free, self-hosted | No API limit | 256×256 face region | ZH, EN, JA | Self-host | MIT code; weights commercial OK |
“Not published” means the cited vendor page does not state it.
Talking photos: Hedra, D-ID, Captions
Hedra is the quality pick for a single still image; D-ID is the safer institutional bet; Captions wins if the work happens on a phone. The technical difference is not mouth accuracy: all three map audio to phonemes and back. It is how much head, shoulder and expression movement surrounds the mouth.
Hedra publishes the least detail of the three. Basic is $20/month with 2,000 credits, Pro $50 with 7,200, Ultra $100 with 18,000, Teams $75 with 14,400. Commercial use is listed on every paid tier, unusually generous for an entry plan. What is missing is exactly what you need to budget: no published resolution, no maximum duration, no credit-to-second rate, no named language list.
D-ID is the opposite: verbose and awkward. Lite is $4.70/month annual, Pro $16, Advanced $108. Video runs to 30 minutes on every plan, but translated video caps at 5 minutes on Lite through Advanced and only opens to 30 minutes on Enterprise: the most important number on the page, and almost nobody mentions it.
Second trap: the licence is personal use on Trial and Lite, with commercial use starting at Pro. The $4.70 plan that looks like the obvious cheap option cannot legally produce a client deliverable. Watermarking follows a ladder (full-screen on trial, D-ID watermark on Lite, AI watermark on Pro, custom logo only from Advanced) and billing rounds up to the nearest 15 seconds, with unused minutes void.
Captions is a different product with a similar name. Max is $24.99/month with 500 credits, Frontier $69.99 with 1,400, including AI Twin, AI Lipdub, AI Translate and AI Voice Clone.
“Captions in 100+ languages” is a captioning claim, not a translation claim: separate those before planning localisation. The page also states that “features and prices reflect iOS plans only.”
Real-video retiming: Sync Labs, VEED, Akool
For retiming footage you already own, Sync Labs is the only vendor here that lets you cost the job before you run it.
Sync Labs is the team behind Wav2Lip. Their pricing page publishes metered rates per model per second at 25 fps: the clearest cost model in the category:
| Sync Labs model | Per second (25 fps) | Face resolution | Documented use |
|---|---|---|---|
| lipsync-1.9 | $0.025 | : | Maximum speed on simple videos |
| lipsync-2 | $0.05 | 512×512 | General purpose, preserves speaking style |
| lipsync-2-pro | $0.08325 | 512×512 | Premium quality, enhanced facial detail |
| sync-3 | $0.1334 | 4K native | Professional grade, close-ups and obstructions |
| react-1 | $0.16675 | : | Expressive, emotional, short clips |
Tiers run Hobbyist $5/month, Creator $19, Growth $49 and Scale $249, with rates falling to $0.04/sec on Scale. Dubbing is billed separately at $6.00 per minute of audio on entry tiers, $4.80 on Scale.
Two things make this different. Max clip length scales from 1 to 30 minutes across four tiers, and the free trial gives three generations a month at up to 20 seconds each with no card. More important, the docs say you can “provide your own translated audio in any language for lipsync”: Sync Labs is not gated on a language list, which matters more than any language count if your audio code-switches mid-sentence.
Their April 2026 post announcing sync-3 claims 4K at 60fps and robustness on sharp angles, low light and multi-speaker shots. That is a vendor claim, roughly six months old, not independently benchmarked.
VEED is the broad-suite option that now ships a dedicated Lip Sync API, listed alongside Fabric 1.0, Background Removal and a Live Avatar API in the site footer. The API documentation requires a login; its limits are not confirmed here. The pricing page publishes translation to 50+ languages from Pro and no watermark from Creator up. On 16 October 2026 its regional prices were: ₹490, ₹1,068 and ₹1,722 per user per month billed yearly for Creator, Pro and Studio.
Akool has the widest published language count here at 155+ for video translation, plus an explicit Lip-Sync feature alongside face swap, multi-face detection and re-age. Annual-billed plans start at $12/seat/month for Starter with 200 credits and $21 for Pro with 600 credits. But note the direction of travel: the plan cards lead with Seedance and Kling video generation rather than avatars, and both Starter and Pro are personal licence, with commercial use on Business.
Digital humans: HeyGen, Synthesia, Tavus
HeyGen wins on language reach, Synthesia on governance, and Tavus is the only one built for live conversation rather than pre-recorded scripts.
HeyGen publishes an unusually complete feature matrix: three free videos a month at one minute each, unlimited photo avatars and 175 languages and dialects from $29, scaling through Pro credit tiers to Business at $149 plus $20 per seat.
