- LTX-2.3's text encoder responds to domain-specific shorthand — VFX terms (rotoscope, matte painting, rack focus), animation terms (contact/breakdown poses, ease-in/ease-out), and restoration terms (telecine, gate weave, Kodachrome) — more precisely than generic descriptions.
- Prompts stay under 200 words in one flowing paragraph, starting with the action, and follow seven components: action, movement, appearance, environment, camera, lighting, and any sudden changes.
- When combining domains, rank them — name the primary vocabulary first and use the others only as secondary/tertiary qualifiers — since equal weighting makes the model average between them instead of committing.
Prompt quality determines output quality in LTX-2.3, and the fastest way to raise prompt quality is to switch from generic descriptive language to the vocabulary your target domain already uses.
Compositors, animators, and restorers each have a shorthand (matte, hold, telecine) that names shots and behaviors precisely. This guide maps that vocabulary onto LTX-2.3 prompt structure, gives before/after examples for each domain, and documents the common mistakes that emerge when domain vocabularies get mixed.
Prompting Fundamentals for LTX-2.3
LTX-2.3 is a DiT-based audio-video foundation model with a Gemma 3 multilingual text encoder using multi-layer feature aggregation and learnable registers (thinking tokens).
The text encoder's job is to translate your prompt into conditioning signals for the 48 shared transformer blocks that drive video generation. That means how you phrase the prompt matters as much as which entities you name.
How LTX-2.3 interprets language
LTX-2.3 prompt guidance is explicit: focus on detailed, chronological descriptions of actions and scenes; include specific movements, appearances, camera angles, and environmental details, all in a single flowing paragraph; start directly with the action; keep descriptions literal and precise; think like a cinematographer describing a shot list; keep within 200 words. Automatic prompt enhancement is available via an enhance_prompt parameter on the pipelines.
Use it to expand shorthand prompts, but for domain-specific work, hand-writing the prompt in the target vocabulary usually beats enhancement.
Structure of a strong LTX-2.3 prompt
A strong LTX-2.3 prompt has seven documented components: (1) the main action in a single sentence, (2) specific details about movements and gestures, (3) character or object appearances described precisely, (4) background and environment details, (5) camera angles and movements, (6) lighting and colors, and (7) any changes or sudden events.
Domain vocabulary attaches at components 5 and 6 primarily, with secondary attachment at 4 (environment).
VFX Vocabulary: Prompt Like a Compositor
VFX vocabulary reaches back to optical compositing and forward to modern node-graph tools. Naming shots the way a compositor names them gives LTX-2.3's text encoder tokens the model has seen paired with the corresponding visual behavior during training.
Compositing and matte painting terms
Compositing terms name the layered structure of a shot. "Matte painting background" reads as a large static painted-look environment. "Bokeh-heavy foreground plate" reads as a soft-focus foreground layer. "Rack focus from background matte to foreground character" reads as a specific defocus transition. These aren't marketing phrases. They are the terms a real production shot description would use, and they steer the model toward matching aesthetics.
Rotoscope, tracking, and camera terms
Rotoscope terminology names the isolation of a subject against its background. "Rotoscoped subject on plate" reads as a clean-edged foreground element. Tracking terminology ("locked-off camera", "dolly-in on subject", "handheld tracking shot", "steadicam glide") maps onto documented camera-control LoRAs in the LTX ecosystem: LTX-2-19b-LoRA-Camera-Control-Dolly-In, -Dolly-Out, -Dolly-Left, -Dolly-Right, -Jib-Up, -Jib-Down, -Static. When you use a camera LoRA, explicitly describe the intended camera movement in the prompt.
VFX prompt example: before and after
Before (generic): "A person walking through a forest, cinematic look."
After (VFX vocabulary): "Rotoscoped subject in a wide two-shot, mid-frame, walking screen-left to screen-right across a matte-painted forest background. Locked-off camera at eye level. Volumetric backlight from screen-right through the canopy, mist floor pass at ankle height. Muted teal-and-gold palette. Cut-in on the subject's hand brushing a fern at second two."
Animation Vocabulary: Prompt Like an Animator
Animation vocabulary handles the timing and dynamics of motion. An animator describes a walk cycle in terms of contact, down, passing, and high positions, not "person walking".
