Undo the paywall. Make better media for free.

Official Real-ESRGAN x4plus weights converted to a pinned ONNX (the conversion script ships), run through ONNX Runtime (CoreML/DirectML/CUDA) with 512 px tiles and ramped overlap blending — golden-tested seam-free vs whole-frame. Temporal stabilization warps the previous output by Farneback flow and blends where warp error is low. The engine is a frozen API consumed by both the app and Pivot; every result carries which backend actually ran, and synthesis is always labeled synthesis.

• Per-frame model + temporal stabilization ≠ Topaz's temporal models on fast motion.
• Faces can go uncanny; face restore is opt-in and off by default for journalism.

Replaces: Studio Super Scale ($295) · Topaz Video AI (~$300) — honest ballparks; the paid tools are good tools, you just shouldn't need them.

faster-whisper (CTranslate2 int8) with Silero VAD and word timestamps; diarization via sherpa-onnx (pyannote segmentation 3.0 + 3D-Speaker embeddings, clustered), majority-overlap assigned per segment. Exports are pure, golden-tested writers: SRT/VTT with word-timed re-blocking to caption presets, speaker-colored marker EDLs (Resolve's Timeline Markers From EDL), and CMX3600 selects EDLs with accumulating record TC and handles. NDF timecode throughout, 23.976 drift documented rather than hidden.

• Crosstalk and heavy accents still beat it — a human captioner is the accessibility gold standard.
• Diarization confuses similar voices in roomy recordings.
• Post only — live captions are community-captioner's job.

Replaces: Studio Speech-to-Text ($295) · Descript / Rev / Simon Says ($15–30/mo) — honest ballparks; the paid tools are good tools, you just shouldn't need them.

Ingest races the watch page against yt-dlp for captions and takes the first winner; transcripts live as sidecars any tool can read. The insight engine is counted, not modeled: extractive briefs, pattern entities with conservative spelling folds, word-boundary framing lenses, sentence-level co-occurrence networks. Exports ride one ffmpeg concat graph — cuts, retimes, fades, and Pillow title cards in a single pass. The town's paper trail comes from its own CivicClerk portal, matched by the meeting's date and name.

• YouTube gates caption delivery by IP reputation; the suite routes around it honestly (your proxy, or the community caption relay) but a gated bare IP may need one.
• The AI reel, report narrative, and translations need your own API key — no key ships, nothing requires one.

Replaces: an afternoon of scrubbing (hours per meeting) — honest ballparks; the paid tools are good tools, you just shouldn't need them.

The portal reader speaks CivicClerk's OData dialect defensively: every URL-shaped string in an event is harvested with its field path, video-ish links sorted first, publishedFiles resolved through the portal's own GetMeetingFileStream endpoint. The zoomgov resolver walks share page → meeting id → play info → mp4 with the cookie jar and Referer Zoom expects. Downloads ride the suite's managed yt-dlp nightly, which never routes its own self-update through your proxy.

• zoomgov.com isn't supported by yt-dlp — the built-in resolver walks Zoom's share pages itself, and when Zoom redecorates, each step fails naming itself.
• CivicClerk tenants disagree about where recordings hide; the reader harvests every URL-shaped field and shows its work rather than guessing silently.

Replaces: screen recording the player (quality + hours) — honest ballparks; the paid tools are good tools, you just shouldn't need them.

Index is deliberately boring: a filesystem scan plus the .scribe.json sidecars the other tools already write, queried with plain substring and word matches. The interesting part is the contract — because Scribe, the Highlighter, and Index share one sidecar format, any file transcribed anywhere becomes searchable everywhere, and the FCPXML export turns hits into markers Resolve reads natively.

• Speech search needs words on disk — run Scribe (or the Highlighter's caption read) once per file first.
• It indexes filenames, metadata, and transcripts — not what's visually IN the frame.

Replaces: MAM/DAM systems ($1k–5k/yr) — honest ballparks; the paid tools are good tools, you just shouldn't need them.

The Library is an aggregation, not a database: each meeting's insight sidecar (counted framing, topics, entities, pace) is read through the same door the analyzer uses, then folded — entity spellings join under a conservative match (same word count, same initials, high sequence ratio), discourse rates normalize per thousand words so a seven-hour meeting can't out-shout a one-hour one, and the montage stitcher letterboxes mixed resolutions into one frame with a title card naming each clip's meeting.

• Cross-meeting cards need at least two read meetings; a meeting without words is listed as unread, not invented.
• The optional AI read sends only the counted digest — dates, tallies, names — never a transcript, and only on your own key.

Replaces: institutional memory (someone's whole job) — honest ballparks; the paid tools are good tools, you just shouldn't need them.

Isolation runs the official DeepFilterNet 3 binary as a subprocess (48 kHz, real-time class) with our mix-back blend. De-hum detects the mains fundamental by harmonic-sum test against the spectral floor, then zero-phase notch cascades. De-click flags second-difference outliers (no IIR ringing), interpolates runs ≤ 2 ms, and bails honestly if more than 1% of samples flag. Room tone is random-phase resynthesis of a Welch PSD profile — exact spectral envelope, loop-safe tail. Loudness is BS.1770 via pyloudnorm; if the true-peak ceiling fights the target, it says so instead of limiting.

• Full-wet isolation sounds processed — the default keeps 35% of the room, on purpose.
• No spectral repair painting (a cough under a word is still RX territory).
• Loudness normalize won't limit: if your peaks argue with the target, it says so instead of squashing.