The API is where HeyGen is strongest, and its documentation is clearer than anything else here. The Lip Sync and Video Translation APIs sit on one engine with two modes.
Speed is the default: faster, adequate translation, standard lip-sync quality, suited to faces with little movement and quick drafts. Precision is slower with high lip-sync quality, and HeyGen’s own docs point it at faces with significant movement, side angles or occlusions, and at final delivery: a documented trade-off, not a marketing adjective.
Also worth knowing: Video Translation clones the original speaker’s voice by default, with an Enterprise option to swap in a preset stock voice. HeyGen also defaults enable_dynamic_duration to true, letting output duration shift to natural speech pacing while still billing on source duration.
Synthesia is the governance play: SAML/SSO, audit logs, role and content-access controls, SCORM export, brand kits and moderation tiers, on top of Basic $0, Starter $29/month and Pro $89/month.
The credit maths is where Synthesia is most often misreported. Per their pricing page: video is 100 credits per minute, dubbing 76 credits per minute, the first 10 video minutes included on every plan, and enabling lip-sync “consumes twice as many credits.”
Video length caps at 3 minutes, and lip sync is unavailable on Basic. A widely circulated round-up dated September 2026 claims 240 credits per minute for lip-synced dubbing and 30-minute videos. Both are wrong against the vendor’s own page.
The dubbing docs add substance. Uploaded source accepts .mp4, .webm and .mov, up to 5GB or 2.5 hours, up to 4K (3840×2160), 23.98–60 fps; a YouTube link has no duration limit.
You choose the pacing model explicitly: Adaptive adjusts video speed to fit the translation, recommended for instruction; Original fixes playback speed and adjusts only the voiceover. Self-serve plans get a named list of roughly 130 target languages, Enterprise 135 with regional locales. If your language is not on the list, it does not exist.
Tavus is a different shape: real-time conversational video, not scripted generation. Starter is $59/month plus usage, Growth $397, with 1080p at 24kHz across 30+ languages and overage of $0.37 per conversational minute and $1.00 per generated minute.
Replicas are built from a single image or two minutes of training video, and Tavus documents a white-labelled replica consent workflow: the only consent workflow I found explicitly documented in a pricing table.
The open-source reality check
Only two open-source models are worth planning around today, and three of the most-cited ones cannot legally ship. This is the section every other round-up skips.
The two that work:
LatentSync 1.6 from ByteDance has the strongest accuracy-to-effort ratio available. Version 1.6 shipped on 11 June 2025 and trains at 512×512 specifically to fix the blurriness in earlier versions. Inference needs 8GB of VRAM at version 1.5 and 18GB at 1.6. Two parameters control everything: inference_steps from 20 to 50, and guidance_scale from 1.0 to 3.0, where the docs state plainly that a higher value “improves lip-sync accuracy but may cause the video distortion or jitter.” That is the sync-versus-realism dial, exposed. The paper’s StableSyncNet reaches 94% accuracy on the HDTF test set, up from 91%.
MuseTalk 1.5 from Tencent Music’s Lyra Lab is the real-time option: 30fps or better on a Tesla V100 at a 256×256 face region, supporting Chinese, English and Japanese. MIT-licensed code, with weights the maintainers state are “available for any purpose, even commercially.” A consumer RTX 3050 Ti with 4GB of VRAM takes roughly five minutes for an eight-second clip in fp16; training stage two wants about 85GB per GPU. The authors document their own limits: the face region “has not yet reached the theoretical resolution bound,” identity preservation drops on moustaches, lip shape and colour, and there is visible jitter from single-frame generation. The repository’s news log ends on 5 April 2025.
The three that will bite you:
Wav2Lip is the one everyone names and almost nobody can use commercially. Its README states that because the models are trained on LRS2, “any form of commercial use is strictly prohibited.” The authors point you to Sync Labs. The generated face is 192×288. Wav2Lip stays valuable as the evaluation baseline (the LSE-D and LSE-C metrics it introduced still measure sync here, and need no ground truth) but it is not a product.
LivePortrait is not a lip-sync model at all, and that is the most common error in the category. It animates a portrait from a driving video, not from audio; the audio-driven work built on top of it is third-party. Its paper reports 12.8ms generation on an RTX 4090, and the repository notes adoption by Kuaishou, Douyin, Jianying and WeChat Channels as a June 2025 author claim. Here is the trap: the code is MIT, but the LICENSE file states InsightFace’s models are for non-commercial research purposes only, and that “if you want to use the LivePortrait project for commercial purposes, you should remove and replace InsightFace’s detection models to fully comply with the MIT license.” Nobody mentions this.
