calesthio/generative-media-skills/d-id-avatar-video

D-ID avatar video production

Use this skill when the selected provider is D-ID or when the user asks for a D-ID talking avatar, photo-to-speaking-video, presenter clip, expressive avatar, interactive visual agent, real-time avatar stream, or D-ID-powered localization/support/training workflow.

Do not treat D-ID as a general cinematic video generator. It is primarily for human-presenter video: a face or presenter speaks text or audio with lip sync. For scenes that need object interaction, product handling, full-body action choreography, multi-shot narrative motion, or cinematic camera movement, use a broader video-generation or composition pipeline and use D-ID only for the presenter segments.

Facts in this skill were verified against official D-ID documentation and policies on 2026-07-10. Re-check live docs before quoting prices, enabled models, supported languages, moderation categories, endpoint schemas, or plan limits in user-facing commitments.

Choose the D-ID route before scripting

Pick the route from the deliverable, not from habit:

Official docs describe D-ID video APIs as supporting avatar videos from images, text, and audio, with V4 Expressive Avatars, V3 Pro Avatars, V3 Instant Avatars, and V2 Photo Avatars as the main video families. The V2 Talks endpoint transforms a source photo into a speaking avatar; the V3 Clips endpoint uses presenter avatars and optional drivers; V4 Scenes use created avatars. Source: D-ID Quickstart, Create a talk, Create a clip, Create a Scene, verified 2026-07-10.

Production preflight

Before any paid or irreversible generation:

1. Confirm the D-ID API key and credits. D-ID's FAQ says API keys are generated in Studio and valid credits are required; the API includes GET /credits for remaining and total credit items. Verified 2026-07-10.
2. Confirm commercial rights and plan constraints. Trial/guest outputs may be watermarked or limited to non-commercial use under D-ID terms; paid plans affect watermarking, usage, and credits. Do not promise commercial use until the user's plan allows it.
3. Confirm the person/likeness rights. For any real person, get explicit permission to use their image, voice, and likeness for the intended output, audience, and duration.
4. Confirm whether the input may contain personal data. D-ID's privacy policy says photos, text, audio, video, voice cloning data, and account data may be processed; minimize unnecessary personal data.
5. Choose text-to-speech vs external audio:

• Use text scripts for rapid iteration, localization, and simpler custody.
• Use audio scripts when exact timing, a pre-approved voiceover, a human recording, or an external TTS voice must be preserved.

1. Decide custody: where input files live, where D-ID should write results if using result_url, where webhooks land, and where final MP4s and metadata will be stored after download.
2. Build a QA checklist before generation so the team knows what "good" means: identity permission, facial fit, lip sync, pronunciation, captions, safe content, brand tone, resolution, watermarks/synthetic marks, and final file custody.

Input preparation

Source image for V2 Photo Avatar / Talks

Use a front-facing, well-lit image with one clearly visible face. Crop for a stable head-and-shoulders composition with room around the chin and hair. Avoid extreme side angles, heavy occlusion, sunglasses, open mouths, busy foreground objects, and images that imply a real person without permission.

Production heuristics:

• Keep face detail high enough for close-up delivery. Upscale or choose another photo instead of hoping D-ID will repair a tiny face.
• Preserve natural skin texture; aggressive beautification can create uncanny lip/cheek motion.
• Prefer a neutral or slight-smile expression for business, support, education, and localization. Strong source emotions reduce believable range.
• Use consistent lighting and color if the avatar will be composited with slides or brand graphics.
• If the user asks to animate a celebrity, public figure, employee, customer, child, or private individual, pause until rights and policy fit are explicit.

Presenter selection for V3 Pro / Clips

Use GET /clips/presenters to list supported V3 Pro Avatars and their presenter_id values. D-ID's docs also note that V3 Pro Avatars are used with /clips and that custom V3 Pro Avatars based on the user's own video footage require contacting D-ID. Verified 2026-07-10.

