--- tags: - text-to-image - lora - diffusers - template:diffusion-lora widget: - output: url: images/videoframe_1090.png text: '-' base_model: Qwen/Qwen-Image-2.1-Turbo instance_prompt: null --- # closecum ## Model description Prompt examples Here are a few complete example prompts. Look further down in the prompting section for a few more details. the video shows an erect penis shooting monstrous amounts of cum into woman's face with fast rhythm. multiple thick slimy cum shots. every cum shot sticks thickly to her face. her whole face gets buried in slimy cum. the cum is sticky. the video shows an erect penis shooting abundant amounts of cum into woman's face. the separate cum shots spurt with very fast rhythm from the penis. countless thick slimy cum shots erupt. every cum shot sticks slimily to her face. her whole face gets buried in sticky cum. penis shoots cum. the cum shots are huge and substantial, thick and girthy. the texture and consistency of cum is sticky and gelatinous. the cum shots explode forcefully from the penis. the cum shoots with quick steady rhythm. slimy cum hangs from her face. Recommended settings 🛈 All the recommended settings given here are with Lightx2v being used. I have not explored settings without it. Strength: 1.0 1.0 works best for both the high and low-noise models. You can go higher for the high-noise model to exaggerate the concepts — at cost of strange behavior and image quality. CFG High-noise: 1.0 - 2.0 1.0 is the baseline. Increase to get thicker, more exaggerated cum. It seems higher resolution generations require higher CFG to maintain this effect. Issues appear at high values — usually related to lighting. I've had 4.0 and above work for some prompts, but it depends. Low-noise: 1.0 For the low-noise model I leave it at 1.0. I haven't seen much difference when changing it. Steps: High 4 + Low 4 4 steps for both high and low works well. High-noise: 4 I consider 4 steps the minimum for the high-noise model. At lower counts some issues may appear: 💀 If cumshots "erase" each other you may have too few steps for the high-noise model. Try increasing. 💀 Similarly if cum appears on the model's face out of nowhere while no cum shoots from the penis — increase the step count for high-noise. Low-noise: 2 - 6 4 steps is a good middle-ground for the low-noise. Additional steps can give more detail in the cum. I've not seen much improvement from going higher than 6. Fewer steps (2-3) can make it appear more voluminous but with less detail. Sampling shift: 2.0 - 12.0 You can leave this around 2 - 12. The model will generate slightly differently based on this value. The valid range goes up to 100 but results degrade the higher you go in my experience. If the cum doesn't look or behave the way you want you can try adjusting the shift. I usually change it 1.0 at a time with fixed seed, exploring the 2.0 - 8.0 range to see if results improve. Prompting Prompt using natural language and use keywords from the tags list (see tags used below). Describe the scene and other things as you would normally for Wan 2.2. Cumshot rhythm / cadence It usually responds to phrasings like these: slow rhythm cumshots fast rhythm cumshots very fast rhythm cumshots Cum amount It usually responds to phrasings like these: monstrous amounts of cum abundant amounts of cum huge cum shots big shots of cum Start image without penis To have a penis appear during the video, prompt it into existence. For example: penis enters from the left penis appears on right My workflow I create and render my start images in Blender. Many great 3d resources are available at Smutbase and PolyHaven. I include a penis that shoots the first cumshot. By changing the size of the start shot you often control the size of shots in the video. I use Blender's motion blur to control the force/velocity of the shots. Creating your start image from a 3d scene also gives full control over lighting. 🛈 Most of my published videos have the workflow embedded. The released model uses different file names, so if my workflows are missing the LoRA files simply replace them with the high and low from this release. Known issues When the penis is far from the target's face cum streams can become suspended mid-air. This LoRA is very keen to connect streams between penis and face because it doesn't know it shouldn't happen over longer distances. Generated penises don't look good. Overcome this by using a penis LoRA or having one in the input image. The trained concepts don't work equally well from all camera angles and framings. It does some things a lot better from certain angles. This is due to my small dataset. Sometimes the model wants to make cum appear from above. To stop this avoid words like drench, splash, spray and similar. over, above and the like may also be problematic. My guess is it's triggered by words related to water concepts in the base model. Lowering CFG on the high-noise model can also help. I want to improve this LoRA but I only have limited time, so don't expect frequent updates. About the dataset The dataset was 26 clips of 5 second length. High-noise model was trained at 256 x 256 and the low-noise at 512 x 512. The average number of tags is 11.8 per sample video and the average number of words is 35. Tags used These are the most frequently used tags of the dataset. Use words from these in your prompts; it's not necessary to use the whole tags verbatim. cum spurts from penis cum shoots from penis cum drips from penis cum is sticky cum on face much cum on face cum on tongue man strokes penis penis touches mouth woman is happy woman is calm woman smiles woman tongue is extended woman mouth is open woman eyes are closed woman blinks eyes woman eyes flutter woman grimaces woman looks up woman looks into camera top-down view side view profile view front view static shot handheld camera erratic camera movement Roadmap ## Download model [Download](/fghui/closeupcum/tree/main) them in the Files & versions tab.