RigSuite 0.97 Release Notes

• The freeze drone is rebuilt as its own voice: press FREEZE and the delay grabs what you just played and holds it, without the loop's own losses and without thinning, for as long as you like. The held level follows your playing instead of a fixed target, the delay keeps echoing live over the drone, and engaging, switching voices and re-grabbing are all clean. Measured on the real engine, a held three-note chord keeps every note for three minutes.
• Four voices play like four instruments: Vapor is air (weightless and wide, with a far, shimmering halo two octaves above the chord), Vesper is rays (pure, still partials shining down into the chord over a clear pad), Vault is the chamber and the ceiling (the low register beneath, a hollow resonant octave above, breathing together), Waver the dream (the pitch wanders, the texture drifts in and out of focus, the image leans slowly left and right). Each is its own signal path with its own kind of top layer, and a new SHEEN control sets how much of that layer you get.
• SWELL and STACK: SWELL sets how fast a grab fades in, from an immediate hold to a slow four-second bloom, and the release mirrors it. STACK changes what a tap does while frozen: instead of melting into a new capture, the new grab layers on top of the held ones, up to four, and the pad gets denser rather than louder.
• An AFTERGLOW section on the AFTERECHO pedal: every hold control now lives in its own labeled section along the bottom of the pedal, named for what it holds (the part of the echo that stays lit), and the whole section lights while a hold is running. FREEZE moved in there as a single button (the old RUN label is gone: the delay never stopped running).
• Footswitch control is complete: tap to freeze, tap again to melt or stack a new capture, quick double tap to release, and new in 0.97, hold the switch about a second to let the whole drone go. Release and CLEAR are learnable MIDI commands of their own too.
• Full-rig captures next to a stereo cab pair: loading a full rig on one side already bypassed that side's cab. If the other side has no amp, its cab is now bypassed as well, so the rig's own cab is never run through a second IR. Load a real amp there later and its cab comes back on its own.
• Any alphabet in a file name: models, cab and reverb IRs, presets and folders whose names use Cyrillic, Turkish, accented or East Asian characters now appear in the browsers and load. They used to be skipped or fail to open on Windows.
• Big libraries open fast: the first open of the NAM library in each session used to re-read every model file, which could take minutes on a large collection. RigSculpt now remembers what it learned between sessions and reads ahead in the background while you play, so the library is ready when you open it.
• Solo and hide in the library: every row of the Explorer's Authors, Amps and Packs lists now carries an S and an X chip at its right edge. S solos the row (only those models exist until you clear it), X hides it, the S and X keys do the same to the row under the mouse, and the strip at the foot of the list shows what is on and clears it with a click. A solo is not a selection. You still click through packs and amps while it holds, and every list honors it: browse, search, favorites, recents, the archetype tiles, Similar, Partner and More like this.
• More like this gives sharper answers: a targeted move (brighter, more gain, bigger lows and the rest) now keeps to the model's own archetype first. Models from another archetype follow after, marked with a !, and every row shows its similarity percent next to how far it moved on the axis you asked for.
• The reference stays pinned: Similar, Partner and More like this keep the model you started from in a REF band at the top of the list, however far you scroll. Click it, or press R, to hear the original again, then click back to the candidate. Right-click inspects it, and it drags to A or B like any row.
• GPU inference (Windows): Settings gains a GPU switch (every instance in the DAW at once, like Playback Mode) that runs the neural amp models on your graphics card through Direct3D 12, and a separate Render GPU policy for offline renders (AUTO hands renders to the card from 8x up, where it measured faster; the processor keeps 1x and 2x). It adds a little latency, reported to your DAW (about 5 ms at 1x and 2x oversampling, 8 ms at 4x, 12 ms from 8x up), and buys the most headroom: 8x oversampling costs the card about what 1x does. The sound is the same to floating-point precision, the setting belongs to the computer rather than the preset, and if the card ever stops responding the model falls back to the processor and the Settings line says so. With the card on, the oversampling slider gains 16x and 32x, which run on the card only and belong to this computer rather than the preset; the line below the switch reports how long a block takes at that rate on your card and buffer before you commit.
• The Workshop analyzes on the graphics card by default: the Workshop now runs the probe inference on your graphics card through Direct3D 12 (the RigSculpt VST's own kernels) and analyzes several models at once. A model that took nine seconds takes about two, and a twelve-hundred-model library comes in under an hour, with the same results: the first model of every run is rendered both ways and compared, and a mismatch keeps that run on the processor.
• The tour has new cards: the guided tour (About, Tutorial) grows from 15 to 22 cards. New cards cover the three library sorts and the solo and hide scope, search, favorites and sort, auditioning against the pinned reference on a real list, AFTERGLOW on the Reverie delay, the MIDI CONTROL list opened for real, and the CABS page with the full-rig rule.