Sonic (CVPR 2025) has a non-commercial licence, needs a single 32GB GPU on Linux, and points commercial users to Tencent Cloud. SadTalker (CVPR 2023) is Apache-2.0 with the restriction removed, but its changelog ends on 12 June 2023 and it uses Wav2Lip internally for mouth generation. EMO (Alibaba, ECCV 2024) is the clearest citation-only release: the GitHub repository holds a paper reference and nothing else. Ditto (Ant Group, ACM MM 2025) is the most permissive of the new wave: Apache-2.0, TensorRT and PyTorch models released.
What things actually cost per finished minute
The figures below are calculated from list prices published on 16 October 2026, in USD before tax. Unconfirmed costs are marked.
| Tool and plan | Published price | Published rate | Cost per finished minute |
|---|---|---|---|
| D-ID Lite | $4.70/mo annual | 10 min/mo included | ~$0.47/min: personal use only, D-ID watermark |
| D-ID Pro | $16/mo annual | 15 min/mo included | ~$1.07/min: commercial licence, AI watermark |
| D-ID Advanced | $108/mo annual | 100 min/mo included | ~$1.08/min: custom logo, manual moderation |
| Sync Labs lipsync-1.9 | $5/mo + usage | $0.025/sec | ~$1.50/min |
| Sync Labs lipsync-2 | $5–19/mo + usage | $0.05/sec | ~$3.00/min |
| Synthesia Starter | $29/mo monthly | 1,250 credits, 10 free min | ~$3.82/min lip-synced dubbing |
| Sync Labs lipsync-2-pro | $5–19/mo + usage | $0.08325/sec | ~$5.00/min |
| Tavus Starter (generated video) | $59/mo | 10 min included | ~$5.90/min fully used; $1.00/min overage |
| Sync Labs sync-3 | $5–19/mo + usage | $0.1334/sec | ~$8.00/min |
| HeyGen, Captions, Hedra | : | credits published, rate not | Not computable from public information |
Synthesia’s arithmetic is worth showing. At $89/month for 6,000 credits, minus the 10 included video minutes, 5,900 remain. Lip-synced dubbing costs 76 credits per minute doubled to 152: roughly 38 finished minutes, about $2.29 each. That last row is the finding: four of the eleven vendors publish a credit balance without publishing what a credit buys. Ask, in writing, before you sign.
Five failure modes that decide your workflow
Drift, occlusion, multi-speaker audio, singing and code-switching are where lip-sync breaks. None are fixed by a pricier tier; they change how you work.
Drift over long clips is a segmentation problem before a model problem. Every retiming tool bills by duration, so chunking costs the same and lets you redo one bad section. Watch the pacing settings: Synthesia’s Adaptive mode lets video speed shift to fit the translation, Original fixes playback speed and retimes only the voiceover.
Side profiles and occlusions are the documented failure case in HeyGen’s own comparison table, and the reason Precision mode exists. A face at 40 degrees with half the mouth hidden is a different problem from a head-and-shoulders interview.
Multi-speaker footage needs speaker assignment before it needs a better model. Synthesia detects speakers per segment and lets you reassign, rename or create one, and creating a speaker builds a new voice clone from that segment’s audio, which is a consent question, not a technical one.
Singing is the worst case. Viseme models trained on speech do not model the mouth tension pitch changes produce. For a music video, budget for manual repair or accept the artefact.
Code-switching defeats language lists. Synthesia publishes a named list of target languages; Akool a count; Sync Labs simply accepts your audio. A presenter moving between two languages mid-sentence will make a language-gated translator refuse or switch voices mid-word.
Deepfake, consent and right of publicity
Three separate regimes are live now, and all three target exactly what these tools produce.
The EU AI Act. Article 50(4) requires deployers to disclose that content constituting a deep fake was artificially generated or manipulated, and Article 50(5) requires that disclosure at first interaction or exposure. Article 3(60) defines a deep fake as AI-generated or manipulated content resembling existing persons, objects, places, entities or events that would falsely appear authentic. Lip-syncing a real person’s photograph to speech they did not record falls inside that.
The United States. The FTC began enforcing Section 3 of the TAKE IT DOWN Act on 19 May 2026. Covered platforms must remove nonconsensual intimate images (including “digital forgeries” created using software, an app or AI) within 48 hours of a valid request, along with known identical copies, with civil penalties of up to $53,088 each. You are not the platform, but this determines what gets pulled down and how fast.
