Wiki

Key ideas

A few ideas explain most of the settings in mAInframe. Read these once and the rest of the guide makes sense.

How an AI image is made

The model does not paint the picture from left to right. It starts with a frame of pure random noise, like TV static, and removes a little of that noise at a time. Each pass it guesses what the picture "should" look like, guided by your prompt and your capture, and nudges the noise toward that guess. After enough passes the static has become a clean image. Video works the same way, for every frame at once.

Each of those passes is a step. The method used to take each step is the sampler. The random static it starts from is set by the seed.

Your capture is the guide

AI Render does not invent a new scene. It keeps the layout, camera move and shapes of your LightWave shot and repaints the surface: materials, lighting, look. The capture passes (Beauty, Clay, Normal, Depth, Outline) are different ways of describing your shot to the model. Depth tells it what is near and far. Outline gives it the edges. Normal gives the direction each surface faces. Beauty is the full shaded picture, so it carries your colours and lighting too, which can be good or can stop the style from taking over.

• PromptA description of the look you want, in plain words. Describe materials, lighting, era and mood ("rusted steel, overcast light, 1970s film photo"). You do not need to describe what is in the shot; the capture already supplies that.
• Negative promptThings you want kept out, such as "blurry, text, watermark, extra fingers". It pushes the result away from those words. Leave it short; a long list rarely helps.
• SeedThe number that creates the starting static. Same seed + same settings = the same picture every time, so you can change one thing (the prompt, say) and compare fairly. A different seed gives a different take on the same idea. When you get a result you like, keep its seed. -1 or Random means "surprise me" each run.
• StepsHow many clean-up passes the model makes. Too few and the image looks smeary or unfinished. More steps add refinement, but past a point nothing improves and it only takes longer. The models in mAInframe are tuned for few steps (Flux Klein looks finished at about 4), so the default is usually right. Raise it a little if fine detail looks mushy.
• CFG (guidance)How hard the model is pushed toward your prompt. Low values give a natural, relaxed result that may drift from the prompt. High values follow the words more literally but go harsh: oversaturated colours, crunchy edges, burnt-in contrast. The fast models used here are built to run at about 1, so change it in small amounts.
• SamplerThe maths used to take each step from noise toward the image. They all reach a similar picture; they differ in speed, sharpness and how many steps they need. See the sampler list below.
• DenoiseHow much of the starting picture is replaced. 1.0 starts from pure static and only your capture guides it, so the style can change everything. Lower values start from the existing image and only repaint part of it: 0.3 is a light touch-up, 0.7 a strong restyle that still keeps the original's colours and details.
• LoRAA small add-on file that teaches the model one specific thing: an art style, a character, a type of product. Adding one makes that look much easier to get than with words alone. Each LoRA works with one model family only (an image LoRA will not work on a video engine). Too many at once tends to muddy the result.
• Weights / modelsThe trained "brain" files the AI runs on. They are large (several GB) and download once in Models. A tool that says weights are missing just needs that download.
• ResolutionThe size of the picture the model draws. Models are trained around a certain size (often about 1 megapixel, e.g. 1024 x 1024 or 1280 x 720); going far above it can give repeated or doubled subjects, far below it loses detail. Sizes snap to multiples of 32 (images) or 8 (video) because the model works on a shrunk-down grid internally. For big final images, render at a normal size and upscale.
• UpscaleMakes the finished image or video larger. This is not a plain resize: an AI upscaler (SeedVR2) redraws the picture at the bigger size, adding sharp edges, texture and fine detail that were not there. That looks much better than stretching, but the new detail is invented, so tiny text or faces can change slightly. It costs extra time and VRAM.
• VRAMThe memory on your graphics card. AI models must fit in it. Bigger resolutions, longer videos and bigger models use more. If a job fails with "out of memory", lower the resolution or length, turn on tiling, or close other GPU programs.
• VAE / latentTo save memory the model does not work on real pixels. It works on a compressed version called the latent (about 8x smaller each way). The VAE is the part that converts between pixels and the latent at the start and end. "Latent preview" is a rough look at the work in progress before that final conversion.
• TilingConverts the result back to pixels in pieces instead of all at once. Uses far less VRAM on big frames or long videos, costs a little time, and on rare occasions leaves a faint seam.
• Precision (BF16 / FP16)How many digits the GPU keeps for each number. Both halve memory compared with full precision and look the same. BF16 is more stable on RTX 30-series and newer cards; use FP16 only on older cards that lack BF16 support.

Samplers, in plain words

• EulerThe simplest method: look at where the image is heading and take one straight step that way. Fast, predictable and clean. A good default, and the one the fast models are tuned for.
• "A" samplers (Euler A, DPM-2 A)"Ancestral". They add a bit of fresh random noise at every step. Results are softer and more varied, but never settle: changing the step count changes the picture. Avoid them for video or when you want to repeat a result.
• Heun, DPM-2Check their work with a second look each step. More accurate, but each step takes about twice as long.
• DPM++ 2MRemembers the previous step to take a smarter next one. Sharp, detailed results in few steps. A good choice when Euler looks a little soft.
• SDE samplers (DPM++ SDE, DPM++ 2M SDE)Add some randomness like the A samplers. Often more fine texture and detail, but less repeatable and usually need more steps.
• DDIM, LMS, DEIS, UniPC, UniPC BH2Older or alternative methods. DDIM is fully repeatable. UniPC is designed to look good in very few steps. Worth trying if a result has artefacts, otherwise not needed.

If unsure, leave the sampler on its default and change steps, CFG or the prompt instead. Those make a far bigger difference.

Start

Each card is one tool. The dark area inside a tool is the preview. It fills in after a capture, a dropped file, or a finished result.