The freeze drone is rebuilt

The 0.96 freeze held a chord by locking the delay loop: input muted, feedback forced to one. It worked, and it learned to re-grab, but that single fact caused every complaint it drew. The loop could only hold whatever the echoes happened to be at the press, the held sound went on through the loop's own filters and lost a little more of itself every pass, and while it held, the delay could not echo anything else.

In 0.97 the drone is a separate voice. At the press, RigSculpt captures one period of the delay (the chord as you heard it in the echoes, matched to the TONE you had set) into its own hold, spliced seamless and recirculated without loss, and the main delay keeps running underneath. Three things follow from that by construction rather than by tuning:

• The level follows your playing. Every grab is measured and placed a few dB under the echo it replaced, so RUN to FREEZE, and every re-grab after it, lands at a level that makes sense for what you were just doing. Soft grabs sit a little lower in the window, hot grabs a little higher; nothing lands outside it. The center of that window was re-tuned by ear during this round, so LEVEL at 0 dB reads as a pad under the playing rather than on top of it.
• The hold is lossless, and it stays put. The 0.96 BODY control traded definition against wash by letting the hold wander; in testing it was almost always left at 100, and the wander was what made the voices read as wavy, so the hold now simply stays put and the held level never sags. Measured over three minutes, the weakest note of a held chord stays 13 dB clear of the bed.
• The delay keeps working. Live playing echoes over the drone exactly as it does when nothing is frozen; a held pad is never a reason to stop playing through the delay.

The faint scratch at the moment of engage is gone, and it took three fixes, because it had three causes. The delay loop stores the raw input in its newest lap and only filters on the way back around, so a verbatim grab was one TONE pass brighter than the echoes you heard, and the shimmer then lifted that extra brightness into a short bright cluster; the capture is now tone-matched. Switching voices used to step the drone's read position, which the recirculating hold then repeated every lap; voices now glide. The young pad's machinery carried a little extra top for its first second or two; a final filter at the drone's output now bounds anything the hold can emit to the brightness of the echoes themselves, the same guarantee the old in-loop freeze had by accident. Each of these is held by a dedicated offline test, including a difference measurement that isolates the drone from everything else the delay is doing.

Four voices that play like four instruments

The 0.96 voices were four flavors of one engine, and they shared a misty character because they shared a signal path. In 0.97 each voice is its own path on the drone's output. None of it recirculates, so the hold itself is untouched, and every layer glides when you switch voices, so a switch never clicks.