California. AB 2602, signed September 2024 and codified at Labor Code section 927, makes a contract provision unenforceable where it allows creation or use of a digital replica of an individual’s voice or likeness in place of work they would otherwise perform in person, without a reasonably specific list of uses, and without representation by counsel or a union. AB 1836 amends Civil Code section 3344.1 to require prior consent from an estate before creating a digital replica of a deceased personality, with penalties of the greater of $10,000 or actual damages.
Both statutes define “digital replica” to cover work in which the person did not actually perform or in which the “fundamental character” of the performance has been “materially altered.” That second clause is the one to think about: re-dubbing a performer so their mouth says different words arguably lands inside it.
My read, treated as a read: get a signed likeness release naming specific uses, and put a visible disclosure on the video rather than relying on metadata. D-ID’s January 2024 ethical pledge commits to building “ethical use” clauses into its terms and requiring that licensees hold proper rights to source material: a commitment, not an indemnity.
General information, not legal advice. If a campaign, a union-covered performer or an EU-facing audience is involved, get counsel.
The workflow that produces a shippable result
- Decide which of the three jobs you actually have. Still image, existing footage, or script-to-video. Choosing wrong costs the project.
- Get consent in writing before you generate anything. Named uses, channels and duration. A vague “all media” release will not survive AB 2602.
- Prepare the source. Front-facing, evenly lit, minimal head movement, 25fps if self-hosting MuseTalk or LatentSync: both trained at that rate. Trim silence; you pay per second for it.
- Segment long material. Ten videos per batch on Synthesia, up to 500 generations per batch on Sync Labs Scale. Chunking costs the same and makes repair possible.
- Choose the quality mode deliberately. HeyGen Speed for drafts and still faces, Precision for movement, side angles or occlusions. On LatentSync,
guidance_scalebelow 1.5 if you see jitter, above 2.0 if you see lag. - Decide the pacing model. Adaptive lets picture speed move; Original holds speed and retimes voice. Instruction: Adaptive. Fast persuasive edits: Original.
- Proofread the transcript before spending credits on the final render. Synthesia’s review step is free, but self-serve plans skip it, so budget for regeneration mistakes.
- Check the artefact at 100%, scrubbing on hard consonants. Plosives reveal drift faster than vowels. Then check the local player; a desynced export that looked fine in-app is the most common delivery failure.
- Ship with disclosure. Visible label plus machine-readable provenance. The EU rule is in force; a buried metadata flag is not compliance.
Which tool, by use case
- One photo, one presenter, social or explainer content → Hedra, or D-ID Pro if you need translation in the same tool.
- Localising real footage you own, at volume, with an engineering team → Sync Labs. Per-second pricing, 30-minute clips, 500-item batches, language-agnostic audio.
- Dubbing training material into many languages → Synthesia on Enterprise. Named language list, transcript review, SCORM export and SSO decide it.
- Live conversational video: interviews, reception desks, language practice → Tavus, the only one built for real-time two-way conversation.
- Everything on your own infrastructure → LatentSync 1.6 for accuracy, MuseTalk 1.5 for speed. Budget for an upscaler and a human pass.
- Face swap or re-age alongside the lip-sync → Akool, on a Business tier if the output is commercial.
Several other tools serve different needs. Wondershare Virbo now markets itself as a product and UGC ad generator; Steve AI gates its API behind Enterprise and offers no self-serve custom avatars; Video Tap generates blog posts and clips. Higgsfield’s pricing page requires a login; its current rates are not confirmed here.
Frequently asked questions
What is the best AI lip-sync generator overall? There isn’t one, because the three categories are not comparable. Hedra produces the most natural talking-photo result, Sync Labs the clearest pricing and longest clips for retiming, HeyGen the widest language coverage. Pick the category first, then the tool.
Which is cheapest per minute? Sync Labs’ lipsync-1.9 at $0.025 per second, about $1.50 per minute at 25fps. D-ID Lite is nominally cheaper at roughly $0.47 per minute but is personal-use only with a D-ID watermark. D-ID Pro at roughly $1.07 per minute is the cheapest commercially licensed option with published arithmetic.
Can I use Wav2Lip commercially? No, not from the open-source weights. The repository states commercial use is strictly prohibited because of the LRS2 training data. The authors now run Sync Labs, and their API is the sanctioned commercial path. Wav2Lip stays valuable as an evaluation baseline, since the LSE-D and LSE-C metrics it introduced still underpin how this field measures synchronisation.
What is the best tool for re-dubbing a mismatched video? Sync Labs if you want control and predictable cost; Synthesia if volume justifies Enterprise and you need transcript proofreading; HeyGen if you want translation and lip sync from one vendor. All three accept your own audio. Only Synthesia caps per-video length at 3 minutes, a real constraint for long-form material.