• AccountSigns you in to your LightwaveLlama account. The license is tied to the account, not a key, so signing in on a new computer is all it takes to activate it.
• StoreWhere you buy the separate tools and plugins, and where updates for what you own appear.
• ModelsDownloads the AI weights each tool runs on. Nothing works until its weights are here; see Key ideas, Weights. You only download the tools you use.
• GalleryEverything you have saved so far, grouped into captures, styled stills and rendered videos, so you can find or reuse an earlier result.
• ProjectTies a folder to the current scene so every capture and render is saved and numbered there automatically. Set it once per scene and you never pick a save location again.
• SettingsPreferences for every tool: how much graphics memory is used, the default seed, the assistant, theme and updates. The memory settings are the ones to look at if renders fail.
• Report a bugSends a ticket straight to the developer. Including the console log shows exactly what the engine was doing when it failed, which usually makes the difference between a fix and a guess.
• All toolsBack to the tool list. Esc does the same from anywhere.

AI Render

Four steps across the top: Capture, Style, Video, Output. Do them in that order. The camera size and frame rate come from the scene.

1. Capture

• Capture previewPlays your scene camera through the frame range and records every frame as all the passes at once. It uses Layout's OpenGL viewport, not a full render, so it is quick. Your lighting and textures only need to be good enough for the AI to read shapes and colours. Capture again whenever the animation or camera changes.
• Still frameCaptures only the frame Layout is on and jumps to Style. Use it to try looks before committing to a whole shot, or when a single image is all you need.
• CameraWhich scene camera is recorded. Its framing, lens and movement become the framing and movement of the AI result, so set the shot up in LightWave first. Active follows whichever camera is current in Layout.
• TimelineSets which frames are captured. Trim to just the part you need: every captured frame is a frame the video engine has to render later, so shorter captures render much faster. I / [ and O / ] set the in and out at the current frame.
• PlaybackPlays the capture so you can check it before styling. Look for viewport clutter or popping objects; the AI will faithfully copy anything that is visible.
• PassesDifferent views of the same shot that guide the AI. Beauty: your normal shaded render, so colours and lighting carry into the result. Clay: plain grey shapes, so the AI is free to invent all materials. Depth: near is light, far is dark, which locks the 3D layout. Normal: the direction surfaces face, which locks the shape and fine form. Outline: just the edges, the loosest guide. Use Clay or Depth when the style should replace your look entirely; use Beauty to keep your colours.
• AdjustOpens the capture range and the frame settings. The button text shows the current size and frame rate.
• StartFirst frame captured. Same as the in handle on the timeline.
• EndLast frame captured. Same as the out handle.
• WidthComes from your LightWave camera so the result lines up with your shot. The size the AI actually draws at is set separately on Style and Video (see Key ideas, Resolution).
• HeightAs Width: read from the camera so proportions match your shot.
• FPSFrame rate from the scene. The video engines render at this rate (MiniMax always at 24), so motion timing matches your animation.
• Hide OpenGL UIHides grids, handles, bounding boxes and other viewport overlays while capturing. Leave it on: the AI cannot tell an overlay from scenery and will paint it into the result.

2. Style

• PresetA ready-written prompt for a well-known look (anime, oil painting, film noir). It is just text: pick one as a starting point, then edit it or add your own words after it.
• Preset promptThe text of the chosen preset. Edit it freely; changes apply to this scene only. It is placed before your own line, and earlier words carry slightly more weight.
• Your promptYour own description, added after the preset. Spend words on what the capture cannot show the AI: materials, lighting, era, film stock, mood. You do not need to describe what objects are in the shot.
• Save as presetStores the current text as your own preset so you can reuse a look across scenes and keep a consistent style across shots.
• User presetYour saved looks. Picking one adds it after the built-in preset, so you can combine, for example, a built-in 'film still' with your own 'brand colours'.
• LoRAAdds an image LoRA: a small add-on that teaches the model a specific style or subject (see Key ideas, LoRA). One or two at a time works best; more tends to muddy the look.
• ResolutionThe size the still is drawn at. Flux Klein is trained around one megapixel, so sizes near that look best; far larger can produce doubled subjects. Scene matches your camera. For large output, keep this normal and use Upscale still. Sizes snap to multiples of 32 because the model works on a shrunken grid.
• Upscale stillAfter the styled still is made, an AI upscaler (SeedVR2) redraws it much larger, up to 8K, adding crisp detail. Use it for final stills or print, not while you are still trying looks: it is slow and needs its own download. The extra detail is invented, so check small text and faces.
• Upscale size (long edge)The final width (or height, if the image is tall) in pixels. 3840 is 4K. Bigger takes longer and uses more VRAM.
• StepsHow many clean-up passes the model makes (see Key ideas). Flux Klein is built to finish in about 4, the default. Try 6-8 if detail looks soft; beyond that it rarely improves.
• CFGHow hard the result is pushed toward your prompt (see Key ideas). Klein works best near 1.0. Raise it slightly if the prompt is being ignored; if colours burn or edges go crunchy, it is too high.
• SeedSets the starting random noise, which decides the particular "take". Keep a seed you like so you can tweak the prompt and compare; change it to see a different variation.
• Seed behaviorWhat happens to the seed after each render. Fixed: repeat the same take while you adjust other settings. Increment / Decrement: step through neighbouring variations one by one. Random: a new variation every time, good for exploring.
• SamplerThe method used for each step (see Samplers in Key ideas). Leave it on the default unless a result looks soft or has artefacts.
• ModelShown when more than one Flux Klein size is installed. Bigger models understand prompts better and give finer detail, but are slower and need more VRAM.
• SourceWhich capture pass guides the AI (see Passes above). This choice has a big effect on how much of your original look survives.
• BlendMixes a second pass into the guide, for example Beauty with some Depth to keep colours while firming up the 3D layout. Mix sets how much of the second pass is used.
• FrameWhich captured frame to style. Pick a frame that shows the important elements clearly, because the video engines use this still as the reference for the whole clip.
• CompareWipe between the styled still and the original capture to check what changed: did shapes drift, did details vanish, did the style take hold?
• Custom inputStyle any picture instead of the capture, for example a sketch or a photo. Edit markup lets you paint changes on it first.
• Video engineWhich AI animates the shot. TeleStyle: fastest; copies the look of your styled still across every frame and follows your motion closely; no prompt. Wan 2.2: higher quality and can be steered with a prompt; slower. LTX 2.5: carries the styled first frame through the whole clip, good balance of speed and control. MiniMax H3: learns the look from up to 5 reference stills instead of a start frame. Start with TeleStyle to check a look quickly, then switch for the final.
• First frameThe styled still that shows the video what the look should be. The engine spreads this look over the rest of the shot, so a good first frame matters more than any other video setting.
• Last frameOptional styled still for the end of the shot. Useful when the shot ends on something the first frame never shows, so the engine is not left to guess its look.
• Style referencesMiniMax H3 only. Example stills of the look you want. Several angles of the same style help it stay consistent as the camera moves.
• Style first and lastAlso styles the first and last reference stills before the video starts, so the video's start and end match the look of your preview instead of the raw capture.
• Generate previewRenders the styled still. Try a few seeds and prompts here before going to Video: a still takes seconds, a video takes minutes, and the video can only be as good as this still.
• Save imageSaves the still. With a project set, it goes into the project's styled folder automatically and is numbered so nothing is overwritten.

