(I've suddenly reached a new level of understanding for Wan2.2 I2V prompts. With these new prompts, I feel like I might not even need the v2.1 version of the LoRA anymore.)

My total step count is 8, with 3 high-noise steps and 5 low-noise steps. The high-noise and low-noise CFG scales are 2.5 and 1, respectively.

prompt： POV shot，handheld camera，The man's ????? rapidly moves out/slides out from the woman's ??????. The man fully exposes his ????? to the camera, and the woman exposes her ??????. The man's ????? then rapidly inserts into the woman's ?????? again. The ????? gradually disappears from tip to base in front of the camera. The man's ????? continuously and repetitively thrusts back and forth, from inside to outside, within the woman's ??????. A woman is straddling a ?????, engaging in ??????????????????. The woman is grinding on top of the ?????. The woman is rubbing her ????? on the ?????., she rides a ?????, moving her body up and down as the ????? slides in and out of her vagina，Viewing angle, nsfwsks,penis goes in and out of vagina，The ????? rapidly thrusts in and out of the vagina.,

————————————————————————————

Reward LoRA V1.0

The fused LoRA consists of the following components:

Wan2.2 I2V Reward LoRA

Wan ???? Posing ???? (Author: The_Cook)

Wan 2.1 - ????? ?????, ????, ????, LoRA, I2V (Author: aipinups69)

Video 1: Reward LoRA enabled. Video 2: Reward LoRA disabled.

（Video 1 has not yet been approved by the platform.）

Setting the Reward LoRA strength to 1 or higher might be better.

(This reward LoRA provides structures related to ??????? and ??????, so please use it in combination with other LoRAs.)

——————————————————————————————

V2.1

I have re-merged several models. The models I combined are: CubeyAI ((wan 2.2 experimental) WAN General ???? model), LocalOptima (POV Insertion [WAN 2.X]), and Kierkegaard420 (Wan2.2 - ????? (T2V/I2V)).

I performed this merge because, in some original images, the man's ????? was outside the woman's ??????. The goal of this merge was to ensure the man's ????? successfully enters the woman's vagina.

I recommend using the merged model at an intensity of 0.9 or below.

Additionally, I recommend using the v2.1 version of the merged model in conjunction with the LocalOptima model (??? FOV Slider [WAN 2.2]). When using it with the combined models listed above, it is essential to use them together only within a low-intensity range.

The specific prompt (or 'trigger word') may need to include the ??????. You can refer to the video I uploaded for prompt ideas.

prompt：(The man's ????? quickly retreated, and the man's ????? quickly removed from the camera, the woman exposed her ??????, and the ????? quickly moved into the camera and quickly inserted into the woman's ??????, and the ????? quickly pumped in the woman's vagina.nsfwsks)

——————————————————————————————————————

Previously, someone asked how the video sound effects are generated, so I’ve uploaded the workflow. If you include prompts like a girl ??????? during sex, I recommend not adding other NSFW-related sound effects at the same time.

With the hunyuanvideo_foley model, prompts really matter a lot.

——————

無駄無駄無駄無駄無駄！！

I'm having a lot of trouble resolving the shape mismatch between them.

It seems fine-tuning and knowledge distillation are indeed the only options.

v3.0

I decided to drop the idea of a “merge” and instead just forced 10KE10.safetensors (wan2.1 LoRA) to fit into wan2.2’s structure.
Here’s the rule I used:

• If a key exists in both → keep wan2.1’s version

• If a key only exists in wan2.1 → discard it

• If a key only exists in wan2.2 → keep it

So basically, it’s not really a merge, more like making wan2.1 conform to wan2.2.

Please set the strength of this LoRA around 0.35.
After a few small tests, I found the lighting effect was too strong, so I reduced the intensity of self_attn × 0.35.

When using this LoRA, please also add the following to your negative prompt 👉（harsh lighting, spotlight, overexposed, flashlight effect, glowing skin, strong highlights, dramatic lighting）Be sure to include similar prompts in the negative prompt; otherwise, the lighting effects will become as strange as in the uploaded video.

And for the positive prompt, make adjustments like 👉（2.5D anime, nsfwsks, ????? goes in and out of ??????, The man’s forceful, rhythmic thrusts drive his ???? into her, each movement causing his ????? to slide completely in and out of her ??????. Her ???????, soft and supple, bounce with his motions, their curves stretching and contracting with every thrust.）

————————————————

v2.0

The script is just too complicated-I'm throwing in the towel on trying to manually fuse this LoRA. It even started showing that bizarre Black-Box Transfer phenomenon. I've officially hit my limit, I surrender, nya🐱🏳️!

​もう無理。

(Honestly, training a new LORA is simply superior to merging the current LORA weights.)

————

In the previous LoRA merging process, shape differences were not taken into account, so only keys with identical shapes were retained. To better transfer the motion features from Wan2.1, I plan to perform reprojection on the matching keys and then fine-tune on Wan2.2 to improve the results.

————

For the merging process:

• Shared keys are combined using weighted averaging.

• Keys with mismatched shapes are retained from LoRA B (????-22-H/L-e8) to ensure compatibility.

• Keys unique to either model are preserved without modification.

————

v1.0

It seems that some users are experiencing incompatibility issues when using certain LoRAs, which are likely related to mismatched shapes. I'm currently working on resolving this.

Please do not download the v1.0 version of the file. I have re-uploaded high-noise and low-noise LoRAs. For this specific wan2.2 LoRA, I used a script to merge the wan2.1 LoRA (original author: basedbase) and the wan 2.2 LoRA (original author: CubeyAI).

————

My total steps were 8, with high-noise end steps at 3. The high-noise and low-noise CFGs were 2 and 1, respectively.

The sampler is Euler Ancestral, and the scheduler is normal.

I set the model strength to 1 or 0.9.

————

Key Comparison Result

• LoRA A (10KE10): 720 keys

• LoRA B (????-22-H-e8): 800 keys

• Shared keys: 600

• Keys only in A: 120

• Keys only in B: 200

Fusion Logic

• Shared keys: Perform weighted merging
  merged = α * A + (1 - α) * B

• Keys only in A: Keep A’s values

• Keys only in B: Keep B’s values

• Default α = 0.5

Source LoRAs:

10KE10.safetensors by basedbase – Enhances motion in ???? scenes, improving natural and smooth character movement in videos or multi-frame images.

NSFW-22-H-e8.safetensors by CubeyAI – ???? LoRA that strengthens ?????? scene details and style.

Fusion Logic:

Common keys (layers modified by both LoRAs) are merged via weighted averaging (default α = 0.5), balancing motion enhancement and ???? features.

Keys unique to 10KE10 are retained to preserve enhanced motion characteristics.

Keys unique to ????-22-H are retained to preserve ???? features.

The resulting LoRA combines motion enhancement with ???? reinforcement, ready for use in ComfyUI and other LoRA-supported pipelines.

Recommended Usage:

Load the merged LoRA alone to achieve both motion and ???? features; loading the original ???? LoRA is optional.

————

​I am using Kijai's wan2.2 i2v fp8 model. When testing it myself, I compared two generated videos with the seed and prompt fixed, one with a LoRA enabled and one with it disabled. When the LoRA was enabled, there was a noticeable improvement in the pumping/thrusting action of the ????? during the ?????? interaction of the characters.