Natural Lift Motion – LTX 2.3 / Eros10 LoRA [Beta]

This is my first LoRA and my first serious deep dive into training motion behavior for LTX 2.3 / Eros10.

The goal of this LoRA is to create a natural image-to-video lifting motion where an adult woman grips the lower edge of her top, pulls it upward, the fabric creates brief tension/compression, and then releases into natural soft motion with short damped settling.

This is a Beta release. It already works very well in many cases, especially for 3s and 5s image-to-video clips, but results still depend on the source image, prompt, LoRA strength, seed, clip length, clothing structure, and workflow setup.

Trigger word

physicsdropfx

What this LoRA is designed for

This LoRA is designed for:

• natural top-lifting motion

• hands gripping the lower edge of the top

• upward fabric movement

• brief natural compression/tension

• clean release when the fabric clears the chest

• realistic weighted motion

• short soft damped settling

• improved long-top / pullover behavior

• reduced clothing morphing

The main focus is not just “movement”, but the full transition:

grip → pull → tension/compression → release → weighted motion → soft settling

Important Beta note

This is a Beta, not a perfect universal one-click solution.

It can already produce very good results, but the final quality depends heavily on:

• source image

• prompt

• LoRA strength

• clip length

• seed

• clothing structure

• camera angle

• workflow settings

• distilled LoRA used in the workflow

Prompting has a major influence on the result. If you clearly describe the clothing, grip point, body shape, and intended motion, the output can improve significantly.

Base / workflow notes

Main base model:

• LTX 2.3 / Eros10

Distilled LoRA used during testing:

• LTX-2.3-22b-distilled-lora384-1.1

Important:

Different distilled LoRAs can produce noticeably different results. Motion strength, fabric stability, timing, and overall behavior may change depending on which distilled LoRA is used in the workflow.

This LoRA was mainly tested with LTX-2.3-22b-distilled-lora384-1.1, so results may vary if you use a different distilled LoRA.

Recommended LoRA strength

There is no single perfect strength for every image.

A good starting point is:

0.90

General range:

0.75 – 1.0

Suggested guide:

• 3s clips: 0.75 – 0.90

• 5s clips: 0.85 – 1.0

Shorter clips usually need lower LoRA strength.
Longer clips can often handle slightly higher strength.

If the motion looks too aggressive, slightly telekinetic, or if the clothing starts to morph, lower the LoRA strength.

If the motion looks too weak or underpowered, raise it gradually.

Prompting tips

This is a motion LoRA, so prompt structure matters more than with a simple style LoRA.

The LoRA responds best when the prompt clearly describes:

• the hands gripping the lower edge of the top

• the hands staying in contact with the fabric

• the top being pulled upward

• the brief tension/compression phase

• the release once the fabric clears the chest

• soft damped settling

• stable clothing structure

Breast size and body proportions can also be influenced through the prompt. If you want a specific look, describe it clearly, but keep in mind that the source image still has a strong influence.

Use only with adult characters / adult-looking subjects.

Example prompt

physicsdropfx, an adult woman grips the lower edge of her top with both hands and pulls it upward in one clear natural motion. Her hands remain in contact with the fabric during the pull and move upward together with it. As the top slides upward over the chest, natural tension and brief compression build across the chest. When the fabric clears the chest, that tension releases and the breasts fall free naturally with realistic weight and momentum, followed by a soft damped settling phase. The clothing stays structurally stable and coherent, with no morphing, warping, or fusion. Static camera, stable framing.

Example negative prompt

hands sliding over the chest, rubbing motion, stroking the body, hands staying low, pulling pants, lifting waistband, fused top and pants, merged clothing layers, clothing morphing, fabric warping, fused fabric and skin, floating fabric, telekinetic clothing motion, unstable clothing structure, broken fabric edges, bad hands, extra fingers, twisted arms, chaotic motion, jelly motion, exaggerated bounce, asymmetrical motion, flickering details, warped torso

Audio prompt note

If your workflow supports audio, you can add something like:

Natural room ambience only, subtle fabric rustling, no background music.

