ROADMAP: OmniAvatar Render¶

Status: Living document Date: 2026-07-16 Related: PRD.md, TRD.md, PLAN.md

v0.2.0 (current — see PLAN.md)¶

Rename omniavatar-core/avatar → omniavatar-core/live.
New omniavatar-core/render interfaces (Generate/Status/Download, AudioUploader capability, Wait helper).
Render providers: HeyGen, Tavus, bitHuman.
Audio upload: bitHuman (file API) and HeyGen (new heygen-go/asset package wrapping upload.heygen.com/v1/asset) implement AudioUploader; Tavus remains URL-only (no upload API).
Symmetric registry API: *LiveProvider* / *RenderProvider*.

Evaluate a generic S3/GCS-presigned-URL uploader helper for providers without hosting (Tavus), likely as a separate utility package so omniavatar-core stays dependency-free.
Evaluate WAV support for HeyGen uploads (only MP3/audio/mpeg is documented as an audio asset type today); fall back to transcoding guidance in videoascode if HeyGen keeps the restriction.

Work lands in grokify/videoascode (consumer), tracked here for coordination:

[x] Concatenate per-slide narration (including pause gaps) into one normalized MP3 before generation (pkg/avatar.ConcatManifestAudio).
[x] vac avatar generate → presenter.mp4, cached by hash of (audio content + avatar config), matching the existing TTS-cache philosophy.
[x] vac avatar compose — FFmpeg circular mask + corner overlay with optional border ring; narration audio remains the authoritative track (never the avatar video's audio).
[x] Orchestrator integration: --avatar-* flags on vac slides video run the overlay as an integrated final stage (orchestrator.AvatarConfig). Upload-capable providers only (heygen, bithuman); Tavus stays on the decoupled --audio-url flow.
[ ] Slide-level overrides (visible: false) driven by the timing manifest (fade avatar in/out).

D-ID (Create a talk API: photo + audio/text) — validates the interface against a photo-first provider; likely requires a d-id-go SDK repo first.
Local/self-hosted renderer (e.g., open-source lip-sync models) for private deployments — exercises the interface without any network provider.

HeyGen lipsync-precision path (Path B): neutral presenter source video + replacement audio for frame-accurate sync and full visual control (framing, wardrobe, camera). Requires wrapping the lipsync endpoint in heygen-go and adding a SourceVideoURL field or extension to GenerateRequest.
Webhook/callback completion: all three providers support callbacks (HeyGen callback_id, Tavus callback_url, bitHuman webhooks). Add an optional capability interface so long jobs don't require polling.
Avatar normalization helpers: standardize provider output (square crop, fixed FPS/codec/duration alignment) — possibly in videoascode rather than here, to keep omniavatar free of FFmpeg dependencies.
Transparent output: Tavus supports transparent_background (WebM, fast mode); evaluate alpha-channel passthrough in GenerateRequest.Background once a second provider supports it.
Multi-language batches: one render job per language from per-language narration, reusing videoascode's JSON transcript support.

Real-time use of render providers (LiveAvatar/CVI already cover this via the live surface).
Provider-side full-presentation rendering (slides stay a local, deterministic FFmpeg composition).