• Vapor is air: weightless, wide, euphoric. The low mass lifts away entirely, the pad passes a second diffusion stage for a sense of immense space, and its top layer is a halo: the chord two octaves up and nothing else, read from the darker body of the pad and veiled again so it never turns glassy, its pitch drifting a few cents each side at two slow, unequal rates so it shimmers without ever beating in step, blooming with the chord's own rise and fall, smeared through its own diffuser and drifting slowly across the stereo field. Where Vesper's rays are close and still, the halo is far and alive. A very slow, broad tilt drifts faintly through the tone.
• Vesper is rays: light shining from above into the chord. Pure, consonant partials (an octave and two octaves on one side, an octave and a twelfth on the other, so the width comes from the intervals and never from beating), polished so they sit clearly above the pad and above BLOOM, and completely still: nothing in this layer moves. The pad under it is the cleanest and the stillest of the four, a little lighter in the bass, with only a trace of an open vowel underneath.
• Vault is the chamber and the ceiling. The sub-octave foundation sits prominent under the chord, never quite still, with a slow pressure building and relaxing beneath it and two inharmonic resonators ringing quietly under that like distant metal under tension; and above the chord, its own ceiling: a single octave colored by a hollow stone-room resonance, breathing with the same pressure as the floor below, so the two move together against the still chord between them. Low and high at once. The BLOOM shimmer still stacks on top of the foundation, like drawing a second organ stop.
• Waver is the dream: beautiful and unstable. The pitch wanders (a slow periodic wow with a random drift riding it), the texture breathes in and out of focus, the stereo image leans slowly left and right, and a broad low-pass wanders across the top, as if the hold were heard underwater or through heat haze. Tape-style aging mellows it over a long hold, and a low head-bump keeps it intimate.

One design rule runs through all four: the motion is glacial (the fastest cycle is about twelve seconds) and the moving filters are broad tilts, low-passes and phase sweeps, never a narrow resonance on the move. A moving resonance is what made the old Vesper's vowel sweep read as a mouth; these read as weather.

Listening tip: switch between the voices on a held chord with BLOOM at zero first, which is where the four paths separate most clearly, then bring BLOOM up and hear it stack differently on each. SHEEN sets how much of each voice's top layer you get; at zero, Vapor, Vesper and Vault show their bodies alone, and the three bodies are deliberately different too: Vapor thin, wide and smeared; Vesper clean and still; Vault weighted. Each top is a different kind of layer as well, not the same layer at a different level: the halo has no octave in it where the rays are built on one, and the ceiling is a single hollow octave.

SWELL and STACK

The 0.96 drone had a CHOIR depth control that scaled each voice's intensity. It never sounded right, and the reason turned out to be simple: every axis it moved was pitch or spectral motion, which a held drone amplifies into seasickness. 0.97 replaces it with two controls that are verbs rather than intensity dials.

• SWELL is how fast a grab fades in, from 0.15 seconds (an immediate hold) to 4 seconds (a slow bloom out of nothing). The release tail mirrors the setting, so one pill gives the drone both its entrance and its exit. The default reproduces the 0.96 envelope exactly. The melt between an old drone and a new capture is deliberately not on this control; that crossfade stays the one you approved in 0.96.
• STACK changes what a FREEZE tap does while already frozen. MELT, the default, is the 0.96 behavior: the held drone melts into the new capture. STACK layers the new grab on top of the held ones instead: grab a low note, grab a chord tone over it, and build a drone one note at a time, up to four layers. The drone's level rides down as layers are added, so the pad gets denser, never louder (three stacked layers measure 1.5 dB above a single one). Beyond four, the quietest layer melts under the new grab. Release lets the whole stack go, on the SWELL's release tail.

Under the hood, the drone keeps five hold buffers for four layers, so a new grab always lands in an idle buffer; a buffer that is still sounding is never overwritten while you can hear it, even under rapid tapping.

The AFTERGLOW section and the FREEZE button

The hold's controls now live in their own labeled frame along the bottom of the AFTERECHO pedal, named AFTERGLOW for what it holds: the part of the echo that stays lit. Inside: FREEZE, the four voice chips, SHEEN, SWELL and LEVEL pills, and the MELT | STACK chip. The frame illuminates while a hold is running, the same "engaged" convention the chain strip uses, so you can see a hold from across the room.

• FREEZE moved in. The header used to carry a RUN | FREEZE pair. Since the delay never stops running, RUN said nothing, and the pair is replaced by a single FREEZE button inside the section that lights while engaged. The parameter underneath is unchanged, so existing MIDI mappings and automation keep working.
• The pills print real units. SHEEN in percent, SWELL in seconds, LEVEL in dB (it used to show its trim as a bare percentage), on a brighter face; the old near-black pills read as disabled.
• The interval chips moved right, to sit over the Diffuse and Bloom knobs they modify, instead of floating in the middle of the mode-chip row.