Optional negative audio terms:

background music, song, singing, loud music, dramatic music, distorted audio, robotic voice, unnatural voice, excessive talking, long dialogue

For clean motion tests, I recommend avoiding long dialogue.

Clothing morphing note

In some cases, slight clothing morphing can still occur, depending on the source image, clothing structure, seed, prompt, clip length, distilled LoRA, and LoRA strength.

If that happens, adjusting the LoRA strength and refining the prompt usually helps a lot. Prompting has a major influence on the final result, so clearly describing the grip point, clothing, and intended motion can significantly improve stability and motion quality.

Long tops and pullovers were specifically refined, but they can still be more sensitive than shorter tops. If the motion becomes too aggressive, telekinetic, or clothing morphing appears, try lowering the LoRA strength and making the prompt more precise.

Best results

Best results are usually achieved with:

• clear source images

• readable torso / clothing structure

• understandable lower edge of the top

• stable camera framing

• good prompt description

• LoRA strength adjusted to clip length

• 3s and 5s image-to-video clips

• LTX-2.3-22b-distilled-lora384-1.1 in the workflow

Extreme camera angles can work, but may require more prompt and strength tuning.

Known limitations

Known limitations:

• slight clothing morphing can still happen in some cases

• results can vary by seed

• prompt quality matters a lot

• some images need lower or higher LoRA strength

• difficult clothing structures may need more tuning

• long or visually ambiguous tops can be more sensitive

• 8s clips are not the main target and may be less stable

• very unclear source images may require more manual prompt adjustment

• different distilled LoRAs can change the final result, especially motion strength, fabric behavior, and timing

Training overview

Base model:

• LTX 2.3 / Eros10

LoRA rank:

• Rank 32

Training resolution:

• mainly 512 × 512

Training clip formats:

• early phased clips used shorter frame structures

• later refinements mainly used 73-frame clips

• main target behavior was optimized for 3s and 5s image-to-video clips

This LoRA was trained through multiple small refinement stages instead of one large training run.

The workflow included:

• base motion learning

• phased motion refinement

• clothing stability refinement

• caption-fix refinement

• long-top / pullover refinement

• final motion / settling polish

The strongest improvements came from small targeted refinement passes with low learning rates and only a few steps. This helped fix specific weaknesses without destroying the good parts of the motion.

Tools used

• ComfyUI for generation and testing

• LTX / LTX-2 trainer environment for preprocessing and training

• LTX 2.3 / Eros10 as the main base

• LTX-2.3-22b-distilled-lora384-1.1 during testing

• manual dataset selection

• manual filtering

• repeated recaptioning

• multiple small fine-tuning passes

Hardware used

This LoRA was trained on:

• RTX 5090

• 64 GB DDR5 RAM

• Ryzen 7 9800X3D

This setup was a good compromise between performance and quality for my workflow. Higher or heavier settings quickly pushed VRAM usage to the limit and caused OOM issues, so I focused on controlled datasets, Rank 32, 512×512 training, and small targeted refinement passes instead of brute-forcing larger settings.

Release notes

Beta 1 release notes:

• first public beta release

• trained for LTX 2.3 / Eros10 image-to-video workflows

• optimized mainly for 3s and 5s clips

• tested mainly with LTX-2.3-22b-distilled-lora384-1.1

• improved clothing stability

• improved long-top / pullover behavior

• improved tension → release → soft settling motion

• reduced clothing morphing through multiple small refinement passes

• recommended strength: start around 0.90 and adjust between 0.75–1.0

Final note

This is my first LoRA, and a lot of this process was me learning, testing, breaking things, fixing them, and slowly understanding what each change actually does.

It is still a Beta, but it already works very well when used with a good source image, a clear prompt, the right distilled LoRA, and the right LoRA strength.

If you take a little time to work with it, adjust the strength, and describe the motion clearly, it can produce very good results.