VapourWiki — Audion VS Engine presets

The 28 presets of Audion VS Engine in 4 palettes: which one for which footage, what each does, its key settings and the encoder that suits it. In the window (Start.exe) a preset is a switch in a group of its palette; in the console it is vs_cli.cmd run --operation <id> --set <field>=<value>. How to work in the window and the console — USER_GUIDE_EN.md; every field with its range and choices — OPERATIONS_EN.md.

Which preset for which task

MLRT marks a neural preset: it needs vs-mlrt (section «The neural presets and MLRT»).

Chains

One run takes one preset. A chain is several runs one after another: the result of a step lands in <Target>\<Palette>\<preset>_<encoder>\, and that folder is the Source of the next step. Write the steps in between with prores_lt: a lossy file between steps passes its losses on to every next step. The final encoder goes on the last step only. A step run often is worth saving as a profile (section «PROFILES»).

Grain and film looks go after the upscale: grain laid before it is enlarged together with the picture.

PRECISION

Technical preparation: denoise, deband, grain. The groups follow the order of the work: Denoise, then Deband, then Master, which does both and lays grain. The grain of every preset starts from a fixed seed: one file gives the same frames on every run.

Denoise

Mild Denoise (mild_denoise)

DFTTest, a spectral denoiser, on the CPU: the same result on any machine, no video card or OpenCL needed. The first pick when little is known about the footage: light, even noise, a small cleanup at low risk.

Encoder: h264_crf14 for the archive, h264_crf17 for a final file.

Shadow Denoise SOTA (shadow_denoise_sota)

BM3D on 32-bit float cleans the whole frame; a smooth luma mask then keeps the cleaned luma in the shadows only, and the midtones and highlights stay as they were. The colour planes are cleaned everywhere: colour noise is ugly at any brightness. A micro-grain goes back over the cleaned shadows so that they do not look plastic. For night and low-light digital footage whose noise sits in the dark. BM3D runs on the video card or on the CPU, as the PROCESSING block says (section «BM3D on the GPU»).

Encoder: h264_crf17 or prores_lt.

Chroma Cleanup (chroma_cleanup)

DFTTest on the two colour planes only. The luma is not touched, so the detail stays as it was. The colour takes a heavier sigma than the luma without a visible loss, as it is stored at a lower resolution. For colour blotches and a dirty blue channel: old compact cameras, low-bitrate AVCHD.

Encoder: h264_crf17.

Deband

Deband Safe (deband_safe)

neo_f3kdb with light settings on 16 bit, no grain added: the steps in skies, walls and night light go, the edges stay sharp.

Encoder: h264_crf17 or h265_crf17. Both write 10 bit, so the smoothed gradients do not band again.

Deband Fine Grain (deband_fine_grain)

The same neo_f3kdb, then AddGrain lays a fine grain over the smoothed gradients. The picture keeps a natural texture instead of the too-clean look of cheap compression.

Encoder: prores_lt for grading.

Master

Filmic Rebuild (filmic_rebuild)

Three stages in one run: BM3D takes the digital noise out, neo_f3kdb the banding, then grain is laid by brightness — heaviest in the shadows, as on real film, lighter in the midtones, lightest in the highlights. The zones fade into each other over 0.1 of brightness. One preset for filmic footage. BM3D runs where the PROCESSING block says.

Encoder: prores_lt or h265_crf17.

Archive Clean (archive_clean)

A light DFTTest and a light neo_f3kdb, no grain: noise and banding go, no texture is added. A neutral flat master for long storage, delivery or a clean intermediate. Grading works better on footage with grain: when grading is planned, take pregrade_prep or filmic_rebuild.

Encoder: prores_422hq or prores_422hq_mxf. The built-in profile archival_master writes prores_422hq.

Pregrade Prep (pregrade_prep)

A very light DFTTest and nothing else: no deband, no grain. The footage goes to Resolve or Premiere almost as it was, only without the edge of the shot noise, so the colourist does not grade on top of it. Deband is left out on purpose: it can fight the colour tools; deband after the grade if it is needed.

Encoder: prores_422hq (the built-in profile resolve_handoff) or prores_lt_mxf for Premiere and Avid.

FILM LOOKS

The looks of film stocks. Every look is built of the same parts: MTF softening, a halation of the highlights, grain, a gamma curve, lifted blacks, desaturation. intensity scales the whole look; it stands on every page of the palette.

Film stock

Film 35mm (film_35mm)

Kodak 35 mm with the stock to choose. The stocks differ in grain, contrast, halation and colour cast: 250D — daylight, neutral, fine grain; 500T — tungsten, faster, a moderate grain, a cooler tilt; 50D — daylight, the cleanest grain.

