Faster! Harder! Shake Harder! | H3 Motion Booster

To be completely transparent: rendering and training large video LoRAs like H3 on cloud GPUs is an absolute money burner 💸. High-tier cloud compute costs add up so quickly with every experiment and epoch iteration.

Every single bit of support helps cover those heavy cloud GPU rental fees, and it directly serves as the motivation and fuel to keep this Bunny training and bringing more high-quality models to you all! 🐰⚡

Ref2VA Edition

A few users have reported that Dynamic V2.0 already works very well with the FL2VA/Ref2VA Hybrid model, so this version is not intended to replace it.

This is a native Ref2VA-trained variant. In my testing, it provides better reference character consistency, with some trade-offs in image quality and overall stability.

Since it's already finished, I'm releasing it so you can try both and choose whichever works better for your workflow.

Use at the start:<Picture 1> is the adult woman.

The woman from <Picture 1> is etc.

Please let me know your testing results. The two versions were trained with different training methods and captioning logic, so I'm very interested to see how they behave across different workflows, prompts, and reference images.

This LoRA is provided as a motion-generation tool. Users are responsible for how they use their reference images and generated outputs.

H3 Motion Booster V2

V2 is a major upgrade over the original Motion Booster — broader motion coverage, stronger impact response, more natural secondary body movement, and added audio support.

What’s New

• Expanded motion coverage – more movement patterns beyond simple forward-backward / vertical motion

• Stronger impact & rebound – better compression, separation, recoil, and visible body bounce

• Improved secondary motion – hips, thighs, torso and chest now react with natural delayed follow-through

• Audio support – trained with audio for better contact sounds, breathing, and motion-linked effects

• Better rhythm control – handles steady, fast, compact, and heavy impact-driven motion more reliably

Usage

Trigger: dynv2 Recommended weight: 0.6–0.8 (start at 0.7)

Works best with I2V. Start from a correctly posed reference image, then describe motion direction, rhythm, contact, rebound, and audio.

I2V performs significantly better than T2V (T2V needs more rerolls).

When paired with NSFW ALLINONE, it can also handle fast oral motion well.

Recommended prompt structure:

integrated_multimodal_description:
overall_soundscape:
non_diegetic_music:

I2V Example

integrated_multimodal_description:

dynv2. The two subjects from preserve the exact starting position, body alignment, and interaction shown in the reference image.

His cock is inserted deep inside her vagina and stays clearly penetrating her for the entire duration.

From the very first frame he/she drives extremely rapid, high-cadence vertical hip motion with short, aggressive strokes — minimal dwell time at the bottom, instant separation, and immediate rebound. The pace is blistering and relentless with zero build-up, maximum visible speed maintained from start to finish. Her hips lift and drop in a continuous, dense rhythm while her breasts bounce heavily with every impact.

extremely rapid high-cadence pelvic motion, compact short-stroke movement, instant rebound, uninterrupted maximum speed, no slow-motion, no acceleration phase. Clear wet penetration is visible as the cock slides rapidly in and out of her vaginal opening.

overall_soundscape: Loud, dense, rhythmic wet skin-on-skin impacts are tightly synchronized with every rapid stroke. Heavier wet slaps land on the deeper thrusts. Continuous heavy breathing mixed with soft muffled moans and short exertion sounds from the woman. Mattress creaks under the intense motion. No spoken dialogue.

non_diegetic_music: None.

Tip: Select he or she depending on who should lead the motion.

H3 Motion Booster V1

This is an experimental motion enhancement LoRA for MiniMax H3, designed to boost action amplitude, fluidity, body recoil, and secondary motion (especially chest movement) in dual-person interactions, fixing the base model's over-soft movement.

Recommended Workflow & I2V / T2V Guidelines

The prompt is optional. The shorter prompts shown in the example images,This trigger just there to make it easier for me to apply them. This trigger has already been written as a generic template, you don’t have to use it—you can simply describe the intensity and amplitude of the shake.

ONLY work with fl2va 8int.

Both I2V and T2V are supported, but I strongly recommend using I2V for this version.

My preferred workflow is to first create a clean, anatomically correct two-person starting image using KREA2 or an image-editing tool, then feed it into H3 I2V along with this LoRA.

Using single-person images or relying solely on T2V often leads to unstable positioning, incorrect interactions, and flawed anatomy.

Trigger Words & Motion Directions

Choose the corresponding trigger word based on the primary direction of movement in your starting frame.
The prompt has been set to a general format; you can modify it to be more specific as needed, as long as you keep the key trigger words.

• Forward / Backward Motion (dynfb1): Best for standing, doggystyle, leaning forward, or any front-to-back reciprocating motions to enhance hip push and body recoil.

integrated_multimodal_description:
[Shot 1] dynfb1, keep the characters, clothing, environment, and overall composition from the first frame consistent. The two figures remain in close interaction while performing extremely strong, rapid, repeated forward-and-back body motion at maximum visible intensity. The hip movement is large, forceful, and high-amplitude, with deep repeated thrusting motion and very fast rhythmic pacing. The bodies show powerful momentum, strong recoil, and aggressive weight transfer with every cycle. The torso, shoulders, chest, waist, hips, hair, and limbs all display intense secondary motion, violent follow-through, and pronounced shaking from the force of the movement. The chest bouncing becomes very strong and highly visible, and the entire upper body reacts dynamically to the repeated high-force motion. The pace continues to build, becoming faster, harder, and more energetic over time. Preserve stable anatomy, keep both figures clearly separated, and keep the camera mostly steady so the body motion remains the main source of intensity.

• Vertical / Up-and-Down Motion (dynvt1): Best for cowgirl, reverse cowgirl, or vertical bounce/rising poses to amplify vertical amplitude and upper-body secondary motion.

integrated_multimodal_description:
[Shot 1] dynvt1, keep the characters, clothing, environment, and overall composition from the first frame consistent. From the very first frame the two figures are already locked in extremely fast, maximum-intensity vertical rising-and-dropping motion with zero build-up. The repeated upward lift and downward drop run at the highest possible speed and frequency for the entire duration, delivering relentless, non-stop, high-amplitude pounding without any slowing or pause. Every cycle is large, forceful, and aggressive, creating continuous deep recoil, violent weight transfer, and explosive body momentum. The hips and lower body drive the motion at blistering pace, while the torso, shoulders, chest, waist, hair, and limbs shake violently with strong secondary motion and heavy follow-through on every single stroke. Chest bouncing is extreme, constant, and highly visible from start to finish. Preserve stable anatomy, keep both figures clearly separated, and keep the camera mostly steady so the ultra-fast body motion remains the sole source of intensity through the final frame.

For more simplified or variant prompts, see the example image.

Current Status & Known Issues

This 1600-step proof-of-concept model was trained on RTX 4090D (48GB) using 20 video clips, aimed at validating motion feasibility rather than acting as a fully generalized solution.

Audio dynamic performance remains experimental and inconsistent, which will be further refined alongside dataset expansion and motion decoupling in future releases.