Markup

After a capture, Edit markup opens on the frame. Paint what should change, then Accept. Those marks are part of the style render.

• MoveMoves what you have drawn.
• SelectionLimits painting to an area so you do not spill onto neighbouring parts. Magic wand selects by similar colour; Shift adds to the selection, Alt removes.
• BrushPaints guidance onto the frame. The AI treats painted colour as an instruction: paint a red patch on a wall and the result will have something red there. Rough shapes are enough; the AI turns them into proper detail.
• EraserRemoves paint and returns that area to the capture.
• RemovePaint over an object and the AI rebuilds the background behind it, as if the object was never there.
• AI fillPaint a mask, describe what should appear there ('a neon sign', 'ivy'), and the AI paints it in, matched to the surrounding light and perspective.
• ColorThe brush colour. The swatches are sampled from this frame so painted fixes blend with the existing colours.
• Size, Hard, OpacityBrush width, edge softness and strength. Soft, low-opacity strokes are a gentle hint; hard, full-opacity strokes are a firm instruction.
• Place an imageDrops a picture onto the frame, for example a logo or a reference object. The AI blends it into the scene.
• SplineDraws clean lines and shapes (arrows, outlines) instead of freehand strokes. Fill closes the shape.
• ToleranceHow similar a colour must be to be included by the magic wand. Higher grabs more of the image.
• EdgeHow soft the edge of the selection is. A soft edge blends the change into its surroundings; a hard edge gives a crisp boundary.
• AllProtects part of the image from the brush: All paints anywhere, Background only behind objects, Foreground only on objects.
• AI fill promptWhat should appear in the masked area. Describe the thing itself, not the whole scene.
• AcceptKeeps the markup. It becomes part of the instructions for the style render on the frames it covers.
• DiscardThrows away the current edits.
• Markup tracksEach accepted edit applies over a range of frames. Drag the ends to set which frames it guides, so a fix painted on one frame can hold through a whole move.

3. Video

TeleStyle follows the styled frames. Wan 2.2, LTX 2.5, and MiniMax H3 add a prompt, as shown here for LTX 2.5. Change, on the right of the engine name, returns to the engine menu on Style.

• SourceWhich capture pass guides the motion and shapes of the video. Beauty keeps your colours and lighting in play; Depth or Clay let the styled first frame decide the look. See Key ideas, Your capture is the guide.
• BlendMixes a second pass into the guide, for example Beauty with some Depth to keep colours while holding the 3D layout firm through camera moves.
• ModeHow the two passes are combined, as in Photoshop layer modes. Blend is a straight mix. Multiply darkens (good for laying Outline lines over Beauty). Screen lightens. Overlay and Soft light add contrast from the second pass.
• StepsClean-up passes per frame (see Key ideas). Video cost grows quickly with steps, so the defaults are kept low. LTX 2.5 fixes this at 8 because it is trained for exactly that.
• CFGHow hard the video is pushed toward the prompt. Video engines are sensitive to it: keep it near the default. Too high gives flicker, harsh colour and jittery detail.
• SeedSets the starting noise for the whole clip. Keep it fixed while you tweak other settings so you can compare renders fairly.
• ResolutionThe size the video is rendered at. Video is far heavier than stills: doubling width and height means about four times the work and VRAM. Render at 720p and use Upscale video for bigger output.
• Max Length (frames)How much of your capture is turned into video. Longer clips need much more VRAM and time; if a render runs out of memory, shorten this first.
• Upscale videoAfter rendering, an AI upscaler redraws every frame larger and sharper. It invents detail, so it looks far better than resizing, but adds a lot of time.
• Upscale methodSeedVR2: smaller and quicker, cleans and sharpens. LTX detailer: a huge model that adds richer new detail, but needs about 54 GB of downloads and lots of VRAM.
• Upscale size (long edge)Final width (or height, for tall video) in pixels. 1920 is Full HD. Auto picks the next standard size above your render.
• ChangeGoes back to the engine choice on Style.
• Video referencesThe styled stills that tell the engine how the clip should look. They matter more than any other video setting: the engine spreads their look across every frame.
• LoRAAdds a video LoRA for this engine (see Key ideas, LoRA). Image LoRAs do not work here.
• Save optionsHow the result is written. MP4 is a single playable file. PNG sequence writes one lossless image per frame, which is better for compositing or bringing back into LightWave as an image sequence.
• Use timeline frame numbersNumbers the PNG files by scene frame (e.g. 0045 to 0120) instead of from 0001, so they line up with your LightWave timeline when you load them back.
• Render videoStarts the video render. This is the slow part; do a short, small test before the full render.
• Live previewA rough look at the frames while they are still being cleaned up (see Key ideas, Latent). Blurry by nature; it shows whether motion and composition are on track so you can cancel early.