Replaces: Studio Voice Isolation ($295) · iZotope RX ($400–1,200) — honest ballparks; the paid tools are good tools, you just shouldn't need them.

Three noise estimators run per frame (temporal difference, fine and coarse Laplacian) into robust histograms, producing separate spatial and temporal sigmas plus a 16-band brightness gain curve. The temporal stage compares 3×3 patches against up to ±3 frames with a hierarchical ±8 px motion search; a hard-knee gate reaches exactly zero past the knee, so mismatches cannot ghost by construction. Spatial cleanup is noise-adaptive NLM with a size-banded Noise EQ, bounded by Detail Rescue. The refine stack (v3.6) rebuilds optical character: coring, acutance, chroma-speckle, and film-matched grain. GPU kernels (Metal/CUDA/OpenCL) are line-by-line ports of one CPU reference, held to ~2×10⁻⁵ agreement by a parity test suite.

• No motion-compensated temporal NR — Studio keeps an edge on fast pans.
• Classical (NLM), not AI: very competitive at normal noise, less magical on starlight footage.

Replaces: Resolve Studio NR palette ($295) · Neat Video (~$160) — honest ballparks; the paid tools are good tools, you just shouldn't need them.

SAM 2.1 hiera-small propagates prompted masklets bidirectionally through a shot at 720p analysis resolution (PyTorch on Metal/CUDA; ONNX port planned). Confidence per frame is the model's mean in-mask probability — the amber QC strip. Post chain: 3-frame morphological majority (de-flicker), despeckle, grow/shrink, gaussian feather; mattes upsample to source res on export. Shot-bounded by the shared cut detector so propagation never crosses an edit.

• Soft-matte quality on hair and motion blur — it's a roto assistant, not a keyer.
• Heavyweight install (PyTorch) until the ONNX port lands.
• Prompt granularity is real: click a face, get a face; click a jacket, get a jacket.

Replaces: Studio Magic Mask ($295) · Mocha Pro (~$700) — honest ballparks; the paid tools are good tools, you just shouldn't need them.

MiDaS-small (256 px, MIT) behind a model-agnostic engine — Video-Depth-Anything-Small is the planned v0.2 backend. Temporal EMA resets at every cut; per-shot normalization uses 2/98 percentiles so hot pixels can't crush the range. Upsampling is a He-style guided filter against the source luma, keeping depth edges on image edges. Export packs 0–1 depth into 10-bit gray ProRes 4444 (near = white by default, invertible, gamma-mappable).

• Relative depth, not metric — mattes and atmosphere, not measurement.
• Mirrors, glass, and scale tricks fool it.
• v0.1 backend is per-frame with temporal smoothing; the video-native model is planned.

Replaces: Studio Depth Map + Relight ($295) — honest ballparks; the paid tools are good tools, you just shouldn't need them.

Speak models the photochemical chain rather than filtering the picture. A negative stage applies closed-form Hurter–Driffield characteristic curves, printer lights time the color in genuine printer points, and a print stage lands the result — all plotted live from the same math the pixels take, so the scope cannot disagree with the render. Saturation happens in the density domain (dye density adds where light multiplies), with an inter-image coupler for dye cross-talk. Halation is injected as exposure before the curve on a multi-scale scatter field with a power-law skirt, making it physically self-limiting. Grain is per-RGB-layer density noise, bandpassed at a physical pitch and boiling every frame, optionally keyed to Hush's clean-confidence matte. Metal on macOS, OpenCL on Windows/Linux, all agreeing with a CPU reference to ~2e-5; identity at default and alpha always passes through untouched.

• Early beta (v0.2): machine-verified module by module, but not yet field-tested through a full Resolve grade — don't cut a feature on it this week.
• macOS binaries only so far; the Windows/OpenCL build follows.
• No film-stock preset library yet — you build the look from the curves themselves.

Replaces: Dehancer (~$380) · FilmConvert Nitrate (~$140) — honest ballparks; the paid tools are good tools, you just shouldn't need them.

Analysis decodes once at 360p: shot detection (adaptive luma-diff), YuNet face detection every 2nd frame, YOLOX-s persons only where faces are absent, greedy IoU/distance tracking. Per shot, a subject is selected (size × centrality × persistence) and solved: shots under 2 s or under the motion floor become static punches at a trimmed median; the rest get the follow controller — deadzone with hysteresis, 12-frame lookahead anticipation, braking-parabola velocity setpoint clamped by v-max/a-max. Overshoot is pinned ≤ 4 acceleration quanta by golden tests. Renders are frame-accurate (EOF flush handled) with audio stream-copied, never re-encoded.

• Faces and people only so far — sports and wildlife will miss (saliency model on the roadmap).
• Punch-ins past native resolution are honest resampling until you enable the Rise engine.

Replaces: Studio Smart Reframe ($295) · auto-clip SaaS tools ($15+/mo) — honest ballparks; the paid tools are good tools, you just shouldn't need them.

Typography renders through Pillow at the output's own pixel size (no scaling pass), composited to premultiplied-safe straight alpha, then encoded by the suite's LGPL ffmpeg: ProRes 4444 for the editor, PNG for stills, palette-optimized GIF for the web. Countdowns render per-frame from the same layout engine, so the type never shifts between seconds.

• Static graphics and countdowns — kinetic type still belongs to Resolve's Text+.
• Font discovery uses the fonts your system already has; it won't fetch webfonts.

Replaces: hand-rolled every time (hours) — honest ballparks; the paid tools are good tools, you just shouldn't need them.