mAInframe Documentation
mAInframe

AI style transfer for LightWave 3D. Capture a preview from Layout, stylize a frame with an AI image model, then render the whole clip in that style with a video diffusion engine - all without leaving LightWave.

Native to LightWave

Runs as a Layout plugin. Capture straight from your viewport, with camera selection and frame range control.

Three Video Engines

TeleStyle, Wan 2.2 VACE and LTX 2.3, each tuned for different speed, quality and VRAM budgets.

Self-Contained

Ships with its own Python, ComfyUI runtime and ffmpeg. One installer, no system-wide dependencies.

Live Feedback

Latent previews during image and video generation, live console, progress and GPU/VRAM monitoring.

How it works

mAInframe is a guided four step pipeline:

  1. Capture - record a preview animation of your scene from Layout.
  2. Style - pick one frame and stylize it with the AI image engine using a preset or your own prompt.
  3. Video - render the full captured clip, using your styled frame as the look reference.
  4. Output - review the result side by side with the input and save it.

Requirements

ComponentRequirement
Operating systemWindows 10 or 11, 64-bit
GPUNVIDIA GPU. 16 GB VRAM or more recommended; cards down to 6 GB work with the Low VRAM mode (see the Low VRAM Guide)
LightWaveLightWave 3D 2018 or newer
Disk spaceAbout 90 GB free including downloaded models
InternetRequired once, for the installer and model downloads
WebView2Microsoft Edge WebView2 runtime (preinstalled on current Windows versions)

Installation

Everything ships in a single folder. After unzipping you will see:

mAInFrame\
  INSTALL.bat          <- run this once
  mAInFrame.p          <- the LightWave plugin
  README_FIRST.txt
  USER_GUIDE.md
  CUI\                 <- AI runtime (embedded Python + ComfyUI)
  server\              <- local UI server
  models\              <- AI models live here
  input\               <- frames/videos handed to the AI engine
  output\              <- raw engine renders
  temp\                <- working files
  1. Unzip the whole folderPut the mAInFrame folder somewhere permanent, for example C:\mAInFrame. Do not move individual files out of it.
  2. Double-click INSTALL.batThis downloads and installs the Python packages the AI engine needs (about 5 GB). It takes 10 to 30 minutes depending on your connection, and you only ever do it once. It is safe to re-run if it gets interrupted - it picks up where it left off.
  3. Add the plugin to LightWaveSee LightWave Setup below.
  4. Download the AI modelsSee Downloading Models below.

Different CUDA version? The installer defaults to CUDA 12.6 wheels of PyTorch. To target another build, open a command prompt in the mAInFrame folder and run set CUDA_VARIANT=cu121 (or cpu, cu118, ...) before launching INSTALL.bat.

LightWave Setup

  1. Open LightWave Layout
  2. Add the plugin fileMenu: Utilities > Add Plugins, then pick mAInFrame.p from your install folder. LightWave will report several plugins added - that is normal, the file contains the whole plugin family.
  3. Add the master pluginGo to Utilities > Master Plugins and add mAInframe. This keeps the plugin attached to your scene.
  4. Launch the UIRun mAInframe Launcher from the plugin list. The main window opens and the local engine starts automatically.

If the launcher tells you to add the master plugin first, do step 3 and try again.

Downloading Models

Models are not bundled - you download them once through the built-in Model Manager.

  1. Open the Model ManagerRun the mAInframe Models plugin from LightWave, or open it from the main window.
  2. Download what you needModels are grouped by engine (Image, TeleStyle, Wan 2.2, LTX, shared components). Each group has its own Download All button, or grab individual files. Downloads resume automatically if interrupted.
  3. Copy any manual modelsA few models marked copy manually came with your purchase download. Copy those files into the models folder shown at the top of the Model Manager window.

A green dot next to every model in a group means that engine is ready to use. You only need the groups for the engines you plan to use - if you never render with LTX, you can skip its models entirely.

The models folder location can be changed in Settings. Handy if you keep large model files on a separate drive.

Quick Start

The main window walks you through four numbered tabs. The short version:

  1. CapturePick a camera, set your frame range with the timeline handles, hit Capture. LightWave renders a preview and mAInframe grabs it.
  2. StyleChoose a style preset (or write a prompt), scrub to the frame you want as your look reference, click Generate Preview. Repeat until you love the frame.
  3. VideoPick an engine, set the length with the frames slider, click Render Video. Watch the latent preview as it cooks.
  4. OutputCompare input and output players, then save the result as MP4 or PNG sequence.

Step 1 - Capture

The Capture tab records a preview animation from LightWave and loads it into mAInframe.

Controls

ControlWhat it does
CameraChoose which camera in your scene to capture from. Active uses the currently selected camera.
CaptureStarts the capture. LightWave runs MakePreview, mAInframe converts the result to MP4 and extracts the first frame.
Timeline handlesDrag the in and out markers to set the frame range you want to work with. The frame count here drives the Video tab length slider.
Width / Height / FPSScene values read from LightWave, shown for reference.

Auto-capture on launch can be enabled in Settings, so opening mAInframe immediately grabs a fresh preview of your scene.

