| name | ltx23-prompting-workflow |
| description | High-precision prompting protocol for LTX 2.3 (Lightricks) video generation, optimized for ComfyUI VFX pipelines and local control. Incorporates cinematic prose, camera motion selectors, and the community-standard 3-stage sampling hack. |
| model | LTX 2.3 (Lightricks) |
ltx23-prompting-workflow
Overview
This skill provides a standardized framework for generating cinematic, high-fidelity prompts specifically designed for the LTX 2.3 model architecture. It leverages director-level specificity to maximize motion dynamics and temporal coherence, optimized for ComfyUI VFX production on local hardware (e.g., RTX 3090).
Core Principles
- Cinematic Prose: Avoid generic descriptors; use professional filmmaking terminology (lens types, lighting sources, film stocks).
- Motion Priority: The model rewards explicit camera movement and physical interaction descriptors.
- 3-Stage Sampling Requirement: All LTX 2.3 workflows should implement the community-standard "3-stage sampling" to ensure micro-expressions and dynamic depth.
- I2V/V2V Constraint: When using reference images/videos, describe only motion, camera, emotion, and changes. Do not re-describe the static subject.
- Specificity Floor: Every prompt must name the camera behavior, subject motion, environmental motion, light continuity, and one explicit artifact risk in the negative field.
- Anti-Living-Photo Rule: A prompt that only says "preserve identity," "gentle parallax," or "cinematic motion" is invalid. It must describe what visibly changes frame-to-frame.
Prompt Architecture (The 6-Element Structure)
Every prompt should be structured as a single cinematic paragraph following this sequence:
- [Genre / Film Stock / Era]: e.g., "Vintage 1985 dark fantasy film", "Kodak Portra 400 aesthetic".
- [Subject Description]: Detailed physical characteristics and clothing.
- [Precise Action + Emotion]: What is happening and the underlying feeling (e.g., "releasing a paper lantern with a mournful expression").
- [Camera Motion]: Explicit movement (e.g., "slow lens push from medium close-up to close-up", "snap zoom", "Dolly In").
- [Lighting Source + Atmosphere]: Physical light descriptions (e.g., "soft side light from the morning sun through fog", "rim lighting").
- [Stylized Tone / Framing]: Final aesthetic polish (e.g., "central framing with telephoto compression, melancholic tone").
Implementation Standards
T2V (Text-to-Video) Example
A silver-haired elderly woman in a loose linen dress stands at the edge of a misty forest lake at dawn, her weathered hands releasing a paper lantern onto the still water; the camera uses a slow lens push from medium close-up to close-up, capturing her peaceful yet mournful expression; soft side light from the early morning sun filters through the fog, casting a warm glow and rim light around her silhouette; the scene is composed with central framing and uses a telephoto lens to compress the depth of the misty forest; stylized with a poetic, melancholic tone and naturalistic visual style.
I2V/V2V (Image/Video-to-Video) Protocol
Focus on the first 20-40 words for maximum impact.
- Rule: Do not re-describe the image.
- Goal: Describe how the pixels move.
- Example: "The subject's eyes widen in sudden terror as the camera performs a rapid snap zoom; heavy rain begins to lash against their skin, creating realistic water ripples and splashing."
Negative Prompting (Critical)
Always include a negative prompt field to mitigate known failure modes:
negative: no drum sticks, no stiff motion, no flat lighting, no artifacts, no morphing, no flickering
Keep the negative prompt in the dedicated negative field. Do not bury it inside the positive prompt paragraph.
Advanced Techniques & Troubleshooting
| Issue | Solution |
|---|
| Stiff/Flat Motion | Use 3-Stage Sampling and more explicit Camera Motion descriptors. |
| Loss of Precision | Integrate ControlNets (Canny, Depth, or Pose) into the ComfyUI workflow. |
| Character Drift in V2V | Use Start/End frame images + IC-LoRA "SCENE LOCK" technique. |
| Unwanted Audio Artifacts | Ensure a robust negative prompt is present if audio generation is active. |
3-Stage Sampling (Community Standard)
Run the KSampler in three sequential passes with progressively decreasing denoise values rather than a single full-denoise pass. This prevents flat motion and builds micro-detail in layers:
- Pass 1 — Structure (
denoise: 1.0, ~10 steps): Establishes composition, subject, rough motion
- Pass 2 — Motion Refinement (
denoise: 0.65, ~10 steps): Refines movement and temporal coherence
- Pass 3 — Detail Polish (
denoise: 0.35, ~5 steps): Adds micro-expressions, texture, and fine motion
In ComfyUI: chain three KSampler nodes in series, each taking the latent output of the previous. Use the same seed across all three passes for coherence.
IC-LoRA SCENE LOCK (Video Extension)
IC-LoRA (Image Conditioning LoRA) is a technique for extending clips beyond their native length without scene drift. SCENE LOCK refers to anchoring the extension by conditioning on both the last frame of the prior clip and a target end frame.
Implementation in ComfyUI:
- Load the final frame of your generated clip as a start-frame reference
- Optionally provide a target end-frame image to direct where the extension lands
- Apply IC-LoRA weights to the base LTX 2.3 model before sampling
- Prompt describes only the continuation motion — do not re-describe the locked scene
This preserves lighting, wardrobe, environment, and character appearance across clips that would otherwise drift.
Workflow Integration
For professional VFX, chain this protocol as follows:
- Reference Generation: Use Flux/SDXL for high-quality reference stills.
- Motion Injection: LTX 2.3 I2V/V2V using the structured prompt above.
- Temporal Extension: Use IC-LoRA SCENE LOCK to extend clips while preserving scene continuity.
- Upscaling: Final pass with RTX Video Super Resolution (x2/ULTRA) post-VAE decode in ComfyUI.
Skill Version
v1.5 (Updated: 2026-04-17)
Author: Hermes Agent Team
Compatibility: LTX 2.3 models only (LTX0203022b0dev0fp8, ltx-2.3-22b distilled)