TeleStyle

Engine options for TeleStyle. The prompt block stays hidden. TeleStyle follows the styled frames.

• PrecisionHow precisely numbers are stored on the GPU (see Key ideas). BF16 is the safe choice on RTX 30-series and newer.
• SchedulerThe stepping method, like a sampler (see Key ideas). FlowMatchEuler is what TeleStyle is trained with and the safest choice.
• Fast modeTakes shortcuts that make rendering much quicker with little visible loss. Turn it off only for a final render if you see small artefacts.
• Enable tilingConverts frames back to pixels in pieces to save VRAM (see Key ideas). Keep it on unless you have a lot of VRAM and see seams.

Wan 2.2

Wan shows the prompt block and an Advanced group. Shift, Denoise, and Reference strength are the sliders in that group.

• Prompt presetReady-written motion and style text. Video presets describe camera and movement; Style looks describe the appearance.
• Preset promptThe preset's text, editable. Placed before your own prompt.
• Your presetsVideo prompts you saved, for reuse across shots.
• PromptDescribe both the look and the motion: 'handheld camera, slow push-in, dust drifting in light'. Write asks the assistant to draft one from your style, still and capture. From Style copies your Style prompt as a starting point.
• Negative promptThings to push out of the video, such as "flicker, blur, distorted hands".
• ShiftDecides where the model spends its effort. Higher values spend more of the steps on big shapes and overall composition (bolder changes, steadier motion); lower values spend more on fine surface detail. The default 5.0 suits most shots.
• DenoiseHow much of the video is repainted (see Key ideas). 1.0 is a full restyle; lower keeps more of the capture's original colours and texture.
• Reference strengthHow closely the video copies the look of your styled first frame. Lower lets the prompt take over more and gives looser, more varied frames.

LTX 2.5

The styled first frame is carried through the whole capture. Steps stay an 8-step first pass plus a 2-step latent upscale. CFG stays editable and starts at 1.0.

• Guide strengthHow tightly the video follows your capture's motion and shapes. 1.0 follows it closely. Lower values give the engine freedom to change movement, which can help a style that needs to reshape things, but the result drifts from your animation.
• PromptSame prompt tools as Wan. Describe the look and any motion details the capture does not show (hair moving, smoke, light flicker).

The page also keeps an older LTX block. It is shown only when the engine value is ltx. The engine menu does not list that value. Its Advanced sliders are Guide strength, starting at 0.67, and Style strength, starting at 0.5. The note on that block reads LTX 2.5: two-stage AI rendering.

MiniMax H3

MiniMax uses style references instead of a first and last frame. Negative prompt is hidden. The note under the options says 24 fps, length snaps down to the 17k+5 grid so the whole capture is covered, resolution sets the canvas around the trained 1 megapixel frame, and Turbo renders in 4 steps.

• Turbo (4-step LoRA)A LoRA that lets MiniMax finish in 4 steps instead of about 20: several times faster with a small loss of fine detail. Keep it on while testing; turn it off for a final render if detail matters.
• X2 Detail VAEConverts the result back to pixels at twice the width and height using a decoder trained to add fine detail. Sharper output at little extra cost, without a separate upscale pass. Needs its own download.
• Reference detailFeeds your style stills through the same detail decoder so fine texture in them (fabric, brush strokes) carries into the video more faithfully. Works only with X2 Detail VAE on.
• AttentionAttention is the part of the model that relates every part of the frame to every other part, and it is where most of the time goes. These choices are different code for that same maths. Sage and Comfy Kitchen are faster; Standard is the slowest but works everywhere. Use Standard if a faster one gives an error.
• Veda sparse attentionA speed-up that skips the parts of the attention calculation that matter least (see Attention). Much faster on long or large clips, with a small risk of lost detail. It only works at the sizes it was trained for.
• Veda trained resolutionVeda only works at these five sizes, so the free size box is replaced by this list while Veda is on.
• Style referencesExample stills of the look you want. Several angles of the same style help MiniMax stay consistent as the camera moves; mixing different styles confuses it.
• PromptDescribe the look and motion. For MiniMax the preset and your line are merged into one prompt.

4. Output

• PlayPlays the result. Watch it full size at least once: flicker and swimming textures are easy to miss in a small preview.
• Result, Split, WipeCompare the render with your capture. Split puts them side by side; Wipe slides between them. The quickest way to spot drift from your animation or lost detail.
• Save captureSaves the unstyled capture, useful as a reference or for a before/after.
• Save resultSaves the rendered video. With a project set, it goes to that engine's output folder and is numbered.
• RecentEarlier renders, so you can compare versions or recover one you did not save.

3D Model Builder

Turn a picture or a sentence into a textured mesh, then send it to Layout as a LightWave object.