Encoder: prores_lt or prores_lt_mxf. The built-in profile filmic_warm_35mm is 250D at intensity 1.0 into h264_crf14.

Film 16mm (film_16mm)

The organic 16 mm feel: softer than 35 mm, as a smaller frame resolves less; more visible grain; a stronger halation; lifted blacks. For a documentary or indie look on digital footage. One field: intensity.

Encoder: prores_lt.

Super 8 (super8)

The most extreme of the stocks: the heaviest grain, the softest optics, the strongest halation, lifted blacks, a lowered contrast — a home movie of the 1970s. For music videos and stylised inserts. One field: intensity.

Encoder: h264_crf17: the grain is part of the look, and a lossy final file keeps it.

IMAX 70mm (imax_70mm)

Large format 65/70 mm, the opposite of Super 8: the resolution is kept, there is almost no grain (on a large frame each grain is tiny against the picture), a gentle halation touches the brightest highlights only, and a barely visible gate weave takes away the too-clean digital look. For shots that should feel epic rather than filmy.

Encoder: prores_lt or h265_crf17.

Look

Cinematic (cinematic)

A universal subtle film look, not tied to a stock: a mild MTF softening, a light bloom of the highlights, a fine grain. No heavy grain, no gate weave, no flicker. A safe first look for any modern digital footage. One field: intensity, which scales the grain, the halation and the softening together.

Encoder: prores_lt or h264_crf17.

Bleach Bypass (bleach_bypass)

The look of a print whose silver was not bleached out: a black-and-white image lies over the colour one. High contrast, silver greys, hot highlights, little colour. A strong stylistic choice. One field: intensity — higher gives more contrast, less colour and a little more softness.

Encoder: prores_lt.

Anamorphic Scope (anamorphic_scope)

The cinemascope look of anamorphic lenses: long horizontal blue streaks from the bright highlights only, and cool teal shadows. The work is done in 16-bit RGB, so the streaks do not clip: a highlight mask, a horizontal blur of the flare length, a blue tint, added over the picture — a streak only adds light. Then a teal shift in the shadows (brightness below 0.4) and a light grain. The frame is not cropped to 2.39:1: a true scope frame is cropped in the edit.

Encoder: prores_lt for the colourist, h265_crf14 for the final file.

RETRO

The analog character of video and photo. Every preset has intensity, as in FILM LOOKS: 0.0 … 2.0, default 1.0.

Video

VHS CRT (vhs_crt)

Composite video, a VHS tape, a CRT TV: a heavy sideways bleed of the colour (the signature of the look), soft optics, a slight misalignment of the colour, mild noise. One field: intensity. The built-in profile vhs_dreamy runs it at 1.2.

Encoder: h264_crf17.

Camcorder 90s (camcorder_90s)

A consumer camcorder of the 90s: softer than VHS, slightly overexposed, a mild colour bleed and tape noise — less a broken signal, more a home video. One field: intensity.

Encoder: h264_crf17.

Photo

Polaroid (polaroid)

A Polaroid SX-70 instant photo of the 70s and 80s: soft optics, a candy bloom of the highlights (even bright skin glows a little), blacks that never reach black, warm cream midtones, a little less colour, dark corners. Softer and less grainy than Super 8.

Encoder: h264_crf17.

RESTORATION

What editors do coarsely or not at all: deinterlacing, inverse telecine, denoise along the motion, compression artefacts, scaling, new frames. Four presets of the palette run on neural networks (MLRT).

Interlace

QTGMC Deinterlace (qtgmc_deinterlace)

QTGMC of havsfunc: every field is enlarged by NNEDI3 and smoothed in time along the motion MVTools finds. The result is progressive frames without combing, flicker or ghosting on fast pans. For DV, HDV, captured VHS, S-VHS and U-matic, broadcast TS. The preset does not denoise: denoise as the next step.

Encoder: prores_lt for further work, h264_crf17 for a final master.

TIVTC IVTC (tivtc_ivtc)

Inverse telecine: TFM matches the film fields back into frames, TDecimate drops the repeated frame of every cycle. NTSC 29.97 fps with 3:2 pulldown goes back to the progressive 23.976 fps film master. For NTSC DVDs, broadcast prints, digitised film. The frame count drops by rdrop frames of every cycle: by a fifth at the defaults; the length of the clip stays. A source below 29 fps is refused: it runs at film rate already, and decimation would drop every fifth real frame.

Encoder: prores_422 or prores_422hq_mxf.

Noise

MVTools MCDegrain (mvtools_mcdegrain)

MVTools follows every block across the frames before and after, and MDegrain averages each pixel with its neighbours along the motion. The noise goes; still texture stays sharp, and so does moving texture, as it is averaged along its motion, not across it. For high-ISO and low-light footage where skin and fabric must keep their texture.