2D vs. 3D animation directives
2D animation vocabulary reads as flat: "hand-drawn linework", "cel-shaded flat color", "limited animation with held frames", "ink-and-paint style". 3D animation vocabulary reads as depth: "posed on threes", "rigged character walk cycle", "sub-surface skin scatter", "arc of motion following the ball path". The two vocabularies produce visibly different outputs from LTX-2.3, and mixing them produces mixed results. Pick one and stick with it per shot.
Timing, easing, and pose terminology
Timing terminology names the motion curve. "Ease-in to a hold, ease-out into the next action" describes a paced two-beat motion. "Snap to pose, hold two frames, drift back" describes a sharper entry with a soft settle. Pose terminology ("contact pose", "extreme pose", "breakdown pose") attaches to the pose-control LoRA family (LTX-2-19b-IC-LoRA-Pose-Control) and reads directly to the encoder.
Animation prompt example: before and after
Before (generic): "A cartoon character throws a ball."
After (animation vocabulary): "Cel-shaded 2D character in mid-shot, flat colors, thick ink linework. Anticipation frame at frame one, character coils back, arm cocked behind head. Snap to release pose at frame six, arm extended fully. Follow-through drift for six frames past release, arm relaxing. Ball trails a stretched motion streak along an arc from screen-left to screen-right. Locked-off camera, three-quarter angle. Warm afternoon palette."
Restoration Vocabulary: Prompt Like a Colorist
Restoration vocabulary handles the material properties of the medium: the film stock, the tape era, the color science.
Grain, telecine, and film-emulation terms
Grain terminology names the noise character. "Fine grain, Kodachrome-tight" reads as low-noise saturated. "Coarse silver-halide grain, black-and-white 16mm" reads as heavy monochrome noise. Telecine terminology names the transfer process. "Telecine flicker, gate weave, edge halation" reads as archival-transfer artifacts. Film-emulation terms ("Ektachrome", "Fuji Reala", "Portra") pull specific color-response curves the model has seen in training data.
Color science and palette directives
Color science vocabulary is precise. "Log-encoded, wide-gamut, muted contrast" reads as a professional-grade base image. "sRGB-clipped highlights, crushed blacks" reads as a consumer-grade final image. "Complementary teal-and-orange grade" reads as the widespread modern feature-film palette. These terms pull period-accurate and style-accurate palettes from the encoder.
Restoration prompt example: before and after
Before (generic): "An old-looking video of a couple dancing."
After (restoration vocabulary): "Kodachrome-emulation palette, warm mid-tones, faint gate weave. Two dancers in mid-shot, three-quarter angle, filmed on 16mm in the mid-1960s. Fine silver-halide grain overlay. Rack focus from the woman's face at frame one to the man's face at frame twenty. Telecine flicker every ninety frames. Complementary teal-and-orange grade, no supersaturated primaries."
Building a Cross-Domain Prompt Template
Cross-domain prompting layers the three vocabularies into a single 200-word prompt when a shot legitimately combines them: a rotoscoped VFX plate composited over an animated background, restored with a period-accurate palette, for example. The layering has to stay disciplined or the encoder receives conflicting signals.
A workable template: open with the domain-primary vocabulary (whichever dominates the shot), attach the second domain as an environmental or material qualifier, and reserve the third domain for lighting or grade only.
Example: "Rotoscoped subject (VFX-primary) on a hand-drawn cel-shaded background (animation-secondary), Kodachrome-emulation palette with fine grain overlay (restoration-tertiary)."
Common Mistakes When Mixing Domain Vocabularies
Mixing domain vocabularies produces the most common failure mode: the model averages between the vocabularies rather than committing to one.
Vocabulary-strength mismatch. Naming three domains equally in the prompt leaves the encoder with no primary. Rank the domains and name the primary first and most.
Anachronistic combinations. "1920s Kodachrome grade" is anachronistic — Kodachrome wasn't introduced until 1935. Verify period-vocabulary consistency before prompting.
Camera-vocabulary-and-LoRA misalignment. When using a camera LoRA, explicitly describe the intended movement in the prompt to match.
Summary
LTX-2.3 prompt quality rises when the prompt uses vocabulary from the target creative domain. VFX vocabulary names layers, mattes, and rotoscope structure. Animation vocabulary names timing, poses, and animation principles. Restoration vocabulary names film stock, color science, and telecine artifacts.
Prompts should stay under 200 words, use a single flowing paragraph, start with the action, and match the loaded LoRAs. Rank vocabularies when combining domains and name the primary first. See the LTX-2 prompt guide for more, or github.com/Lightricks/LTX-2 for example prompts alongside each pipeline entry point.