• From imageBuilds the object from one picture. The AI looks at it and imagines the sides it cannot see. Best input: one whole object, centred, plain background, soft even light.
• From textFirst makes a picture from your description, then builds the object from it. Handy when you have no reference image.
• Optimize for 3DAdds words to your prompt so the generated picture is one centred object on a clean background, which is what the mesh builder needs. Turn off only if you are deliberately describing that already.
• EngineThe AI that builds the shape. Trellis2: the all-rounder, clean shapes and good textures. Pixal3D: a different reading of the picture; try it when Trellis gets a shape wrong. Hunyuan3D 2.1: strong detail, but its licence excludes the EU, UK and South Korea. UniForge3D: builds from several photos of one object; not for commercial use.
• Mesh qualityPresets that set poly count, texture size and steps together. Draft is for checking quickly whether a picture will work; Standard for most uses; Detailed and Extreme for hero objects seen up close, at the cost of time and a heavier object. Custom keeps your own values.
• Poly countHow many faces the finished mesh has. More faces hold finer shape (sharp edges, small details) but make a heavier object that is slower in Layout. Background props: 10-50k. Close-up objects: 150k or more. The AI's shape has a limit to its detail, so very high counts mostly add weight, not detail.
• TexturePaints the colour onto the mesh as an image map. The size is the image width in pixels: 1024 is fine for distant objects, 4096 for close-ups. Off gives a plain grey mesh.
• UV layoutUVs are how a flat texture image is wrapped onto the 3D surface, like the pattern pieces of a sewn toy. Automatic cuts the surface into many small pieces; fine for any object, but hard to paint by hand. Organic makes a few large, even pieces; better for characters and creatures you may want to repaint later. Adds about a minute.
• Bake ambient occlusion (slow)Darkens creases and corners in the texture, where light would naturally struggle to reach. Makes the object read better under flat lighting, but the shadow is painted in permanently, so it can look wrong under your own scene lighting. Leave off if you will render with GI.
• Remove backgroundCuts the subject out of the picture first so the backdrop is not built into the mesh. Turn off only for a picture that already has a clean or transparent background.
• Project source photosCopies the real pixels of your picture onto the texture, which is sharper than what the AI paints. If the built shape does not match the photo (e.g. a leg is in a different place), the projection lands in the wrong spot; turn it off then.
• Back, Left, RightExtra photos of the same object from other sides. They fill in texture the front view cannot see, so the back is not guessed. For UniForge3D they also shape the mesh.
• AdvancedSteps and Seed work as described in Key ideas. More steps give a cleaner, more faithful shape at the cost of time; change the seed to get a different interpretation when a shape comes out wrong.
• ShadingHow the preview is drawn. Wireframe shows the polygon flow; useful for judging whether the mesh is clean enough for your use.
• GenerateBuilds the mesh. The line underneath names any weights still missing; download them in Models.
• Saves toWhere the object files are written.
• Send to LayoutConverts the result to a LightWave object and loads it. Weld points joins duplicate points so the surface is one piece (needed for smoothing and subdivision). Merge triangles turns pairs of triangles into quads, which are easier to edit. SubD mesh tries to make it usable as a subdivision surface. Diffuse 100% makes the texture show at full brightness.

3D Texture Engine

Select an object in Layout. The tool captures it from several angles, styles each view, and projects the result back onto the mesh.

• Target objectThe object to paint. It needs no UVs in advance; the tool makes them.
• Use styled stillUses your latest AI Render still as a look reference, so the object matches the style of your shot.
• Texture promptDescribe the material: 'worn red leather with brass studs', 'mossy stone'. Material words (worn, glossy, rusted) do more than colour words.
• Reference imageA picture of the material or look you want. The AI copies its colours and texture instead of relying only on words.
• Orbit viewsHow many angles are photographed and painted. More views cover hidden areas and reduce stretched or blurry patches, but each view is another AI image to make. 8 suits most objects; 12 to 16 for complex shapes with many hidden areas.
• Texture resolutionWidth of the final texture in pixels. 1024 for objects seen at a distance, 2048 for close-ups.
• Use scene cameraUses your own camera keyframes as the views instead of an automatic orbit. Use it to make sure the angles the audience will see get the best coverage.
• Also write Specularity + Bump mapsAlso writes a shininess map and a surface-relief map worked out from the painted colour, so the surface reacts to light more believably in LightWave instead of looking printed on.
• Review captures before texturingStops after photographing the views so you can check them. Catches badly framed or blocked views before you spend time on painting.
• Load textured .lwo into Layout when doneSwaps the textured object into your scene when finished.
• Use these viewsThe views look right; continue to painting.
• Retake viewsAdjust the object or camera in Layout, then photograph again.
• Generate texturesPhotographs the views, paints each one, and blends them into one texture on the object.
• 3D, Color, Spec, BumpThe result: the textured object, and each map on its own. Check the colour map for blurry or stretched patches where views did not reach.
• Import to LayoutLoads the result if it was not loaded automatically.

Auto UV

Select a mesh in Layout. Unwrap writes a new object with packed UVs and replaces the one in the scene.

• Atlas sizeUVs are how a flat image wraps onto the surface, like the pattern pieces of a sewn toy. This sets the texture size the pieces are packed for, which decides how much gap is left between them so colours do not bleed across at that resolution.
• Output nameName for the new object. The original file is kept.
• Unwrap selectedCuts the selected mesh into pieces, flattens them and packs them into one square, then swaps the result into your scene.

Motion

ARDY, MotionBricks, and UniMate are three engines. One runs at a time, and it shares the GPU with AI Render. Launch the one you opened. Esc returns to the tool list.

