Upload 2 files

Triggers. This LoRA activates with the following trigger word: The video begins with a shot of a person, the video then jumpcuts to the same person furiously masturbating at the same location wearing the same clothing, they are rapidly stroking their penis and sliding hand up and down the penis shaft while making an orgasm face, they ejaculate and white liquid cum squirts from the tip of their penis.. Include them in your positive prompt to pull the trained concept into the output. Without the trigger, the LoRA still has some effect, but it is usually much weaker.

Prompting. Wan Video uses natural-language sentences. For image-to-video, use separate narration-style prompts describing motion rather than image descriptions. Text-to-video prefers a more structured, robotic image-describing prompt, while image-to-video should describe action and motion like prose narrative. Prompt length and specificity significantly affect output quality. Add the trigger word (The video begins with a shot of a person, the video then jumpcuts to the same person furiously masturbating at the same location wearing the same clothing, they are rapidly stroking their penis and sliding hand up and down the penis shaft while making an orgasm face, they ejaculate and white liquid cum squirts from the tip of their penis.) alongside your normal prompt.

Weight. A good starting weight for most LoRAs is around 0.8, and the typical useful range is 0.6 to 0.9. From there, adjust based on what you see:

Too strong (bleeding into everything, distorted anatomy): drop to 0.5 to 0.7
Too weak (no visible effect): push to 0.9 to 1.0
Style LoRAs often want lower weights (0.5 to 0.7) than character LoRAs
If the author's description lists a recommended range, prefer that. They know their own training best.

Compatible checkpoints. Any Wan Video-family checkpoint should work. Popular choices include Wan 2.1, Wan 2.2, LTX Video, Hunyuan Video, and FramePack. Results vary by checkpoint, since a LoRA tuned against one finetune's aesthetic may look different on another even when the base family matches.

Base settings. Wan uses flow-matching, which requires different CFG behavior than SDXL: CFG 1.0 with speed LoRAs (lightx2v, CausVid) or CFG 3.5 or higher without speed LoRAs for better detail. Standard generation runs 20 to 30 steps at full quality (the res_2s sampler takes about twice as long but produces superior results). Output clips are limited to 81 frames, roughly 5 seconds at 16 fps, by default; extending beyond needs VACE extension or FramePack. Typical resolution targets are 480p (854x480) or 720p (1280x720), with upscaling via SeedVR2 or tile-based methods. Advanced workflows combine FramePack for longer clips, character LoRAs trained via musubi-tuner for consistency, and spacepxl control LoRAs for depth or tile conditioning on the 1.3B model.

Loading it. Video LoRAs are ComfyUI-only in practice. Save the file to models/loras/ and load it with a Load LoRA node (or Load LoRA Model Only in setups that route model and CLIP separately) placed between the checkpoint loader and the sampler in your video workflow. Chain multiple Load LoRA nodes to stack LoRAs. Automatic1111 and Forge do not run video models.

SuddenlyStroking_HIGH.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:578cffe6305ab1d14cdf8ac8b3d0b78b86605aae3c84f80db4e7c10559a9cf76
+size 153453576

SuddenlyStroking_LOW.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3a43e7733760b5002f2d3b48449d0a922af7de820b7d15b0dd657ef7ec96fef5
+size 153453576