Step 2 - Style

The Style tab generates the styled reference frame that defines the look of your final video. This is where you iterate: tweak prompt, regenerate, compare, repeat.

Picking a source frame

  • Use the frame scrubber to pick any frame from your capture.
  • Or bring your own image: drag and drop, click to browse, or paste from the clipboard.

Styling controls

ControlWhat it does
Preset21 built-in style presets (Anime, Cinematic Photo, Cyberpunk, Oil Painting, Watercolor, Pixel Art and more).
PromptYour own style text. Combined with the preset prompt if one is selected.
W / HGeneration resolution, rounded to multiples of 16. Scene resets to your scene size.
Steps / CFGInference steps and guidance strength for the image engine.
Seed-1 picks a random seed each run. Set a fixed value to reproduce a result.
Generate PreviewRuns the image engine. A live latent preview appears while it generates.

Comparing results

Tick the Compare checkbox under the preview to enable the swipe slider - drag it across the image to wipe between the original frame and the styled result. Untick it to view the styled frame on its own.

Happy with the frame? Click Next to move to the Video tab. You can also Save Image to keep the styled frame itself.

Step 3 - Video

The Video tab renders your full captured clip in the style of the reference frame, using one of three engines.

Common controls

ControlWhat it does
EngineTeleStyle, Wan 2.2 VACE or LTX 2.3. See Video Engines for how to choose.
Prompt / NegativeOptional text guidance for the video stage.
Steps / CFG / Shift / DenoiseEngine sampling parameters. The defaults are tuned per engine; LTX always uses its fixed 8-step distilled schedule.
Max Length (frames)Slider capped at your captured frame count. At maximum it renders the whole clip; pull it down to render a shorter test first.
Seed-1 for random.
Render VideoStarts the render. A latent preview card appears in this tab and updates live during sampling.

TeleStyle options

ControlWhat it does
PrecisionBF16 (default) or FP16 model precision.
SchedulerSampling scheduler: DPM++ (default), UniPC or FlowMatchEuler.
AccelerationAttention backend: SageAttention (default, fastest), Flash Attention, Memory Efficient or Default.
Fast ModeReduced-step fast path. On by default; turn off for maximum quality.
Enable TilingTiled VAE processing - slower but uses much less VRAM. Forced on automatically in Low VRAM mode.

Wan 2.2 VACE options

ControlWhat it does
TypeMotion guidance source: None (original frames), Depth (DepthAnything), Depth (Lotus) or Pose (DWPose).
Ref StrengthHow strongly the styled frame drives the look. Higher is more stylized, lower preserves more of the original.

LTX 2.3 options

ControlWhat it does
Guide StrengthDepth guide influence (IC-LoRA). Controls how tightly motion and structure follow the input.
Ref StrengthFirst-frame styled image conditioning strength.
Process WidthInternal processing resolution. The result is upscaled back to your target size at the end.
Use ReferenceToggle the styled frame conditioning on or off.

Test short first. Pull the Max Length slider down to 8 to 16 frames for a quick look test before committing to a full-length render.

Step 4 - Output

The Output tab plays your captured input and the rendered result side by side.

  • Save Output - copy the rendered video to a folder of your choice, as MP4 or as a PNG sequence (optionally numbered with your timeline frame numbers).
  • Save Input - keep the captured preview clip as well.

Everything you render also appears in the Gallery.

The Plugin Family

Adding mAInFrame.p registers several plugins in LightWave. Each opens in its own window, so you can keep the console on a second monitor while working in the main UI.

PluginWhat it opens
mAInframeThe master plugin. Add this to Master Plugins once per scene.
mAInframe LauncherThe main four-tab workflow window.
mAInframe SettingsStandalone settings panel: VRAM mode, server control, models folder and more.
mAInframe ModelsThe Model Manager in its own window.
mAInframe GalleryBrowser for everything you have generated.
mAInframe ConsoleLive engine log in its own window.
About mAInframeVersion and author information.

Settings

Open mAInframe Settings from LightWave. Changes apply immediately.

Performance

SettingWhat it does
VRAM ModeAuto detects your GPU and chooses sensible behavior. High keeps models resident for maximum speed (24 GB cards). Normal balances speed and memory. Low enables aggressive offloading, tiled VAE and resolution caps for 6 to 12 GB cards.
Use SageAttentionFaster attention kernels for sampling and VAE work. On by default.
Latent previewLive preview frames during image and video generation. On by default; turn off to squeeze out a little extra speed.
VRAM Optimization (Unload Models)Frees model memory between pipeline stages. Helpful on smaller cards.
Clear VRAMButton that immediately unloads everything from GPU memory.

Server

SettingWhat it does
Keep server runningLeaves the engine alive when you close the UI window, so reopening is instant. The server always shuts down with LightWave.
Start / Kill ServerManual control over the background engine process.

Workflow

SettingWhat it does
Models DirectoryPoint mAInframe at a custom models folder, for example on another drive.
Auto-Capture on LaunchGrab a fresh preview automatically every time you open the UI.
Save Styled Previews ToAuto-save every generated styled frame (styled_001.png, styled_002.png, ...) to a folder. Optionally save the raw source frames too.

