Operations

All tools below are local and free. ffmpeg is assumed present (it backs the engine already).

Cut / trim: keep a slice

In-composition trimming usually needs no new file: a clip plays a sub-window via data-media-start + data-duration (see hyperframes-core). Only cut a physical file when exporting/assembling outside the composition.

Reframe / crop: change aspect ratio

For a non-destructive crop, set a clip-path on the element in the composition itself (render-time, source file untouched) instead of re-encoding with ffmpeg.

Montage / stitch: join clips

Silence-cut / highlight: trim dead air, grab the best moment

Transforms with a quality choice (process)

These have a local option AND a higher-quality HeyGen-CLI option. Run the local one for free/offline; use the HeyGen CLI when quality matters. Showing the user a side-by-side (local vs HeyGen) is the honest way to let them choose.

After any op: resolve --from out.ext --type <type> to register the derived asset (it records provenance and auto-promotes to the global cache).

ponytail: media-use doesn’t re-wrap ffmpeg/heygen here, that’s deliberate (OP1). The value it adds is the ledger + global reuse on the output, via --from. Add a thin process verb only if agents repeatedly fumble these recipes.

Exact error-diffusion dither

Use the local processor when the requested look specifically calls for Floyd-Steinberg, Atkinson/Macintosh, Jarvis-Judice-Ninke, Stucki, Burkes, or a Sierra variant. These are sequential error-diffusion algorithms, not the realtime Bayer effects.dither shader.

Available algorithms: floyd-steinberg, atkinson, jarvis-judice-ninke, stucki, burkes, sierra, sierra-lite, and two-row-sierra. The default is balanced Floyd-Steinberg with a black/white palette. Palettes contain 2-6 #rrggbb colors in authored dark-to-light order; reversing the order intentionally inverts the mapping. --point-size controls 1-20px blocks; --brightness and --contrast accept 0.5-2; --detail accepts 0.1-1.

The processor supports ordinary SDR images and MP4 video, preserves video audio, and emits BT.709 MP4. It rejects tagged PQ/HLG input rather than silently tone-mapping it. To animate the transformation, keep the original and processed files as two real media layers and use the seek-safe GSAP timeline to reveal or crossfade between them. Use the realtime Bayer shader instead when the dither amount itself must animate continuously.

Transcription (default: Parakeet, better than whisper.cpp)

transcribe.mjs is the default local transcription path. It runs NVIDIA Parakeet-TDT via parakeet-mlx, which beats whisper.cpp on the Open ASR Leaderboard (avg WER ~6.05% vs 7.44%; on NOISY audio 4.73% vs 5.96%, where whisper-large-v3 hallucinated to 308% WER on meetings) and is 5-10x faster. It emits { text, words:[{text,start,end}] } with word timestamps (merged from Parakeet’s sub-word tokens), feeding transcript-cut, captions, and the audio engine directly.

VERIFIED on 24GB: accurate, ~3s (cached) for 8s audio. Parakeet covers English + 25 European languages. For other languages, or when parakeet-mlx is not installed, transcribe.mjs auto-falls-back to whisper.cpp (99 languages) via hyperframes transcribe. --engine parakeet|whisper forces one. (Cohere Transcribe tops the leaderboard on paper but its mlx-audio quants produced garbage and ran 40-70x slower on a Mac in testing, so it is not wired in.)

Text-based editing (transcript cut)

transcript-cut.mjs is a compiler, not a wrapper: it turns word timestamps and agent cut decisions into exact kept segments. It is provided even though the rest of this file is guidance-only.

Use --plan first when you want to inspect the kept segment JSON before encoding.

Ducking (declare in-composition / bake for export)

B1, declare ducking in the composition. audio-duck.mjs emits a volume lane as a data-automation attribute. Add it to the background <audio> element; the source file stays untouched. Lane times are clip-local, so pass --composition to let the script subtract the element’s data-start.

B2, bake ducking only for exported or standalone files.

Declare inside compositions. Bake only for assets leaving the hyperframes pipeline.

Publish loudness

Two-pass loudnorm measures first, then applies the measured values with the target LUFS baked in.

Socials target, -14 LUFS:

Podcast target, -16 LUFS:

Generate: images (local first, cloud upsell)

resolve --type image retrieves from the HeyGen catalog first; on a miss it GENERATES. Two paths, best-for-the-machine picked automatically:

1. Local (default, free, private): mflux (FLUX-on-MLX). resolve spec-checks AVAILABLE RAM and runs the best FLUX-class model that fits, via scripts/lib/local-models.mjs (imagegen ladder) + mflux-provider.mjs. The RAM ladder (agent sees it via describeModelLadder("imagegen", specs)):

   Tier	Model	Needs (available RAM)	Notes
   medium	FLUX.1 schnell int4	~8GB (--low-ram)	~20s/512px on 24GB. VERIFIED. Fast.
   large	FLUX.2 Klein 4B int4	~32GB	higher quality, full-resident
   xlarge	Qwen-Image	~64GB	top quality, 64GB+ Macs only

   Gotchas baked into the table: the official FLUX repos are HF-gated, so it points at non-gated community 4-bit re-uploads; and --low-ram is MANDATORY at the medium tier (without it a 768x512 run swap-thrashed to 90 minutes on 24GB; with it, 20 seconds).