ARDY

• Start engineLoads ARDY onto the graphics card. It shares the GPU with AI Render, so the Style and Video models are unloaded. The first start takes minutes.
• MotionClips are spans of the timeline each with their own text prompt ('walks forward', then 'sits down'). The engine blends one into the next. Replan regenerates from the playhead after an edit, keeping what comes before.
• Restart from hereKeeps the motion up to the playhead and generates a fresh continuation. Use it when the take went wrong partway.
• Restart takeThrows the take away and starts again from frame 0.
• ExamplesInserts a sample prompt to show the kind of wording that works.
• Waypoint modeClick the floor to place a spot the character must reach. The prompt says how it moves; waypoints say where.
• Drop aheadPlaces a waypoint straight ahead of the character, a quick way to make it travel further.
• Frames aheadHow soon the next waypoint must be reached. Too few frames for the distance forces an unnatural sprint.
• Dense pathPins the route along the way, not just the end point, so the character follows the path more tightly.
• Clear constraintsRemoves all waypoints and pins so only the text prompts steer the motion.
• PosePins the body, or one hand or foot, to a position at the playhead: a hand on a rail, a foot on a step. Drag the handle into place.
• Apply pinsRegenerates so the motion arrives at the pinned pose naturally rather than snapping to it.
• Move startMoves where the character starts and which way it faces. Takes effect on Restart take.
• Reset startBack to the origin, facing +Z.
• TransportPlays the take and moves the playhead. Edits, replans and pins all act at the playhead.
• Send to LayoutWrites the take as a BVH skeleton animation and loads it into Layout. To put it on your own character, rig the character in AI Rig with the Core skeleton.

ARDY settings

The gear beside Send to Layout opens View, Model, Advanced, and Export. MotionBricks and UniMate use the same gear. The Setup notes inside it change with the engine.

• MeshShows a body over the skeleton in the preview. Display only.
• OpacitySee-through body makes it easier to judge foot contact and crossing limbs.
• SkeletonShows the bones. Or, in Model: the bone layout the motion is generated for. Core is the standard humanoid AI Rig can bind to.
• Start directionShows an arrow for the start heading.
• Camera followsKeeps the view on the character as it travels.
• Dark viewportDark background for the preview.
• SkeletonShows the bones. Or, in Model: the bone layout the motion is generated for. Core is the standard humanoid AI Rig can bind to.
• HorizonHow many frames are generated per chunk. ARDY builds motion a chunk at a time, each continuing from the last. Longer chunks are smoother but react more slowly to edits.
• AccelerationOptional speed-ups. torch.compile and TensorRT build optimised GPU code first: the first start is much slower, then generation is faster. Motion is identical.
• Load modelLoads the chosen skeleton model.
• SeedPicks one of many possible performances of the same prompt. Change it for a different variation.
• Text weightHow strongly motion follows the prompt, like CFG for images. Higher gives clearer, bigger actions; too high looks stiff or jittery.
• Constraint weightHow strongly waypoints and pins are obeyed. Raise it if the character misses its marks; lower it if it contorts to hit them.
• History cropHow many past frames each new chunk looks back at to continue smoothly.
• Future cropHow far ahead each chunk reads upcoming prompts and constraints, so it can prepare (slow before a waypoint, turn before a new direction).
• Replan bufferFrames of margin kept around an automatic replan. Approximation; the engine's exact rule is not published.
• Replan thresholdHow far the motion may drift from the plan before it replans itself. Lower corrects more often. Approximation; not published.
• Auto replanLets the engine correct itself when the motion drifts from what was planned.
• ApplySends the Advanced values to the running engine.
• BVH file pathWhere Export BVH only saves the file.
• Export BVH onlySaves the animation as a BVH file without loading it, for use in other programs.
• SetupARDY needs about 7 GB of weights from Models. It shares the GPU with AI Render, so one stops the other.

MotionBricks

• W A S DSteers the character live, like a game.
• GaitsThe movement style: walk, crawl, stealth, injured, zombie and so on. Switch while recording for a performance that changes mood.
• FacingFace where I walk turns the character toward its direction of travel. Strafe keeps it facing the camera while it steps sideways or back.
• RecordNothing is kept until you record. Steer while recording, then stop.
• ResetPuts the character back at the start.
• Send to LayoutLoads the recorded take as animated nulls, one per bone.

UniMate

• PromptDescribe the action in plain words.
• GenerateWrites the motion directly onto the rig of the character selected in Layout, using your own bones. It recognises bones by name, so correct unusual names first.
• SetupNeeds its checkpoint from Models. The licence is non-commercial (CC BY-NC 4.0).

Video Mocap

One person, full body, steady camera. The skeleton comes into Layout as nulls.

• 1. VideoThe footage to track. Best: one person, whole body always in frame, steady camera, clear contrast with the background, close-fitting clothes. Depth is guessed from a flat video, so movement straight toward or away from the camera is the least accurate.
• Frame strideTrack every frame, or every 2nd or 3rd. Skipping is two to three times faster but loses quick moves (a fast kick can vanish). Skipped frames are held so timing still matches the video.
• Send to Layout when doneLoads the result as soon as tracking finishes.
• Track & send to LayoutFinds the person's joints in every frame, rebuilds them in 3D, and loads them as an animated skeleton of nulls.
• Send last take to LayoutLoads the last tracked take. It arrives as a separate skeleton; to drive your own character, rig it in AI Rig with Bind to Video Mocap.

AI Rig

Rig a mesh, pose it, and send it to Layout as FBX. The first run downloads its weights from Models.

