Move from avatars to full production
Synthesia turns a script into a stock-avatar video. Hedra Agent and Workspaces produce the whole piece with your own presenter, voice, and footage, on avatar and character models served and post-trained on Hedra's own infrastructure.
export HEDRA_API_KEY="<key_id>:<secret>"
curl -X POST \
"https://api.hedra.com/v3/models/omnihuman-15" \
-H "Authorization: Key $HEDRA_API_KEY" \
-H "Content-Type: application/json" \
-d '{ "input": {
"prompt": "A presenter speaks to camera in a bright kitchen",
"aspect_ratio": "16:9",
"resolution": "1080p",
"start_image": "https://example.com/presenter.png",
"audio": "https://example.com/script.mp3"
} }'A more modern way to work
Skip the model aggregators. Hedra Agent and Workspaces give your team an autonomous agent that supercharges your workflows with a superintelligent understanding of the latest models: brief it in a shared space and it plans, generates, and edits across video, image, speech, and music, with your team and your agents working alongside it.
GPU infrastructure, an inference engine, and post-training research
Hedra operates GPU clusters and the inference engine that runs diffusion and autoregressive media models on them 8x faster. The same research team post-trains models for customers: supervised fine-tuning, LoRA training, and GRPO, on your data.
Hedra and Synthesia: focus at a glance
Synthesia is a finished product for presenter videos. Hedra serves and post-trains the avatar models a product like that is built on.
| Dimension | Hedra | Synthesia |
|---|---|---|
| What it is | GPU infrastructure, inference and post-training research | Avatar video for training and communications |
| Core technology | Hedra's inference engine, its own GPU clusters, and post-training (SFT, LoRA, GRPO) | Stock and custom AI avatars with text-to-speech in 140+ languages |
| Models | Open-weight models on Hedra's engine, models post-trained for you, and partner models through a gateway | Its own avatars inside its product |
| How you connect | API, SDK, CLI, MCP, and Hedra Workspaces | Synthesia app and API |
| Where inference runs | Hedra's GPU clusters, or your own | Synthesia's cloud |
| Best fit | Teams building visual products who need inference, capacity and custom models | L&D and communications teams making presenter videos |
Avatar, character and voice models on Hedra infrastructure
Video rates are per second of generated performance; voice rates are per 1,000 characters. Open-weight models run on Hedra's engine; partner models are reached through the gateway.
| Provider | Model ID | Rate |
|---|---|---|
| HeyGen | heygen-photo-avatar-4 | 10¢/s · up to 1080p |
| ByteDance | omnihuman-15 | 16¢/s · 1080p |
| Kling | kling-ai-avatar-v2 | 11.5¢/s · 720p pro |
| ElevenLabs | elevenlabs-v3 | 12.6¢ / 1,000 characters |
| MiniMax | minimax-speech-25-hd-preview | 12.86¢ / 1,000 characters |
Bring your own presenter, one request.
Synthesia picks the avatar from its library. On Hedra you send your own image and audio.
POST api.synthesia.io/v2/videos
Authorization: $SYNTHESIA_API_KEY
{ "input": [{ "scriptText": "Hello, World!",
"avatar": "anna_costume1_cameraA",
"background": "green_screen" }] }A stock avatar, a script, and a background from the library.
POST api.hedra.com/v3/models/omnihuman-15
Authorization: Key $HEDRA_API_KEY
{ "input": { "start_image": "<your presenter>",
"audio": "<your voice track>", ... } }Your character, your audio. Swap omnihuman-15 for kling-ai-avatar-v2 and nothing else changes; post-train on your performers if you need to.
FAQs
- How do Hedra and Synthesia fit together?
- Synthesia is an end product: pick an avatar, paste a script, get a video. Hedra serves the avatar, character and voice models under that kind of product on its own GPU infrastructure, and post-trains them for customers building presenter features.
- Can Hedra post-train a model for us?
- Yes. Hedra runs supervised fine-tuning, LoRA training, and GRPO on models you own or license, on your data and against your evals, then serves the result on the same engine. The weights are yours.
- Can this run on our own hardware?
- Yes. Hedra Inference deploys on your Kubernetes or bare metal, including air-gapped environments, across NVIDIA, AMD, AWS Trainium, Google TPU, and custom accelerators. Weights, data, and the control plane stay on infrastructure you operate.
