LoRA Overview｜Anime Cumshot Aesthetics – Precision Load
Anime Cumshot Aesthetics – Precision Load is a LoRA specifically trained to enhance the visual and stylistic quality of cumshots in anime-style videos, focusing on density, precision, impact, and aesthetic realism.
This model was developed to complement existing cumshot LoRAs like WAN CUMSHOT by focusing specifically on anime-style aesthetics and short-form timing. While general-purpose models aim for broad adaptability, Anime Cumshot Aesthetics – Precision Load targets refined features such as thickness, directional accuracy, and artistic expression, delivering paste-like cumshots with strong visual impact in tightly timed scenes.
All my video templates are upscaled, smoothed, and interpolated using my V2V (video-to-video) workflow.
About this version 2025/06/23 {UPDATE}
🎞 WAN 2.1 Anime Cumshot – Emotion Edition (LoRA)
🧠 Description
Anime Cumshot – Emotion Edition is a specialized LoRA built upon Anime Cumshot Aesthetics – Precision Load.
This version adds expressive facial and physical reactions during cum scenes, delivering dynamic and emotionally charged outputs that reflect anime-style shock, anger, acceptance, and more.
All my video templates are upscaled, smoothed, and interpolated using my V2V (video-to-video) workflow.
Wan v2v with Upscaling and Smoothing and Interpolation 放大平滑處理 - v1.03 use this version | Wan Video 14B i2v 480p Upscaler | Civitai
📁 Training Details
Clips: 33 original animated clips
Length: 3 seconds per clip
FPS: 16
Style: 100% anime
Trained Reactions: Facial expressions, blinking, flinching, head movement, eye tracking
LoRA Type: Emotion-focused enhancement
✅ Compatibility Notes and Personal Tips
This LoRA is designed primarily for anime-style scenes, but also works reasonably well in semi-realistic settings.
If you are working with realistic characters or lighting, I recommend pairing it with WAN CUMSHOT to achieve more natural pacing and impact.
Wan Cumshot - I2V v1.0 | Wan Video 14B i2v 480p LoRA | Civitai
In my experience, combining with leads to smoother results, especially in timing and shot rhythm.Emotion = 1.0 WAN CUMSHOT = 0.2~0.5 (You can try different weights, but try not to exceed my LORA's range.)
You can also use this LoRA alone without any problems in anime-style compositions. However, depending on your base model and prompt structure, some users may notice that the ejaculation and reaction timing may appear slightly slow or delayed.
This is a subjective issue I’m still investigating — I haven’t found a reliable method to speed up perceived motion yet.
🧪 Suggested Settings
Steps: 20 (You can try increasing the steps. Sometimes I use 22 to 24.)
CFG Scale: 6 (If you want your image to adhere more closely to your prompt, try increasing the CFG. Sometimes I use 6.2)
Sampler: UNI_PC
Scheduler: SIMPLE
🎯 Supported Emotional Reactions
😲 Surprise
・Eye-widening, shocked expression
・Visual Cue: flinch + wide eyes
😧 Disbelief
・Slightly open mouth, frozen/stunned expression
・Visual Cue: disbelief + open mouth
😐 Calm / Still
・Blank face, slow blinking, little to no reaction
・Visual Cue: slow blink + still face
😠 Anger
・Glaring, clenched jaw, tense expression
・Visual Cue: glare + clenched jaw
😊 Soft Smile
・Relaxed gaze, faint smile, subtle head lift
・Visual Cue: soft smile + head tilt
😒 Displeasure
・Subtle frown, quiet rejection or visible discomfort
・Visual Cue: frown + slight recoil
✔️ Additional Supported Reactions:
・Head tilt (slight up/down)
・Turning head (left/right)
・Eye closure / eye tracking
・Subtle recoil / delayed reactions
⚠️ Current Limitations (Work in Progress)
Prompt Compatibility: Works best with fixed, tested prompts. Varying wording might reduce stability.
Blended Emotions (e.g., shock → smile) are not yet stable.
Intensity Control: No fine control over light vs. intense emotions yet. All output is medium strength.
Limited Targeting: Does not yet support fine-tuned cum landing (e.g., thigh-only, chest-only).
💡 Recommended Pairing (if needed)
Anime Cumshot Emotion: 1.0WAN CUMSHOT: 0.3Wan Cumshot - I2V v1.0 | Wan Video 14B i2v 480p LoRA | Civitai
I have not tested compatibility with other LoRAsThis LORA can be used completely standalone without any issues, or you can try mixing it with other LORAs.