• 1. MeshThe model to rig. Works best on a single closed mesh in a neutral pose (T or A pose, arms away from the body). Separate parts such as eyes or armour can get odd weights.
• 2. SkeletonBind to AI Motion (Core) builds the bone layout ARDY uses, so ARDY takes play on your character. Bind to Video Mocap matches the Mocap skeleton. Generate invents bones to suit the shape (tails, wings), but no existing motion will fit them.
• Apply the last motion takePlays your last ARDY or Mocap take on the new rig so you can check the weights in motion before sending.
• Candidate bones per stepThe AI places bones one at a time, each time choosing from likely next bones. This is how many options it considers at each step.
• Sampling thresholdIgnores unlikely bone placements until this share of probability remains. Lower is more conservative.
• VariationLow gives the safest, most typical skeleton; high gives more unusual placements. Lower it if skeletons come out strange.
• Search widthHow many complete skeletons are built in parallel before the best is kept. Wider is more reliable but slower.
• Preserve texture and scalePuts the rig back on your original mesh so UVs, textures and size are kept.
• Voxel skin postprocessRecalculates the skin weights from a solid version of the mesh. Fixes parts that pull on the wrong bone (an arm moving the chest) at the cost of extra time.
• Rig meshPlaces the bones and works out how strongly each bone moves each part of the surface (the skin weights). Needs about 14 GB of VRAM.
• BoneBend bones in the preview to test the weights: bend an elbow and only the forearm should move. The pose is baked into the export, so reset it first if you want a neutral pose.
• 4. Send to LayoutWrites an FBX with the mesh, bones and weights and loads it into Layout.

Models, project, and settings

Models

Groups stay collapsed until you open them. A group is listed when the catalog has files for it. Download all on a group fetches that group. Community alternatives, on a group that offers them, switches in experimental replacements for the text encoder and LoRA.

• Main and LoRAsMain holds the models the tools run on. LoRAs are optional add-ons that teach a model a specific style (see Key ideas, LoRA).
• Models folderWhere all weights are kept. Together they can exceed 200 GB, so choose a large drive; an SSD loads models much faster. Changing it does not move existing files.
• Download all missingDownloads every missing file. Usually better to open a group and download only the tools you use.
• Hugging Face accessHugging Face is the site most AI models are published on. Some authors make you accept their licence first: make a free account, accept on each listed page, create a read token in your account settings, paste it here and Connect.
• ImageFlux Klein, the image model behind AI Render's Style step and the 3D Texture Engine. Needed for almost everything in Render.
• WanThe Wan 2.2 video engine: the highest-quality, prompt-guided option. Large download; only needed if you render with Wan.
• TeleStyleThe fastest video engine. A good first download, since it lets you test looks on video quickly.
• LTX 2.5The LTX 2.5 video engine, a balance of speed and control.
• MiniMax H3The MiniMax H3 video engine, which works from style reference stills.
• VedaOptional speed-up for MiniMax on long or large clips. Only useful if you render with MiniMax.
• MiniMax H3 X2Optional detail decoder that doubles MiniMax's output size with extra sharpness.
• SeedVR2 upscaleThe AI upscaler behind Upscale still and the lighter Upscale video option (about 4 GB).
• LTX detailerA very large (about 54 GB) video upscaler that invents richer detail than SeedVR2. Only worth it for final, high-resolution video.
• Trellis2The default engine for 3D Model Builder.
• Pixal3DAn alternative 3D Model Builder engine, worth having when Trellis2 misreads a shape.
• Hunyuan3D 2.1A detailed 3D engine. Its licence does not cover the EU, UK or South Korea.
• UniForge3DBuilds 3D from several photos of one object. Non-commercial licence.
• ARDYThe text-to-motion engine (about 7 GB).
• MotionBricksThe live-steered motion engine (about 2.2 GB).
• UniMateAnimates characters already in your scene. Non-commercial licence.
• Video MocapThe body tracker used to turn video of a person into animation.
• AI RigThe automatic rigging model. Needs about 14 GB of VRAM when running.
• Qwen3.5-4BThe language model behind the AI Assistant. Runs entirely on your computer.
• ImageStyle add-ons (LoRAs) for stills.
• LTXStyle add-ons for the LTX video engine.
• MiniMaxStyle add-ons for MiniMax.
• WanStyle add-ons for Wan.

Project

• Project nameThe start of every saved file name. Blank uses the scene's file name.
• Project folderOne folder per scene, with sub-folders for captures, styled stills, markup and each engine's videos, so every result has a predictable place.
• PreviewCaptures, with all passes. Numbered so each recapture is kept.
• StyledStyled stills. The ff/lf numbers in the name record which first and last frame they belong to, so you can tell versions apart.
• MarkupMarkup paintings, so edits can be reloaded later.
• LTXLTX 2.5 videos, in their own folder so engines' results never mix.
• MinimaxMiniMax videos.
• WanWan videos.
• TelesyncTeleStyle videos.
• Incremental filesOn: each save gets the next number, so every version is kept. Off: that type of file is overwritten, saving disk space but losing earlier versions.

Settings

• Enable the AI AssistantTurns on a chat assistant that runs on your own computer (nothing is sent online). It can write prompts and explain or change settings. It unloads before renders so it does not take graphics memory from them.
• Assistant runs onGPU answers quickly but uses graphics memory. CPU is slow but leaves the GPU entirely to rendering.
• Keep the assistant loaded while idleFaster replies, but its memory is not given back while idle.
• Think before answeringThe assistant reasons step by step before replying. Slower, sometimes better for complex questions.
• Release channelBeta gets new features earlier, with more risk of bugs. Stable gets only finished releases.
• Check for updatesChecks now instead of waiting for the automatic check.
• Open StorePlugins to buy, the ones you own, and updates.
• Read tokenYour Hugging Face token, for models that require accepting a licence. Stored encrypted for your Windows user.
• Download diagnosticsSaves versions, logs and settings to a file you can attach to a bug report. No images or license files are included.
• Styled preview folderAutomatically saves every styled still you generate into this folder, numbered, so no experiment is lost.
• Autosave filenameName used for automatically saved stills. Blank uses the scene name.
• Capture automatically on launchCaptures the shot as soon as the window opens, saving a click when you always start that way.
• Save the original capture tooKeeps the unstyled capture beside each render for before/after comparisons.
• Legacy capture (MakePreview)An older capture method using LightWave's MakePreview. Slower; use only if the normal capture misses something in your scene.
• Keep the engine running after closing this windowClosing the window does not stop a render in progress, and reopening is instant because models stay loaded.
• VRAM modeHow much of each model stays on the graphics card versus being moved to system memory when idle. High keeps everything on the card: fastest, needs 20 GB or more. Low and Maximum CPU offload let big models run on 8-12 GB cards, but much slower. Auto picks for your card.
• Reserve VRAM (GB)Graphics memory held back for the working data of a render. Raise it to 3-4 if large stills run out of memory.
• Dynamic VRAMLoads model parts only when they are needed instead of all at once, so larger models fit on smaller cards. Restart the video worker after changing.
• Async weight offloadMoves model parts between system memory and the card in the background while it works, hiding some of the cost of offloading.
• Wan model precisionHow many digits each number in the Wan model keeps. fp8 uses half the memory of bf16 and looks almost the same. Keep the recommended setting unless results look wrong.
• Tiled VAE For VideoConverts video frames back to pixels in pieces to save graphics memory (see Key ideas, Tiling). Keep it on unless you see seams.
• Default seedThe seed new scenes start with (see Key ideas, Seed).
• Default seed behaviorWhat happens to the seed after each render, for new scenes. Fixed for repeatable results, Random for variety.
• Unload models after each Style/Video stepFrees graphics memory when a job ends, so the next one starts clean. Slower, because the next job reloads its model. Turn on if switching between Style and Video runs out of memory.
• Unload models between pipeline stagesWithin one job, frees each part of the model as soon as it is done. Slowest, but lets the heaviest jobs run on small cards.
• Reset GPU after a failed renderA failed render can leave graphics memory in a bad state; this restarts the video engine so the next try works.
• Latent preview during generationShows rough frames while rendering so you can see early whether it is going well and cancel if not. Costs a little speed.
• ThemeThe colour scheme of the interface.
• Show announcementsShort notes from LightwaveLlama such as maintenance windows or sales.
• Start serverThe engine is the background program that runs the AI models. Start it again if it was stopped or crashed.
• Stop serverShuts the engine down and frees all GPU and system memory it was using.
• AccountOpens the sign-in screen.
• Free GPU memoryUnloads every model now, for example before using another GPU-heavy program.

Gallery

Open it from the Gallery button, or from Menu, Gallery. It lists files already saved. Click a thumbnail to open it.

• OutputsRendered videos.
• StyledStyled stills.
• InputsCaptures.
• RefreshReloads the tab you are on.
• Show in folderOn an open item. Reveals that file.

Store

Open it from Menu, Store, or from Settings, Open Store. Purchases use your LightwaveLlama account.

• BrowsePlugins you can buy.
• LibraryPlugins you own. The count sits on the tab.
• UpdatesUpdates for what you own. The count sits on the tab.
• RefreshReloads the catalog and your purchases.
• CloseLeaves the Store.
• PayCard details go straight to Stripe, the payment processor; mAInframe never sees your card number.
• Back to StoreShown while an installed plugin page is open.

Account

Sign in with your LightwaveLlama account. mAInframe II unlocks when that account has an active LWMF2 license, and this computer is activated automatically. No key is typed in here.

• EmailThe account email.
• PasswordThe account password.
• Confirm passwordShown while you are creating an account.
• Sign inSigns in. Create an account shows Confirm password and changes the button to Create account. The note then reads Use the email associated with your purchase. The other button reads Already have an account? Sign in.
• Forgot password?Uses the email you typed. If that email has an account, a password reset link is sent.
• Check purchase againLooks up your license again, for example right after buying.
• Sign outSigns this computer out.
• ContinueShown once you are signed in. It reads Continue when the license is active, and Continue to Store when it is not. It closes the screen. A signed-in account without a license opens the Store.
• Buy mAInframe IIOpens the LightwaveLlama store page. A new account opens the in-app Store. Rendering and animation tools are a separate purchase.
• Manage purchasesYour LightwaveLlama account page, where you can move the license to another computer.

AI Assistant

The Assistant button in the title bar opens a column on the right and pushes the tool aside. Turn it on under Settings, Enable the AI Assistant. It runs Qwen3.5-4B inside mAInframe.

• MessageAsk in plain words: 'write a prompt for a rainy cyberpunk street', 'why is my video flickering', 'what VRAM mode should I use'. It can see your current settings.
• ApplyThe assistant only suggests; nothing changes until you press Apply, so you can review every change first.
• Auto applyLets it fill in prompts and settings without asking. Renders still wait for you.
• NewStarts a new conversation.
• Add an imageShows it a picture to describe or match, e.g. a reference you want your shot to look like.
• CloseCloses the column. The same button in the title bar toggles it.

Report a bug

Open it from Menu, Report a bug, or from Report bug on the console. The console also has Copy, Follow, and Clear.

• SubjectA short summary, e.g. 'Wan render fails at frame 40'.
• What happenedWhat you did, what you expected and what happened instead. Steps to reproduce help most.
• Include the console logAttaches the engine's log. This shows exactly what failed and is the most useful part of a report.
• Images or videoA screenshot or screen recording of the problem.
• Send reportSends the ticket to LightwaveLlama. Cancel closes the form.
• SupportWhen a reply arrives, Reply opens it. The thread is listed, then a Reply box, Send, and Back.