🚫 Usage of this LoRA on external sites is NOT permitted.

A full bukkake in just 5 6 seconds, for wan 2.2 I2V.

v2

It's the same recipe as v1, it produces similar outputs, but this time it was trained at higher resolution, on longer videos, with an expanded dataset.

Noticeable improvements against v1:

🟢 anatomy looks much better.

🟢 the issue with backgrounds is fixed.

🟢 it became (a bit) more controllable.

The v2 structure is as follows:

Intro (5 frames) => Cumshot 1-4 (4*16 frames) => Outro (28 frames)

Prompting

A typical prompt would look like this:

A series of sh00tz. 
Scene 1: A woman smiles at the camera.
Scene 2: sh00t on forehead, from above. 
Scene 3: sh00t on cheek, from the left. 
Scene 4: sh00t on face, from the right. Extra penis on the left.
Scene 5: sh00t on tongue, from the left. Tongue out.
Scene 6: c0hf, she looks around her, laughing. 

Or simply copy and paste this into a wildcard node (repetitions are intentional):

A series of{| dutch angle| sideview| wide-angle| close-up} sh00tz.
Scene 1: A woman {speaks|fixes her hair|smiles at the camera|dances|walks}.
Scene 2: sh00t {on face|on face|on hair|in mouth|on cheek|on nose|on mouth|on teeth|on forehead|on eye}, from {the left|the right|the left|the right|above}. {|Hand {on her hair|under her chin}.}
Scene 3: sh00t {on face|on face|on hair|in mouth|on cheek|on nose|on mouth|on teeth|on forehead|on eye}, from {the left|the right|the left|the right|above}. 
Scene 4: sh00t {on face|on face|on hair|in mouth|on cheek|on nose|on mouth|on teeth|on forehead|on eye}, from {the left|the right|the left|the right|above}. {|Extra penis on the {left|right}.|Penis out of the frame.}
Scene 5: sh00t {on face|on face|on hair|in mouth|on cheek|on nose|on mouth|on teeth|on forehead|on eye}, from {the left|the right|the left|the right|above}. {|She spits.|She opens her eyes.|Her mouth is closed.}
Scene 6: c0hf, she is {confused|happy|amazed|stunned|dazzled|entertained}, she {looks around her|laughs|fixes her hair|talks|puts a finger into her mouth|puts her tongue out|smiles|clenches her teeth and closes her eyes|blows a kiss}. 

You can customize each line. Those examples should respond well:

Hand {on her hair|under her chin}.
Extra penis on the {left|right}.
Penis out of the frame.
Mouth {closed|wide open}.
Tongue out.
She spits.
She opens her eyes.

However, wan won't render the prompt exactly as it should most of times, and the prompt may bleed between cuts.

Also, writing explicit words like "penis" or "cum" is generally not needed, and may conflict with the keywords.

LoRA Files

There are 3 versions:

• Epoch 65 is slightly undertrained. It's the most flexible version, but issues may arise. Can be useful though, for combinating it with other loras, or if the face changes too much with the other variants.

• Epoch 75 is the default choice ; best balance between flexibility and robustness. Start with this one.

• Epoch 85 is slightly overtrained. It produces more details, but it's more rigid.

Settings

Getting started

This video from the showcase embeds a simple, easily reproducible workflow.

The other videos contain a workflow with latent upscaling between the high-noise and the low-noise stages.

Suggestions

• Frames : 97.

• LoRA Strength: 0.9-1.0.

• Steps: evenly split, like 4+4.

• Shift: between 7 and 8.

• Samplers: euler, or ipndm.

• Schedulers: simple or beta.

• Run the first high-noise steps without speedup loras: you will get better prompt adherence.

• To get less cuts: lower the high-noise lora's strength.

v1

Trained on 24 81-frames square videos, all of which were structured this way:
initial state (5 frames) => cumshots (4x16 frames) => end state (12 frames).

Just start your prompt with the keyword "sh00tz", and wan will know what to do.

On the high-noise model: works best without any lightx2v lora and cfg at 3.5 (tested with 4/8 steps): semen looks more liquid and natural.

On the low-noise model: add dr34ml4y LoRa (for example, at 0.7 strength) for better realistic outcomes.

Consistency is relatively good most of times, so you should see a coherent progression between cuts.

Wan may draw a weird background floor. You can add "within the same background" in your prompt and/or "background changes" in the negative prompt to mitigate that effect, or just describe the location corresponding to your source image.

Since the training data focused mostly on the face, use a vertical format (like 9:16 or even 1:2) to include more of the woman's body within the frame.