Choose a presenter by:

• audience trust fit: executive, educator, peer, support specialist, sales advisor;
• wardrobe/background fit: neutral office, green screen, home, outdoor, branded-safe;
• framing and gesture style: hands visible vs no-hands, energetic vs calm;
• localization continuity: one presenter per language family unless a localized face/persona is intentional.

Do not claim that the presenter can hold products, point to arbitrary UI elements, walk around a set, or interact physically with props unless that capability is explicitly available in the chosen D-ID presenter/driver workflow.

Custom avatar and consent footage

For V3 Instant Avatars, D-ID docs state that a consent verification process is required before unlimited videos can be generated with that avatar. For Express Avatars, POST /scenes/avatars accepts a source_url for the source video and a consent_id for a valid consent video; POST /consents creates a consent flow for supported languages. Verified 2026-07-10.

Capture custom avatar source footage as if it were a miniature talent shoot:

• one person only, stable camera, eye-level framing, even light, clean background;
• no copyrighted music or third-party voices in the source;
• natural speaking pace and a few seconds of neutral face;
• wardrobe and background that can be reused across future videos;
• separate consent capture in the required language, tied to the intended person and account.

If consent fails or is missing, do not work around it with a still photo, lookalike, or cloned voice. Ask for valid consent or switch to a stock presenter.

Script, voice, and language direction

D-ID can animate from text or audio scripts in relevant endpoints; /tts/voices lists voices and can be filtered by provider: Amazon, Microsoft, Azure OpenAI, ElevenLabs, or Google. Some endpoints include an x-api-key-external header for bringing an external ElevenLabs API key for IVC voices. Verified 2026-07-10.

Write for lip sync and trust:

• Use concise spoken sentences. A D-ID presenter is strongest when the face supports the message, not when it races through dense text.
• Keep one clear idea per sentence. Avatar videos punish long clauses because eye contact and facial motion make rambling feel artificial.
• Spell out uncommon pronunciations, acronyms, names, medicine/legal terms, and brand words.
• Add explicit pause strategy in the text or external audio. Use punctuation, SSML-like pauses if supported by the selected script/voice path, or pre-rendered audio with clean silences.
• Localize meaning, not just words. Adjust greeting, idioms, gesture expectations, and trust cues for the target culture.
• Match emotion to route. V4 expressive/sentiment paths can justify more emotional writing; V2 photo avatars should use restrained affect.

Prefer pre-rendered audio when:

• the user has a legally cleared human voiceover;
• the brand voice is already approved;
• pronunciation must be exact;
• another TTS provider is preferred;
• you need one master audio track reused across D-ID and non-D-ID edits.

When using audio upload, D-ID's POST /audios stores an uploaded audio/video input temporarily, converts the result to WAV at 16 kHz, supports audio/ and video/ MIME types, limits the file to 6 MB, and stores it for 24-48 hours unless routed differently by the endpoint parameters. Verified 2026-07-10. Download and preserve any source audio and generated result in your own project storage; do not rely on temporary upload URLs as the only copy.

API lifecycle pattern

Use D-ID's async job model deliberately:

1. Create or select the avatar/presenter:

• Talks: provide source_url.
• Clips: provide presenter_id and optionally driver_id.
• Scenes: create/select avatar_id; Express Avatar creation uses POST /scenes/avatars.

1. Provide the script as text or audio, with voice/provider configuration appropriate to the endpoint.
2. Add name and non-sensitive user_data for traceability. Never put secrets, private user data, or raw customer PII in user_data; D-ID docs describe it as echoed into responses and webhooks.
3. Prefer a webhook for production automation. If no webhook is configured, poll the matching GET endpoint (/talks/{id}, /clips/{id}, or scenes equivalent) until completion.
4. If using result_url, use a controlled HTTPS destination or presigned URL with the right permissions. Otherwise, download the returned video URL promptly into the project's asset store.
5. Store the request payload minus secrets, job id, endpoint, D-ID route, voice/provider, avatar/presenter id, consent id if applicable, source asset hashes, result URL, local final path, created/finished timestamps, cost/credit note, and QA result.

Common endpoint facts, verified 2026-07-10: