MiniMax H3 Kiss LoRA V0.3 (Beta)

Hi everyone!

I'm back with V0.3 of my MiniMax H3 Kiss LoRA!

Just like the previous versions, this is still a Beta release. I'm continuing to experiment with different training approaches to see what works best.

What's New in V0.3?

1. Expanded Training Dataset

I've increased the training dataset from 200 to 300 videos.

The newly added videos are generally clearer and higher quality, with clip lengths extended to a little over 6 seconds.

2. AI-Generated Captions (Codex)

This is probably the biggest experiment in V0.3.

Instead of manually writing captions for the training videos, I used Codex to automate the AI captioning process and completely re-captioned the entire dataset.

Previously, my captions were usually quite simple, such as "two people kissing."

AI-generated captions are much more detailed, describing the actions, interactions, and other visual details.

However, more detailed doesn't necessarily mean more accurate.

I'm still not sure whether AI-generated captions actually produce better training results than my own manual captions. That's exactly what I wanted to explore with this version.

Training Results & Checkpoints

I tested multiple checkpoints throughout the training process.

Some earlier checkpoints occasionally stopped the kissing action or failed to follow the prompts, while later checkpoints sometimes produced unnatural repetitive movements.

After testing, I personally found the 2000-step and 2200-step checkpoints to be the most usable.

So I'm releasing both versions:

• 2000 Steps

• 2200 Steps

I honestly can't say which one is better. Feel free to try both and see which works best for your workflow.

Known Limitations

Some of the issues from previous versions are still present, especially stiff tongue movements, unnatural mouth interactions, and repetitive motions.

Increasing the dataset size and using more detailed AI captions haven't completely solved these problems.

So please consider V0.3 another experimental Beta release, rather than a fully polished model.

Final Thoughts

Training this kind of motion LoRA for MiniMax H3 has been much more challenging than I originally expected.

But I'm still experimenting, learning, and trying different approaches.

Thank you all for your continued support, feedback, and encouragement!

I hope you enjoy V0.3, and I'd love to hear which checkpoint works better for you!

V0.2 Beta Release Notes

This LoRA has gone through many rounds of experimentation, along with plenty of failed attempts. Overall, training a LoRA for H3 has turned out to be much more challenging than I originally expected.

One of the biggest issues in the previous version was the stiff and unnatural tongue movements. While V0.2 brings some improvements in this area, the underlying problem is still far from being fully resolved.

French kissing remains a particularly challenging action for AI to learn. In many training videos, it's difficult for the model to distinguish which tongue belongs to which person, often resulting in confused or anatomically incorrect tongue movements.

For this version, I completely rebuilt the original training dataset and experimented with several changes:

• Increased the video training bucket resolution from 512 to 768, hoping to reduce blurriness and improve fine details.

• Introduced a large number of 16:9 training videos, whereas the previous dataset consisted entirely of 1:1 videos.

• Experimented with image LoRA training, inspired by techniques shared by other creators, to improve kissing poses and image quality.

Unfortunately, despite these changes and the additional training effort, the overall improvements have been relatively limited.

Included Models

This release includes three LoRA models:

Video LoRA:

• cxy_kiss_lora_h3_v02_step1200.safetensors

• cxy_kiss_lora_h3_v02_step1000.safetensors

These are two checkpoints from the same training run, saved at 1200 and 1000 steps.

I've tested both versions, but honestly, I haven't been able to determine which one performs better. Feel free to experiment and see which works best for you!

Image LoRA:

• cxy_kiss_image_v02

This is an additional image-based LoRA trained with a 1024 resolution bucket, intended to improve kissing poses, visual details, and overall image quality.

This is my first attempt at incorporating an image LoRA into this workflow, based on training approaches shared by other creators. Its effectiveness is still experimental.

Recommended LoRA Strength

cxy_kiss_image is an image-trained LoRA, designed to enhance visual quality, improve fine details, and provide a wider variety of kissing poses.

It is meant to be used together with cxy_kiss_lora, not on its own.

In other words, you need to load both LoRAs at the same time:

• cxy_kiss_lora – Handles kissing motions and animations.

• cxy_kiss_image – Enhances details, visual quality, and pose variety.

You can also combine cxy_kiss_image with the V0.1 video LoRA, but you may need to reduce the LoRA weights to avoid overly strong effects or visual distortions.

Suggested strengths:

• Image LoRA: 0.3–0.8

• Video LoRA: 0.5–1.0

Feel free to experiment with different combinations to find what works best!

Final Notes

For these reasons, V0.2 is still being released as a Beta version.

