🧪 Overview

This is an experimental LoRA trained for the LTXV I2V (Image-to-Video) model, version 0.9.7, with 13B parameters. It transforms static images into fluid, seamless animated loops, with natural motion applied only to flexible or dynamic elements — like hair, clothing, particles, and ambient light — while preserving rigid structure stability (e.g., armor, weapons, mechanical parts).

This is the first version of this LoRA, and results may vary depending on the prompt quality and seed. Better versions may be released in the future as training techniques are refined.

My Workflow

⚙️ Training Details

• 🧠 Base Model: LTXV I2V v0.9.7 (13B parameters)

• 🎞️ Video Dataset: 140 short clips

• ⏱️ Frame Rate: 24 fps

• 🧮 Frames per Video: 49

• 🖼️ Resolution: 512px

• 🔁 Epochs: 250

• 🧮 Total Training Steps: ~35,000

• 📉 Learning Rate: 1e-4

• 📦 Batch Size: 1

• 📐 LoRA Dimension: 32

• ⚙️ Optimizer: AdamW8bit

• 🛠️ Trainer Used: Diffusion Pipe (by tdrussell)

• 🚫 Not using official trainer: LTX-Video-Trainer (by Lightricks)

• Layer Coverage:

  • When trained using Diffusion Pipe, all layers were updated during LoRA training.

  • In contrast, the official trainer from LightTricks (LTX-Video-Trainer) by default only updates attention layers (e.g., to_k, to_q, to_v, to_out.0), making it possible to use higher dim (e.g., 128) while still keeping the file size low (~700MB).

• Initial Loss: High — LTXV I2V is known to require many steps before reaching stability

⚠️ The I2V 13B model begins with a very high initial loss, and convergence is slow — requiring many steps to stabilize below 0.1. Training this architecture is not plug-and-play and takes persistence.

⚠️ Prompting Recommendations

This LoRA is very sensitive to prompt quality and seed variation.

Using short or unclear prompts often causes:

• Rigid elements like weapons or chairs to appear soft or rubbery

• Unintended motion of static parts (e.g., armor bending, background flickering)

These artifacts are not due to the LoRA itself but rather to a lack of motion guidance in the prompt or an unsuitable seed.

✅ To get the best results:

• Use long, detailed prompts that clearly separate moving vs non-moving parts

• Try changing the seed if you're seeing unwanted distortion

You can generate prompts automatically using my custom ComfyUI node:
🔧 Ollama Describer
This node uses a vision-capable LLM to generate motion-aware captions. In my case, I used Qwen2.5-VL-7B to generate all motion prompts during training and testing.

💡 Alternatively, the LTXV Prompt Enhancer from Lightricks' custom node set may also be used for prompt conditioning.

🧠 Recommended Prompt Template

Use this with any vision-enabled LLM like Qwen-VL, Gemini, or GPT-4o:

You are an expert in motion design for seamless animated loops.

Given a single image as input, generate a richly detailed description of how it could be turned into a smooth, seamless animation.

Your response must include:

✅ What elements **should move**:
– Hair (e.g., swaying, fluttering)
– Eyes (e.g., blinking, subtle gaze shifts)
– Clothing or fabric elements (e.g., ribbons, loose parts reacting to wind or motion)
– Ambient particles (e.g., dust, sparks, petals)
– Light effects (e.g., holograms, glows, energy fields)
– Floating objects (e.g., drones, magical orbs) if they are clearly not rigid or fixed
– Background **ambient** motion (e.g., fog, drifting light, slow parallax)

🚫 And **explicitly specify what should remain static**:
– Rigid structures (e.g., chairs, weapons, metallic armor)
– Body parts not involved in subtle motion (e.g., torso, limbs unless there’s idle shifting)
– Background elements that do not visually suggest movement

⚠️ Guidelines:
– The animation must be **fluid, consistent, and seamless**, suitable for a loop  
– Do NOT include sudden movements, teleportation, scene transitions, or pose changes  
– Do NOT invent objects or effects not present in the image  
– Do NOT describe static features like colors, names, or environment themes  
– The output must begin with the trigger word: **lvwpr**  
– Return only the description (no lists, no markdown, no instructions)

🧪 Experimental Status

This is the first public version of this LoRA for LTXV I2V.
If I discover new training techniques, better captioning strategies, or improvements in convergence, future versions will be released with higher quality and better performance.

🙌 Feedback Welcome

If you create something interesting with this LoRA, feel free to share what you’ve made.
I’ll be checking community uploads — and if I find your results particularly impressive, I’ll help give them a boost of Civitai buzz 😉