I created this LoRA by referring to This.
I would like to pay my respects to the creator of that LoRA.
All LoRA versions were created for I2V. Therefore, generating a new penis is not supported.

— New —

The LoRA for MiniMax H3 has been released.
Pls remember again, This lora only supports -I2V-

The uploaded lora was made with a very early version of the training tool, so I have labeled it as a beta version. All currently available tools are still unstable for I2V training, but now trying various methods to achieve better quality.

All sample videos were generated using the default MH3 example workflow.


---- Old News ----

(2026.4.27) Uploaded an LTX 2.3 version LoRA.
The training techniques previously used with WAN were carried over directly.
Since this version is a somewhat experimental LoRA, please refer to my sample video in ComfyUI and test it in various ways.
(This has not been tested in T2V. An NSFW LoRA maybe needed.)
Again, this is an “experimental” version. Since the step count is low, the results may not be very good. Various tests are currently being conducted, and it may be updated after retraining.

First of all, I did not train on any facial data. As far as I know, there are already enough (or many) facial-related LoRAs (such as cum on face, cum_in_mouth, etc.).

If want facial content, please use another great LoRA or use this one together with it (I have not tested the results when using them together).

-- WAN --
Using the TripleKsampler node gives even better results. I highly recommend trying it out.
https://github.com/VraethrDalkr/ComfyUI-TripleKSampler

Prompt used:
She moves hand rapidly up and down the shaft of her penis and she ejaculates a thick and huge amount white sticky liquid into air repeatedly from her penis.

I used 20 videos of 80 frames / 5 seconds each as the dataset.

Regarding liquid ejection, compared to Wan 2.1 the results are different, and it seems to have become more difficult to generate. However, if you use the workflow of the video I uploaded as it is, the results should not be bad.

The noise issue in the previous High_noise has been fixed, and older versions have been removed. The noise no longer occurs, and 121-frame sequences are now supported.

It might also be a good idea to adjust the lightX lora and This lora weights, or change the prompt, to create even better results.