FREEZE from a footswitch

The full gesture set for a momentary switch mapped to "Delay freeze (hold / re-grab)":

• Tap to freeze. The grab lands at the press, which is the whole point of the instrument, so nothing in 0.97 delays it.
• Tap again while frozen to melt the held drone into a new capture, or, with STACK on, to layer the new capture on top.
• Quick double tap to release (unchanged from 0.96).
• Hold about a second to let the whole drone go. New in 0.97. The release follows the SWELL tail, so a slow SWELL gives you a slow, graceful exit. If you would rather release from its own footswitch, "Delay freeze RELEASE" is in the MIDI menu as a separate action.

A latching switch keeps its freeze-while-held behavior; the long press applies to momentary bindings only, where it has a meaning. "Delay CLEAR (flush echoes + drone)" also joins the MIDI menu as a switch action: the same instant flush as the panel's CLEAR button, for the moment you need silence now.

Full-rig captures next to a stereo cab pair

Since 0.96, loading a full-rig capture (one with its cab baked in) on a side automatically bypassed that side's cab IR, while an amp-only capture on the other side kept its cab. One scenario slipped through: one amp and a stereo cab pair, then a full rig loaded where the amp was, with nothing on the other side. A single amp chain feeds both cab sides, so the rig's already-cabbed signal was run through the second IR as well.

0.97 extends the rule to the amp-less side: a full rig on one side with no amp on the other (or with amp stereo off) bypasses both cabs. The bypass is tracked as RigSculpt's own, exactly like the same-side rule, so a cab you switched off yourself is never touched, and the moment an amp-only capture lands on either side, the cab that was bypassed for it comes back on its own.

Any alphabet in a file name

On Windows, a file name is handed to the operating system as a short list of bytes, and the meaning of those bytes depends on which text encoding the program claims they are in. RigSculpt stored its paths in the system's local encoding, which covers the Western alphabet and little else; a model named in Cyrillic, Turkish, or Chinese, or a library inside a user folder with an accented name, could not be represented, so the scanner skipped it (0.96 logged the skip) and a direct load failed.

0.97 keeps every stored path as UTF-8, which represents every alphabet, and converts it to the operating system's own form exactly once, at the moment a file or folder is opened. That rule now runs through the NAM library, the cab and reverb IR browsers, the pedal browser, presets, the preferences and cache files, the factory-content installers and the loaders, and the IR loader gained a path-aware entry point so the conversion reaches it intact. An offline test creates a folder and a model with Cyrillic, umlaut and CJK characters in their names and confirms the library lists the file under its exact name and reads its metadata.

Big libraries open fast

A NAM file is one large text document, and the metadata RigSculpt reads to build its library view sits inside it, after the model's weights. Reading the metadata means parsing the whole file, weights and all, and the library did that for every file on the first open in each session. On a library of a few thousand models that is minutes of reading before the view appears, and a report of five minutes is what prompted this fix.

• The results are kept between sessions. What the reader learns is saved to a small cache file next to your preferences (framing_cache.json), keyed by each file's size and modification time, so an edited or re-embedded model is always re-read and a renamed one is simply learned fresh. The cache is a convenience only; delete it at any time and RigSculpt rebuilds it.
• It reads ahead. A background task warms the cache from the moment the plug-in loads, walking your library folders while you play, so even the first open of a session finds the work already done.
• One reader, three views. The NAM library, the pedal browser and the full-rig views share the same metadata reader and all benefit.

Solo and hide in the library

A large collection needs two things a selection cannot give you: a way to narrow the whole Explorer to a few makers or packs for a session, and a way to take the models you never reach for out of every list without deleting them. Every row in the Authors, Amps and Packs lists (including a single model under an author, an amp or a pack) now shows an S chip and an X chip at its right edge when you hover it. S solos the row: from then on only those models exist, in every view. X hides the row. Both stay lit while they are on, the S and X keys do the same to the row under the mouse, and the strip at the foot of the list names the keys, shows what is on, and clears it all with one click. The right pane's heading shows what is on.