How do I stop lip-sync drifting on long clips? Segment before you generate. Every major retiming tool bills by duration or by second, so rendering in chunks costs the same and lets you fix a bad section without redoing the whole take. Synthesia’s batch limit of ten videos per job and Sync Labs’ batch endpoint of up to 500 generations both suit this. Also decide the pacing mode up front: Synthesia offers Adaptive (video speed shifts to fit the translation) or Original (only the voiceover is retimed).
Which tool has the highest resolution ceiling? Sync Labs publishes 4K native for sync-3 and 512×512 face resolution for lipsync-2. Akool Pro exports 4K, Synthesia 1920×1080. Among open-source models LatentSync 1.6 trains at 512×512 and MuseTalk works at 256×256, and MuseTalk’s authors state this has not reached the bound.
Are open-source models good enough for production? For drafts, internal content and anything under 1080p, yes. For client-facing 4K masters, not without help. LatentSync needs 18GB of VRAM at version 1.6 and exposes a sync-versus-realism dial you will tune per clip. MuseTalk runs at 30fps on a Tesla V100 and documents its own jitter and identity limits.
Do I have to label lip-synced video as AI-generated? In the EU, yes. Article 50(4) has applied since 2 August 2026 and requires disclosure that deep-fake content was artificially generated or manipulated, provided clearly at first exposure. Providers face a machine-readable marking duty with a 2 December 2026 compliance date. In the US, the TAKE IT DOWN Act governs removal of nonconsensual imagery including AI forgeries.
What about songs and music videos? Hardest case here, and no tool solves it. Viseme models are trained on speech, and melodic pitch changes alter mouth tension in ways they do not represent. Use short clips, check every frame, expect manual repair. Sync Labs’ react-1 is marketed for expressive short clips: vendor positioning, not a measured result.
How do I choose between credits and per-second pricing? Per-second if your volume is unpredictable, because you pay only for what you render. Subscription credits if output is steady month to month. Avoid any vendor publishing credits without the credit-to-second rate: HeyGen, Captions and Hedra all do this today, making their cost per minute unknowable until you ask.
Bottom line
Talking-photo tools animate images. Retiming tools fix footage you own. Avatar platforms generate whole videos from scripts, and they are not substitutes for each other.
If you take one thing from this: check the licence before you check the price. D-ID’s $4.70 plan and Akool’s $12 plan cannot produce commercial work. Wav2Lip’s open weights cannot ship at all. LivePortrait’s MIT code is unusable commercially until you replace its InsightFace models.
And if you are shipping to an EU audience, put the 2 December 2026 Article 50(2) deadline in your project plan, not a footnote.
Sources
All pages opened and read on 16 October 2026.
Open-source models and papers
- Wav2Lip: GitHub (Rudrabha)
- Wav2Lip evaluation framework and LSE-D / LSE-C: GitHub
- A Lip Sync Expert Is All You Need: arXiv 2008.10010
- MuseTalk: GitHub (TMElyralab)
- MuseTalk: Real-Time High-Fidelity Video Dubbing: arXiv 2410.10122
- LatentSync: GitHub (bytedance)
- LatentSync: arXiv 2412.09262
- LivePortrait: GitHub (Kling AI Research / Kuaishou)
- LivePortrait LICENSE: InsightFace restriction
- LivePortrait: arXiv 2407.03168
- Sonic: GitHub (jixiaozhong)
- Ditto: GitHub (antgroup)
- Ditto: arXiv 2411.19509
- SadTalker: GitHub (OpenTalker)
- EMO: GitHub (HumanAIGC / Alibaba)
- OmniHuman-1.5: arXiv 2508.19209
Vendor pricing and documentation
- HeyGen pricing
- HeyGen Video Translation documentation
- Synthesia pricing
- Synthesia AI Dubbing documentation
- Synthesia help: How do I dub a video?
- D-ID pricing
- D-ID Pledge for the ethical use of synthetic media
- Hedra pricing
- Hedra Terms of Use (updated 4 September 2026)
- Tavus pricing
- Tavus documentation
- Sync Labs pricing
- Sync Labs API documentation
- Sync Labs: sync-3 model announcement (April 2026, vendor claim)
- Akool pricing
- VEED pricing
- Captions pricing
- Steve AI pricing
- Video Tap pricing
- Wondershare Virbo
Regulation and legal
- EU AI Act, Article 50: Transparency Obligations
- EU AI Act implementation timeline (updated 31 August 2026)
- FTC: Complying With the Take It Down Act (May 2026)
- Fenwick & West: California’s New AI Laws Limit Uses of Digital Likeness
Competitive analysis