I've tried to improve the angles, but I haven't been able to solve very unique angles yet. You can try using it for specific body parts, which I also haven't figured out a solution for yet. For example, scenes focusing only on ejaculating on breasts or on buttocks.
I've tried a lot of source material, but haven't found a good solution. The only way might be to extract frames from videos and focus only on material of breasts or buttocks. I haven't tried that yet, but I'll give it a shot when I have time.
PROMPT:
Mix and match as desired. However, success isn't guaranteed, so please try it yourself or refer to my video examples.
A girl is kneeling
A girl is sitting.
A girl is standing.
A girl is lying down. If you want the man to not be in the frame at the beginning and then enter the frame from off-screen, use this.
A man's lower body and penis enter the frame from below.
A man's lower body and penis enter the frame from the right side.
A man's lower body and penis enter the frame from the left side.Main actions and landing points
The man is stroking his penis. He ejaculates on her face, in her hair, on her chest, into her mouth, on her tongue, stomach,If you want the woman to help the man, use this. For other scenarios where the woman helps the man, like using her breasts, I apologize, but I don't have that material. However, you can try it; theoretically, the ejaculating action shouldn't disappear.
She strokes his penis.Describe the state of the semen （Please use this as a reference to capture a similar feel）
The cum is thick and sticky, clinging like paste before sliding off.
The cum is thick and sticky, sticking to her skin before dripping down slowly.
The cum is thick and gooey, clinging to her lips and tongue before dripping off.
The sticky fluid briefly clings before slowly sliding off her skin.
It is thick and sticky, clinging like paste before sliding off slowly.
The sticky fluid clings briefly before slowly sliding off her skin.
The cum is thick and viscous, sticking tightly before sliding down her body.
Thick and heavy, the cum holds to her body before dripping away.
Emotion：
Here are all my available packages for your reference. You're welcome to mix and match, but I haven't tested too many combinations for success, so please try it yourself!
Her face shifts in an instant. She looks shocked, then tilts her head upward, staring toward the source of the cum.
A sudden jolt crosses her face. She flinches in shock, slightly recoiling as she slightly turns her head to the right.
She reacts immediately. Her eyes widen in shock, then she turns toward the source of the cum, looking very angry.
There is an instant shift in her expression. She flinches in shock, her eyes wide with disbelief.
She blinks slowly, without changing her expression.
A wave of shock washes over her face. She flinches in shock, slightly recoiling.
A sudden emotional shift appears on her face. She looks shocked, then turns toward the source of the cum, looking very angry.
A sudden change crosses her face. She flinches in shock, her eyes wide with disbelief.
There is a subtle shift in her expression. She blinks slowly, frowning faintly, her expression calm but clearly displeased.
A moment of disbelief flashes across her face. She looks incredulous, then tilts her head upward, staring toward the source of the cum.
Her expression remains unchanged.
Her face tenses briefly. She tilts her head upward, eyes following the source of the cum.
A quiet reaction passes over her face. She closes her eyes for a moment, frowning faintly. Her lips press together tightly, her face calm yet clearly displeased.
A quick flash of shock crosses her face. She flinches in surprise, her eyes wide with disbelief.
Initially, she is momentarily stunned. Then, flinching in shock, her eyes squeeze shut, and she instinctively recoils, noticeably turning her head to the right.
Her expression changes instantly. She furrows her brow, her mouth agape in disbelief and rage. She then looks up and glares toward the source of the cum, clearly furious.
As the fluid hits, she flinches in surprise, blinking a few times, then noticeably turns her head to the right.
A sudden wave of emotion strikes her face. She flinches in shock, her eyes wide with disbelief.
Her face is flushed. She stares ahead with hazy eyes, seemingly dazed yet still conscious, her expression calm and unresisting. Her left eye had briefly closed before reopening.
Her features soften. She smiles faintly and keeps her gaze on his penis without looking away.
Her brows twitch faintly. She closes her eyes for a moment, then reopens them with a slight smile.
A flash of shock flickers across her face. She flinches, her eyes wide with disbelief.
Her expression shifts suddenly. She flinches, then turns toward the source of the cum, her eyes wide with disbelief.
Her features soften slightly. With a soft smile, she lifts her head and looks up at him, letting the cum splash across her face and chest.