Encoder: prores_lt or h265_crf17.

DPIR Denoise (dpir_denoise)

A neural preset: it needs MLRT. DPIR, a deep denoise network, keeps the texture of skin and fabric on very dirty footage where BM3D and DFTTest smear it: ISO 6400 and higher, a phone in the dark. The network is told how heavy the noise is (strength). The frame size and the frame count stay.

Encoder: prores_lt for the colourist, h264_crf14 for the final file.

Artefacts

Deblock H.264 Artefacts (deblock_h264_artefacts)

Deblock_QED of havsfunc, for the artefacts of block compression: the edges of 8×8 blocks, ringing around edges, mosquito noise. It finds the real edges first and smooths the borders of the blocks only, so the true detail stays. For old web rips, messenger re-encodes (WhatsApp, Telegram), low-bitrate SD broadcast.

Encoder: h264_crf17.

Derainbow Decross (derainbow_decross)

The rainbow, dot crawl and cross colour of NTSC composite are a fine colour flicker that moves from frame to frame. The motion is found on the luma, and the colour planes are averaged along it over 3 frames — the frame, the one before and the one after. The moving artefacts cancel out, the real colour stays. The luma is not touched, so the sharpness stays. For digitised VHS, U-matic, BetaSP, broadcast masters, S-Video into SDI.

Encoder: prores_lt or h264_crf17.

Dehaze Local Contrast (dehaze_local_contrast)

Local contrast on the luma alone: the picture minus its blur is added back, weighted by a brightness mask that peaks in the midtones and fades out at black and white. The highlights do not burn, the shadows do not go muddy. The colour is not touched, so it does not shift; the math is 16-bit, without halos. For haze and a flat low-contrast picture.

Encoder: prores_lt for the colourist, h264_crf17 for the final file.

Scale

NNEDI3 Upscale 2X (nnedi3_upscale_2x)

NNEDI3 (znedi3, through nnedi3_resample) doubles the width and the height: clean edges, the same result on every run, no network to invent texture, no MLRT needed. For SD and HD archive that must be enlarged predictably. 1920×1080 becomes 3840×2160.

Encoder: prores_lt for handoff, h264_crf14 or h265_crf14 for the final file.

Real-ESRGAN 2x (vsmlrt_realesrgan_2x)

A neural preset: it needs MLRT. Real-ESRGAN doubles the width and the height (1080p into 3840×2160) with cleaner, sharper lines than Lanczos or Spline36. Its models in vs-mlrt are made for drawn animation: on live action they can over-sharpen and smooth the texture away — for live action nnedi3_upscale_2x (no MLRT needed) is the safer upscale.

vs-mlrt carries no general model of Real-ESRGAN (general-x4v3, general-wdn-x4v3): the preset does not offer them, and a model the installed vs-mlrt lacks is refused by name, never run as another.

Encoder: prores_lt for further work, h265_crf14 for the final file.

Soft HD Rebuild 2x (soft_hd_rebuild_2x)

A neural preset: it needs MLRT. For footage that is HD by size but holds far less detail: soft optics, pixel binning, a strong anti-aliasing filter, old codecs, generations of archive copies. A light DFTTest cleanup comes first, so that the network does not enlarge the noise and the blocks; then Real-ESRGAN at twice the width and height; then a careful sharpening of the luma.

Encoder: prores_lt for grading, h265_crf14 for the final file.

PROFILES

A profile is a preset with its settings and its encoder, saved under a name in config\profiles. It runs on a file or on a whole folder and keeps a step of a chain ready.

Run a profile (profile_run)

The PROCESSING block stands on the page and moves whatever the preset of the profile has. The result goes into <Target>\Profiles\<profile>_<encoder>\.

Save a profile (profile_save)

The button «Save a profile» on the page of a preset saves the settings of that page and the encoder of the cards.

The built-in profiles:

Denoise strength by measured noise

Seven denoise presets — mild_denoise, chroma_cleanup, archive_clean, pregrade_prep, shadow_denoise_sota, filmic_rebuild, mvtools_mcdegrain — measure the noise of the source before they start (noise=auto, the default):

• 5 frames spread over the clip, flat areas only: texture, edges and areas within 3 % of black or white are left out;
• the result is a standard deviation in 8-bit codes, the luma and the colour each their own.