Soloing is not selecting. Selecting a pack is where you are looking; soloing a pack is what exists, so you can solo a pack and then browse its amps, search inside it, or open Similar, Partner and More like this on a model and get answers only from that pack. Solos on different lists combine as an intersection (solo a maker and a pack to get that maker's models in that pack); solos within one list add up; a hide always wins. The archetype tiles recount to what survives, and the relational lists are rebuilt through the filter with a wider candidate pool, so a solo'd pack still fills a Similar list. The filter is for the session: reopening the plugin starts clean.

More like this keeps the answer in the family

Ask for something brighter than a modern tight amp and the old list could lead with a clean amp whose EQ fingerprint happened to sit close on everything else. It was a correct answer to the wrong question. The list now keeps to the model's own archetype first (clean headroom, edge of breakup, crunch, saturated lead, modern tight, saggy hi-gain), ranked by similarity within it, and only then lists models from other archetypes, each marked with a ! so you know the move is bigger than the one you asked for. Every row now also carries its similarity percent beside the measured change, so a +1883 Hz at 82% and a +4736 Hz at 61% read as what they are.

One click takes you back to the reference

Comparing a candidate to the amp you started from used to take three moves: Back to model, then the amp, then back to the list. Now every relational list, Similar, Partner and More like this, pins the reference in a band at the top, marked REF, that stays put while you scroll and audition. Click it, or press R, and the original is back in the slot; click the candidate again and you are comparing. The band tells you where the reference is right now (loaded in A, or waiting), right-click reads its measured story, it drags to A or B like any row, and its star works. On a Partner list the band inspects on click instead, because that list fills the other side.

GPU inference (Windows)

Settings gains a GPU switch, for every instance in the DAW at once (like Playback Mode). OFF is the processor, as before. GPU runs the neural amp models on your graphics card through Direct3D 12 and returns each block a little later (about 5 ms at 1x and 2x oversampling, 8 ms at 4x, 12 ms from 8x up, the same figure at any small buffer size: the card's timing jitters more at the big settings and the buffer is sized to cover what was measured), with that latency reported to your DAW, so nothing drifts and the card does the heavy lifting: 8x oversampling costs it about what 1x does. The sound is identical to the processor path to floating-point precision; every model was checked against it. The setting belongs to this computer, never to a preset or a project, so a preset you share never asks for a card the next machine does not have. Models the card cannot take (a few unusual architectures) stay on the processor, and the Settings line says why. A block the card delivers late plays as a short fade to silence instead of stalling your DAW, the caption counts it, and a card that is repeatedly late drops the model back to the processor. Offline renders have their own policy, the Render GPU button beside the switch, decoupled the way Render Mode is from Playback Mode: AUTO (the default) hands a render to the card from 8x oversampling up, where it measured faster than the processor (about 1.7x at 8x, 7x at 32x), and keeps 1x and 2x on the processor, which renders those faster; ON always, OFF never. A bounce has no deadline, so the card is waited for and a render never contains a dropped block; a render at 16x or 32x runs the card one block at a time, its fastest way of working at those settings, with no added latency. If the card stops responding, the model falls back to the processor and carries on.

With the GPU on, the oversampling slider grows two more steps past 8x. They run on the card only, and they are a property of this computer rather than the preset: the saved setting stays at 8x, and every instance whose slider sits at 8x plays at the rate you choose here, wherever the card runs the model. A model the card cannot take runs its plain 8x. The line beneath the GPU switch reports the card's timing at the chosen rate, so you can see the cost before you commit.

The tour has new cards

The guided tour now has 22 cards. The new ones walk the parts of the Explorer that grew this round: the three ways the left column sorts a library and the solo and hide scope on top of them, search, favorites and the Sort button, and a card that opens a real Similar list so you can see the pinned reference and the one-click compare. Further on, a card switches the delay to Reverie and points at AFTERGLOW, one opens the MIDI CONTROL list itself and explains learn, the action groups, the sweeps and the freeze switch's long hold, and a CABS card covers a cab per side and the full-rig rule. Every card you approved before reads exactly as it did.