A sharp reaction crosses her face. She flinches in shock, her eyes wide open, then turns toward the source of the cum.
Her expression remains unchanged, her eyes simply closing.
She shifts her head slightly and watches him quietly, expression calm but aware.
A quiet disapproval flickers across her face. She blinks slowly, frowning faintly, then raises her head, her expression calm but clearly displeased.
As the fluid hits, her eyes widen in shock, and her mouth slightly opens, as she noticeably turns her head to the left.I usually put these supporting prompts below the main prompts I've written. This is a personal habit I find useful; you can decide if you want to use them yourself.
White translucent cum.
anime style, open eyes, eyelids match skin tone, Natural eyelids, anime eyelashes, simple eyelashesHere are my negative prompts for your reference.
blurry, out of focus, low detail, bad anatomy, ugly, overexposed, underexposed, distorted face, extra limbs, cartoonish, 3d render artifacts, duplicate people, unnatural lighting, bad composition, missing shadows, low resolution, poorly textured, glitch, noise, grain, static, motionless, still frame, overall grayish, worst quality, low quality, JPEG compression artifacts, subtitles, stylized, artwork, painting, illustration, cluttered background, many people in background, three legs, walking backward, zoom out, zoom in, mouth speaking, moving mouth, talking, speaking, mute speaking, unnatural skin tone, discolored eyelid, red eyelids, red upper eyelids, no red eyeshadow, closed eyes, no wide-open innocent eyes, poorly drawn hands, extra fingers, fused fingers, poorly drawn face, deformed, disfigured, malformed limbs, realistic penis, thighs, fog, mist, eyelashs, realistic eyelashes, detailed eyelashes, voluminous eyelashes, blush,If something like fog or mist appears, please add it to the negative prompts.
⚠️ Disclaimer:
All characters are AI-generated and do not represent real people.
Content generated by this model is AI-created and may not be entirely accurate or reflect reality. Please do not rely on it as factual.
You are solely responsible for any copyright infringement or other intellectual property issues arising from content you generate using this model.
We do not guarantee the continuous availability or error-free operation of this model. We reserve the right to modify or terminate the service at any time.
You are solely responsible for any content you generate and how it is used.
This LORA is for academic exchange only and has not been shared on any other websites. Please do not repost or reupload it. It is only permitted for use on Civitai
Description
This LoRA is compatible with both live-action and anime styles.
This LoRA is compatible with both live-action and anime styles.
This LoRA is compatible with both live-action and anime styles.
LoRA Overview｜Anime Cumshot Aesthetics – Precision Load
Anime Cumshot Aesthetics – Precision Load is a LoRA specifically trained to enhance the visual and stylistic quality of cumshots in anime-style videos, focusing on density, precision, impact, and aesthetic realism.
This model was developed to complement existing cumshot LoRAs like WAN CUMSHOT by focusing specifically on anime-style aesthetics and short-form timing. While general-purpose models aim for broad adaptability, Anime Cumshot Aesthetics – Precision Load targets refined features such as thickness, directional accuracy, and artistic expression, delivering paste-like cumshots with strong visual impact in tightly timed scenes.
All my video templates are upscaled, smoothed, and interpolated using my V2V (video-to-video) workflow.
---------------------------------------------------------------------------------------------------
About this version 2025/06/23 {UPDATE}
🎞 Anime Cumshot – Emotion Edition (LoRA)
🧠 Description
Anime Cumshot – Emotion Edition is a specialized LoRA built upon Anime Cumshot Aesthetics – Precision Load.
This version adds expressive facial and physical reactions during cum scenes, delivering dynamic and emotionally charged outputs that reflect anime-style shock, anger, acceptance, and more.
All my video templates are upscaled, smoothed, and interpolated using my V2V (video-to-video) workflow.
Wan v2v with Upscaling and Smoothing and Interpolation 放大平滑處理 - v1.03 use this version | Wan Video 14B i2v 480p Upscaler | Civitai
📁 Training Details
Clips: 33 original animated clips
Length: 3 seconds per clip
FPS: 16
Style: 100% anime
Trained Reactions: Facial expressions, blinking, flinching, head movement, eye tracking
LoRA Type: Emotion-focused enhancement
✅ Compatibility Notes and Personal Tips
This LoRA is designed primarily for anime-style scenes, but also works reasonably well in semi-realistic settings.