2. Cloud upsell (better quality): the codex CLI image_gen tool, on the user’s ChatGPT subscription (codex owns auth, no key here, no per-call charge). It is the automatic fallback when no local model fits AND the explicit “make it better” choice on any machine. Users who just want codex can ask for it directly. Verified: prompt -> raster -> frozen + ledgered.

--local-only keeps mflux (once cached) and skips codex (network).

Generate: video (resolve --type video, HeyGen avatar first)

resolve --type video "<intent>" is the default path. It generates a script-driven HeyGen avatar video first (the free-usage allowance — OAuth sessions ride the web-plan free avatar-video quota where eligible, API keys follow normal API billing), falling back to local generative LTX only when HeyGen is unavailable, uncredentialed, or --local-only is passed. The two are non-substitutable outputs (a real presenter vs. a generic generative clip), so treat the fallback as “HeyGen wasn’t reachable,” not “upgrade the quality”:

• HeyGen avatar video (default, free for new API users): heygenVideoGenerate (scripts/lib/heygen-video-provider.mjs) shells the heygen CLI — never the raw API — auto-picking a public avatar and a starfish voice (override with --avatar-id/--voice-id, threaded through as ctx.avatarId/ctx.voiceId). If the CLI reports not_authenticated, the provider prints an onboarding recommendation (avatar video is free for new API users — sign in) to stderr and falls through to LTX instead of hard-failing.
• Local fallback: LTX 2.3 on MLX via dgrauet/ltx-2-mlx, the videogen ladder in local-models.mjs (ltx-video-provider.mjs). Generative clips (t2v), spec-gated to RAM. Verified on 24GB: 512x320 x 33f with audio.

Every generating heygen call from media-use — TTS, avatar video, and catalog search — sends the allowlisted X-HeyGen-Client-Source: media-use header (persistent flag, works on every subcommand) via the shared HEYGEN_CLIENT_SOURCE_ARGV constant (scripts/lib/heygen-cli.mjs), so usage tags correctly in billing/resource meta and shows up in the API dashboards. Read-only discovery (avatar list, voice list) doesn’t need it.

For structured bodies resolve --type video doesn’t expose yet (a specific avatar_id/voice_id combination beyond the ctx overrides, or a pre-recorded audio_url instead of a script), the raw heygen video create recipe below remains the escape hatch:

Avatar videos are deterministic + script-driven (lip-sync from a script or a pre-recorded audio_url), distinct from the generative LTX clips. After a manual recipe renders, resolve --from <downloaded.mp4> --type video to ledger it (not needed when generating via resolve --type video directly — that already ledgers the result).

Image-to-video (animate any still into a talking clip)

Not wired into resolve --type video (deferred — the avatar type covers the default script-driven case). heygen video create takes the raw POST /v3/videos body, so switching type from avatar to image animates any image of a person into a lip-synced talking video, with no avatar/photo-avatar creation step first. Point image at a public URL or an uploaded asset_id, and drive speech with a script+voice_id or a pre-recorded audio_url:

Common optional fields: title, resolution (4k/1080p/720p), aspect_ratio, remove_background, background, voice_settings, motion_prompt + expressiveness (photo-avatar animation), and callback_url/callback_id for webhooks. Don’t hardcode these from memory: the CLI self-documents the full, current body with heygen video create --request-schema (a discriminated union keyed on type), so read the schema rather than trusting a stale field list. For a still you’ll reuse across many scripts, create a reusable Photo Avatar once instead (heygen avatar create). Ledger the result with resolve --from <downloaded.mp4> --type video. Docs: https://developers.heygen.com/image-to-video (opens in a new tab).

HEVC / H.265 sources

HEVC/H.265 sources need no conversion for render (FFmpeg pre-decodes all input video) or for preview (auto-proxy transcodes and caches an H.264 copy on first use, disable with --no-proxy or media.autoProxy: false in hyperframes.json). A manual H.264 proxy via ffmpeg -i in.mp4 -c:v libx264 -crf 18 proxy.mp4, registered with resolve --from, remains available for edge cases (e.g. auto-proxy disabled, or ffmpeg unavailable at preview time).