Console

The Console plugin shows the live engine log: model loading, sampling progress, VRAM events and any warnings or errors. When reporting a problem, this is the first thing to screenshot.

  • Autoscroll is on by default - the log follows the newest line. Toggle the autoscroll button off to scroll back through history while a job is running.

Video Engines

Three engines, three different trade-offs. You can switch per render - the styled frame works with all of them.

EngineCharacterBest for
TeleStyleFast, faithful to the input motion, light on VRAM.Quick turnarounds, longer clips, smaller GPUs.
Wan 2.2 VACEHighest stylization quality, with selectable motion guidance (depth or pose). Heaviest of the three.Hero shots where look matters most. 16 GB+ recommended at full resolution; works on smaller cards in Low VRAM mode at reduced resolution.
LTX 2.3Strong quality at speed, fixed 8-step distilled sampling, depth-guided.The middle ground: better quality than TeleStyle, faster than Wan.

Low VRAM Guide

mAInframe runs on cards down to about 6 GB, with some patience. Here is how to get the best results on a small GPU:

  1. Set VRAM Mode to LowIn Settings. This enables tiled VAE encoding and decoding, CPU offloading for text encoders, full transformer block swapping and automatic resolution caps. On cards under 12 GB, Auto mode applies most of this on its own.
  2. Render fewer framesUse the Max Length slider in the Video tab. VRAM use grows with frame count; a 16-frame test tells you if the look works before you commit.
  3. Keep resolutions modestIn Low mode, Wan renders are automatically capped at 832 px on the long edge and upscaled afterwards. For LTX, reduce Process Width.
  4. Prefer TeleStyle or LTXThe Wan 2.2 stack is a 14B model and the most demanding path. TeleStyle is the lightest.
  5. Close other GPU appsBrowsers, viewport-heavy DCC apps and even Windows itself compete for VRAM. Every few hundred MB counts on a 6 GB card.

Expect low VRAM renders to be slower - block swapping trades speed for memory. That is the deal that makes a 14B model run on a 6 GB laptop at all.

Files & Folders

Everything mAInframe creates while you work lives in the temp folder inside the install:

FileContents
temp\preview_anim.mp4The captured LightWave preview
temp\preview_input.pngSource frame used by the Style tab
temp\preview_styled.pngYour generated styled reference frame
temp\render_output.mp4The rendered video, before you save it
temp\server.logEngine log - attach this when contacting support

Raw engine renders are kept in output\mAInFrame at the install root, and the frames/videos handed to the engine in input\. The models folder (or your custom models directory) holds the downloaded AI models. The CUI and server folders are the runtime - you never need to touch them.

Troubleshooting

The launcher says to add the master plugin first

Add mAInframe under Utilities > Master Plugins, then run the launcher again.

The UI window does not open

  • Confirm the WebView2 runtime is installed (it ships with current Windows; Microsoft offers a standalone installer).
  • Confirm you are running the plugin from inside the full mAInFrame folder - the .p file cannot work alone.
  • Run INSTALL.bat if you have not yet.

Capture fails

  • Check that LightWave can create a preview normally from the same scene.
  • Make sure only one mAInframe instance is running.
  • Re-run INSTALL.bat if ffmpeg is missing (it is installed as part of the package step).

Generation fails with missing models

Open the Model Manager. Any red entries in the group for your chosen engine need downloading. Remember the manual-copy models from your purchase download.

Render fails with out-of-memory (OOM / CUDA error)

  • Set VRAM Mode to Low in Settings and try again.
  • Reduce the Max Length slider and the resolution.
  • Click Clear VRAM in Settings before retrying, or restart the server.
  • See the full Low VRAM Guide.

Render fails even though capture worked

The video stage needs both a capture and a styled frame. Generate a preview in the Style tab first.

It works, but everything is slow

  • Check the VRAM Mode - Low mode on a big GPU leaves performance on the table. Use Auto or High.
  • Make sure SageAttention is enabled in Settings.
  • Check Task Manager (Performance tab, your NVIDIA GPU) to confirm the dedicated GPU is being used and not starved by other apps.

Still stuck?

Contact support with a screenshot of the Console and the file temp\server.log.

FAQ

Does it work without an internet connection?

Yes, after the one-time install and model downloads everything runs fully local. No cloud, no uploads.

Can I use my own image instead of a captured frame?

Yes - drag and drop, browse or paste any image into the Style tab.

Do I need all the models?

No. The image engine models are required, but you only need the video engine groups you actually use.

Can I keep models on another drive?

Yes. Set the Models Directory in Settings and mAInframe will look there.

Why does my first render take longer than the next ones?

The engine loads models on the first job and keeps them warm afterwards. Enable Keep server running in Settings to stay warm between sessions too.

Is my GPU supported?

Any reasonably recent NVIDIA GPU works. 16 GB VRAM or more gives the smoothest experience; 6 to 12 GB cards should use Low VRAM mode and expect longer render times.