Generating avatars with Hedra?

Hedra renders stuck in queue or sync drifting? We unstick it.

Hedra promises a talking avatar from a single image and an audio clip, and most of the time it delivers. The rough days are when the queue backs up for hours, the mouth stops matching the vowels halfway through, or a voice clone bounces back because the consent script was not read verbatim. We get the pipeline moving again and build the guardrails so the next batch of renders does not need a human watching it.

  • Since 2015Fixing technology
  • FreeEstimates
  • US teamReal engineers

hedra-render.log

still failing

What it is

Hedra at a glance

Hedra is an AI avatar video platform built around its Character-3 foundation model, which takes a still portrait and an audio track and produces a lip-synced talking video. The company was founded in 2023 and is headquartered in San Francisco, with a mix of self-serve credits and enterprise API access.

The product covers short-form talking-head video for ads, training content, personalised sales outreach and avatar-driven social posts. Jobs run as asynchronous renders with credit pricing per second, voice cloning requires a recorded consent sample, and finished assets come back through the dashboard or a webhook. The web studio sits alongside a REST API for teams who want to generate in bulk.

Type
AI avatar video generation platform
First released
2024 Character-1 public release
License
Proprietary SaaS, credit based pricing
Maintained by
Hedra, Inc.
Typically used for
Talking-head marketing videos, personalised outreach, avatar presenters, short-form social content

The honest version

Hedra pros and cons

No technology is all upside. This is where Hedra earns its place and where it tends to cause trouble.

Where Hedra is strong

  • Single-image portrait supportYou feed Character-3 one still image and an audio clip and get a usable talking video. Most rival tools want a video reference or a studio capture first.
  • Natural head and shoulder motionThe model moves the head, blinks and shifts the shoulders rather than freezing the face around a moving mouth, which is the giveaway on older avatar tools.
  • Audio-first workflowBring your own voiceover or use a clone, and the render is driven by the audio rather than a separate script track that then has to be aligned.
  • API plus web studioThe browser app covers one-off edits; the REST API covers bulk generation. Teams rarely outgrow one without the other being ready.
  • Credit model is simplePricing is per second of output and shows on the job before you commit. There is no separate charge for voice cloning or face training to budget for.

Where Hedra bites

  • Queue times are unpredictableOn a quiet afternoon jobs finish in minutes; during a product launch the same job can sit queued for hours. There is no priority tier that reliably jumps the line.
  • Lip sync drifts on longer clipsAnything past the 45 second mark can start losing alignment on sibilants and plosives, and the fix is usually to cut the audio into shorter segments and stitch the renders.
  • Voice clone consent is strictThe consent recording has to match the script word for word in the speaker's own voice. Accented readings, missed words or background noise bounce the submission.
  • Content filter is opaqueA perfectly normal corporate video can trip an NSFW or policy flag without a clear reason, and the appeal process is slow when you have a deadline.
  • Portrait requirements are tighter than the marketing suggestsOff-angle photos, strong shadows or busy backgrounds raise the failure rate noticeably, even when the preview thumbnail looks fine.
  • Export options are limitedYou get MP4 at a fixed set of resolutions. Alpha channel, separate stems or lossless masters are not part of the standard export.

Sound familiar?

Common Hedra problems we fix

The Hedra problems that bring people to us, and what you have once each one is fixed.

job stuck in 'queued' state for 4h+

Renders sit in the queue past any reasonable wait

A surge of traffic, a failed webhook retry loop, or a job that silently moved into a dead state without the dashboard reflecting it. We find whether the job is actually progressing or needs resubmitting.

Queue moving, status honest

VoiceCloneRejected: consent mismatch

Voice clone submissions keep bouncing

The consent audio has to be the same speaker reading the exact script in a quiet room. One missed word or a noisy take fails the review every time. We script and record it the way the review pipeline expects.

Clone approved on first pass

lip-sync drift after 00:42

The mouth stops matching the audio halfway through

Longer clips accumulate phase drift and end up noticeably out of sync. Splitting the audio into shorter segments and rendering each piece restores accuracy, with a clean stitch at the join.