Their settings (strength, sigma, thsad) are made for noise of about 2 codes. Measured noise above that raises them: BM3D's sigma and MVTools' thresholds grow with the noise, DFTTest's sigma with its square. A setting is never lowered, so clean footage gets exactly what the setting says. MVTools' scene-change threshold rises with the noise too: on heavy noise every block would exceed the default one, the whole frame would count as a new scene and be left unfiltered.

noise=fixed (--set noise=fixed) takes the settings as they are. What was measured and what the preset made of it go into the report of the run (the field noise of each file), for example:

copy[NOISE] mild_denoise: noise 5.34 / 3.44 / 2.75 (8-bit, Y/U/V) -> DFTTest sigma 57.1

BM3D on the GPU

shadow_denoise_sota and filmic_rebuild are mostly BM3D. bm3dvk runs BM3D on a Vulkan device of the VapourSynth core — NVIDIA, AMD or Intel, with the ordinary graphics driver; no CUDA Toolkit is needed. Without a Vulkan device BM3D runs on the CPU (bm3d).

The PROCESSING block (vs_accel) of these two presets:

• On the video card two runs of one file agree within about 65 dB PSNR: the card sums in a varying order. bm3d on the CPU gives the same frames byte for byte on every run.
• The other filters run on the CPU, and the video card does not help them; so PROCESSING is shown only where it moves something: the two BM3D presets, the neural presets and profiles.
• SERVICE → «BM3D: video card or CPU» (bench_gpu) measures frames a second of each on the first video of the Source; «Doctor» makes a live call and names the device. For a bare vspipe run of a .vpy, AUDION_VS_BM3D=vulkan or AUDION_VS_BM3D=bm3d names one of them.

The neural presets and MLRT

Three presets run on vs-mlrt, neural networks inside VapourSynth: dpir_denoise, vsmlrt_realesrgan_2x, soft_hd_rebuild_2x. RIFE Interpolation is out: on an RTX 5070 TensorRT builds wrong frames between the frames, DirectML is open only through ONNX Runtime of the old plugin interface, which VapourSynth R80+ does not load, and on the processor a new 4K frame takes 1.28 s — about half an hour for a minute of 24p doubled to 48p (Decisions).

• Installation. The program looks for an NVIDIA RTX card with nvidia-smi. With one, it offers MLRT among the additional components at the start and in SETUP (install_mlrt): TensorRT, ONNX Runtime and the models, about 3.5 GB to download, trimmed to this card. MLRT goes into the folder of the engine; nothing is written into Windows. Until it is installed, the page of a neural preset says MLRT IS NOT INSTALLED.
• Without an RTX card and without MLRT the window does not show these presets, and the console refuses them before the start with [PREFLIGHT] FAIL MLRT (exit code 2). MLRT installed on a machine with RTX keeps working when the folder moves to another machine: there the networks run on DirectML or on the CPU.
• Where the network runs — the PROCESSING block (vs_accel):

• The first run of a model builds its TensorRT engine: 30–120 s. The engine is kept on disk, and later runs start at once. A new tile size or a new frame size builds a new engine.
• The tile is the piece of the frame the network takes at a time. A smaller tile needs less video memory and runs slower: when the card runs out of memory, lower tile. tile_pad overlaps the pieces so that their seams do not show.
• Size and frames. vsmlrt_realesrgan_2x and soft_hd_rebuild_2x double the width and the height. dpir_denoise keeps both.

copyvs_cli.cmd run --operation vsmlrt_realesrgan_2x --set model=animejanaiV2-L2 --set vs_accel=gpu --encoder prores_lt

Choosing an encoder

The quality ladder is the same for the software encoders, NVENC, QuickSync and AMF: 14 → 17 → 21, three steps that are clearly apart. The window's default is x.264 on the CPU at 14 in MP4, that is h264_crf14.

x.264 and x.265 on the CPU write 10-bit 4:2:0, in MP4 or MKV. A hardware encoder this machine has not passed (SERVICE → «Hardware capabilities») is off in the window.

Rule. For any step that is not the last — prores_lt or prores_lt_mxf, or DNxHR where the receiving program prefers it. For the final file — h264_crf14 or h264_crf17 (software, the best quality per bit), or the hardware ladder of the machine (nvenc, qsv, amf) when time matters.

Hardware or software

NVENC on RTX 40 and RTX 50 cards gives quality comparable to libx264 medium at the same CQ. On older cards (GTX 10, RTX 20) the software CRF is slightly more efficient at the same quality, and NVENC is still many times faster.

Audio

Only the video goes through the preset; the audio is handled by itself:

• the audio of the source is copied as it is;
• audio that MP4 does not take as it is — PCM, for example, of a ProRes source going into x.264 or x.265 — is encoded to AAC at 384 kbit/s;
• into MXF the sound goes as PCM 24 bit at 48 kHz: its muxer writes no other.

The output file carries the colour range of the source (-color_range): a full-range source stays full range.