There are still many challenges to overcome, especially when it comes to realistic and natural tongue movements.

I plan to continue experimenting with different training approaches, higher-quality datasets, and better ways to improve motion consistency.

Hopefully, future versions will bring more significant improvements.

Thank you for your patience and support!

MiniMax H3 Kiss LoRA V0.1 Released

Hi everyone,

After I released my Wan2.2 Kiss LoRA, it received some positive feedback and support from the community.

However, Wan2.2 is starting to feel quite old to me now. The model itself has many limitations, and after MiniMax recently open-sourced H3, I almost immediately lost interest in continuing with Wan2.2.

H3 is an extremely powerful model, and I personally believe it has a good chance of becoming one of the mainstream open-source video models in the community, while Wan2.2 may gradually fade into the background.

Since H3 was released, I have been experimenting with many of its features, especially its multi-image reference video generation, which I find very interesting and fun to use.

Naturally, I decided to try training my Kiss LoRA on H3 as well.

Training Settings

I used the same video dataset that I previously used for my Wan2.2 Kiss LoRA.

The dataset itself has some limitations. Most of the source videos were collected quite a while ago, and many of them are relatively low-resolution and not particularly sharp.

Training settings:

• Resolution: 512 × 512

• Video length: about 5 seconds

• Frames: 124 frames

• FPS: 24

• Audio: disabled

• Training steps: 1750

• LoRA Rank: 32

Important Note About H3 LoRA Training

Please keep in mind that H3 has only been open-sourced for a very short time, and LoRA training support is still highly experimental.

At the time I trained this V0.1 model, I used AI Toolkit's early MiniMax H3 training implementation.

Interestingly, shortly after I finished this training, AI Toolkit added an alpha version of a dedicated MiniMax H3 Training Adapter.

This may be important because some early H3 LoRA users have reported strange behavior with previous training implementations, such as:

• LoRAs training successfully but having weak effects

• inconsistent behavior between I2V and reference-based generation

• image or audio quality degradation

• LoRAs becoming unstable after additional training

There is currently some speculation that H3's architecture and guidance-distilled training design may require more specialized handling than conventional video LoRA training.

Therefore, this V0.1 LoRA was trained before the new alpha Training Adapter became available.

This is another reason why I consider this release highly experimental.

I plan to test the newer H3 training implementation in future versions.

Does It Work?

Surprisingly, yes.

Despite all of the above, the LoRA does have an effect.

However, its behavior depends heavily on the generation mode.

I2V

For standard Image-to-Video, the effect seems relatively weak or inconsistent.

Sometimes I can see the LoRA influencing the motion, while in other generations the effect is barely noticeable.

I am currently not completely sure why this happens.

Multi-Image Reference / Reference-to-Video

With H3's multi-image reference / reference-based video generation, the LoRA works much more noticeably.

This is currently the workflow where I get the best results from V0.1.

However, the generated video can sometimes look a little blurry.

My current guess is that this is mainly caused by the relatively low quality and limited resolution of my original training dataset, although the still-evolving H3 training implementation may also be a factor.

Trigger Words

As with my previous Kiss LoRA, there is no special trigger word.

Just describe the action directly.

For example:

• kiss

• tongue kiss

• passionately kiss

• spit

• and similar descriptions

Feel free to experiment with your own prompts.

Recommended LoRA Strength

I recommend starting with a LoRA strength of 0.5.

In my testing, this LoRA generally works best at relatively low strengths. I recommend keeping the strength below 0.7 whenever possible.

Suggested range:

• 0.4–0.5: Recommended for most cases

• 0.5: My current recommended default

• 0.6–0.7: Stronger effect, but may start affecting image quality or motion stability

• Above 0.7: Generally not recommended

Higher strength does not necessarily produce a better kissing motion. In some cases, pushing the LoRA too strongly may introduce more blur, artifacts, or unnatural motion.

For this experimental V0.1 release, I recommend starting at 0.5 and adjusting slightly depending on your prompt and reference images.

Future Plans

For now, I will most likely stop updating the Wan2.2 version of this LoRA.

My future development will mainly focus on MiniMax H3.

For the next version, I plan to experiment with newer H3 training methods, including the newly added AI Toolkit H3 Training Adapter, and possibly improve the training dataset with higher-quality video clips.

This V0.1 release should therefore be considered a first experimental H3 version, rather than a finished or fully optimized LoRA.

Feel free to test it, experiment with different H3 workflows, and share your results.

And finally, thank you to MiniMax for open-sourcing such an impressive model.