Sync holds across the whole clip

generation marked 'nsfw_flag'

The content filter is blocking safe videos

A corporate script or a medical explainer flagged for no obvious reason. We rework the portrait and audio to avoid the common triggers and keep the review response on file for later disputes.

Legitimate renders go through

WebhookTimeout after 3 retries

Finished jobs never notify your app

Completion webhooks fail because the endpoint is slow, returns the wrong status, or never got added to the allowlist. We harden the receiver and reconcile missing jobs by polling the API.

Every finished render recorded

InsufficientCredits mid-batch

Bulk jobs stop halfway because credits ran out

A nightly batch of 400 renders that fails on job 220 with no retry logic. Auto-top-up plus a pre-flight cost estimate stops it happening again.

Batches finish or stop cleanly

ImagePolicyViolation: non-front-facing

Source portraits keep being rejected

Headshots at the wrong angle, with sunglasses, heavy filters or cropped foreheads fail Character-3's face detection. We prep the images against the real requirements before upload.

Portraits accepted first time

What we do

Hedra help, from rescue to batch pipelines

Book us for a single failing render, or let us run the whole avatar pipeline.

Rescue
Queue backlogs cleared, rejected clones re-recorded correctly, and stuck jobs identified before you keep paying for them.
Voice clone onboarding
Consent scripts recorded the way the review process expects, with retakes planned if the room acoustics will not pass.
Bulk generation pipelines
API wrappers that batch submit, poll, retry and reconcile credits, so a 500-video run finishes without a human babysitting it.
Portrait and audio prep
Source images cropped and lit to pass face detection, and audio stems cleaned so lip sync stays locked for the full clip length.
Managed Tech
Ongoing avatar production with monitoring on queue health, credit burn, filter rejections, and a team on call when Character-3 pushes a model update.

How it works

From first message to fixed

  1. Look Free estimate

    Tell us what's going wrong. We look at it, then tell you plainly what's wrong and what it will cost.

  2. Fix

    We fix the cause, not the symptom. If something is down, getting it back up comes first.

  3. Harden

    Backups, monitoring, updates and access, so the same problem can't quietly come back.

  4. Hand back or stay on

    You get a plain-language write-up. Or we stay on for hosting, upkeep and new work.

Often found together

Running Hedra alongside something else?

We work on the whole stack, so the problem can be wherever it actually is.

Questions

Hedra questions

Something else? Call+1 737 727 7327

Why do my Hedra renders sit in the queue for hours?

Capacity varies by time of day and by model version. Character-3 jobs tend to backlog during US business hours, and sometimes a job silently moves into a stuck state while the dashboard still shows 'queued'. We check the API status of the specific job rather than trusting the UI, and if it is dead we resubmit against a quieter window.

Why does my voice clone keep getting rejected?

The consent recording has to be the exact scripted phrase, read by the person being cloned, in a quiet environment with no background music or edits. Paraphrasing, accented reads that change a word, or any detectable post-processing all cause a reject. A clean take in a treated room usually passes on the first submission.

How do I stop lip sync drifting on long videos?

Keep each render under 45 seconds of audio, split a longer script at natural breath points, and stitch the output clips in a separate editor. Longer single-shot generations build up phase error that becomes visible on sibilant and plosive sounds.

Can I generate Hedra videos at scale through the API?

Yes, and this is where most teams hit real issues. The API is straightforward to call, but reliable bulk runs need credit balance monitoring, idempotent retries on 429 rate limits, and a reconciliation pass against the webhook so no finished job is left orphaned. We build that wrapper once and reuse it.

Why did my corporate video get flagged as NSFW?

The safety filter is deliberately cautious and sometimes flags innocuous portraits or scripts. The practical fix is to adjust the source image lighting and framing, review the audio for anything that could sound ambiguous out of context, and if it still trips, submit through the appeal route with your business use case documented.

How much does Hedra help cost?

The estimate is free. Tell us the number of videos per month, the models you are using, and what has been going wrong, and we come back with a flat price before anything starts.

Free estimate

Hedra giving you trouble?

Tell us what Character-3 is doing, or not doing, and a US engineer will reply with next steps and a price.

  • No charge to find out what it will take
  • If it's urgent, say so

Replies come from a person. By sending, you authorize us to email you back. No marketing.