If you are working with realistic characters or lighting, I recommend pairing it with WAN CUMSHOT to achieve more natural pacing and impact.
In my experience, combining with leads to smoother results, especially in timing and shot rhythm.Emotion = 1.0 WAN CUMSHOT = 0.3
You can also use this LoRA alone without any problems in anime-style compositions. However, depending on your base model and prompt structure, some users may notice that the ejaculation and reaction timing may appear slightly slow or delayed.
This is a subjective issue I’m still investigating — I haven’t found a reliable method to speed up perceived motion yet.
🧪 Suggested Settings
Steps: 20 (You can try increasing the steps. Sometimes I use 22 to 24.)
CFG Scale: 6 (If you want your image to adhere more closely to your prompt, try increasing the CFG. Sometimes I use 6.2)
Sampler: UNI_PC
Scheduler: SIMPLE
🎯 Supported Emotional Reactions
😲 Surprise
・Eye-widening, shocked expression
・Visual Cue: flinch + wide eyes
😧 Disbelief
・Slightly open mouth, frozen/stunned expression
・Visual Cue: disbelief + open mouth
😐 Calm / Still
・Blank face, slow blinking, little to no reaction
・Visual Cue: slow blink + still face
😠 Anger
・Glaring, clenched jaw, tense expression
・Visual Cue: glare + clenched jaw
😊 Soft Smile
・Relaxed gaze, faint smile, subtle head lift
・Visual Cue: soft smile + head tilt
😒 Displeasure
・Subtle frown, quiet rejection or visible discomfort
・Visual Cue: frown + slight recoil
✔️ Additional Supported Reactions:
・Head tilt (slight up/down)
・Turning head (left/right)
・Eye closure / eye tracking
・Subtle recoil / delayed reactions
⚠️ Current Limitations (Work in Progress)
Prompt Compatibility: Works best with fixed, tested prompts. Varying wording might reduce stability.
Blended Emotions (e.g., shock → smile) are not yet stable.
Intensity Control: No fine control over light vs. intense emotions yet. All output is medium strength.
Limited Targeting: Does not yet support fine-tuned cum landing (e.g., thigh-only, chest-only).
💡 Recommended Pairing (if needed)
Anime Cumshot Emotion: 1.0WAN CUMSHOT: 0.3
I have not tested compatibility with other LoRAsThis LORA can be used completely standalone without any issues, or you can try mixing it with other LORAs.
I've tried to improve the angles, but I haven't been able to solve very unique angles yet. You can try using it for specific body parts, which I also haven't figured out a solution for yet. For example, scenes focusing only on ejaculating on breasts or on buttocks.
I've tried a lot of source material, but haven't found a good solution. The only way might be to extract frames from videos and focus only on material of breasts or buttocks. I haven't tried that yet, but I'll give it a shot when I have time.
PROMPT:
Mix and match as desired. However, success isn't guaranteed, so please try it yourself or refer to my video examples.
A girl is kneeling/ A girl is sitting./ A girl is standing./ A girl is lying down. If you want the man to not be in the frame at the beginning and then enter the frame from off-screen, use this.
A man's lower body and penis enter the frame from below. A man's lower body and penis enter the frame from the right side. A man's lower body and penis enter the frame from the left side.Main actions and landing points
The man is stroking his penis. He ejaculates on her face, in her hair, on her chest, into her mouth, on her tongue, stomach,If you want the woman to help the man, use this. For other scenarios where the woman helps the man, like using her breasts, I apologize, but I don't have that material. However, you can try it; theoretically, the ejaculating action shouldn't disappear.
She strokes his penis.Describe the state of the semen （Please use this as a reference to capture a similar feel）
The cum is thick and sticky, clinging like paste before sliding off.
The cum is thick and sticky, sticking to her skin before dripping down slowly.
The cum is thick and gooey, clinging to her lips and tongue before dripping off.
The sticky fluid briefly clings before slowly sliding off her skin.
It is thick and sticky, clinging like paste before sliding off slowly.
The sticky fluid clings briefly before slowly sliding off her skin.
The cum is thick and viscous, sticking tightly before sliding down her body.
Thick and heavy, the cum holds to her body before dripping away.
Emotion：
Here are all my available packages for your reference. You're welcome to mix and match, but I haven't tested too many combinations for success, so please try it yourself!
Her face shifts in an instant. She looks shocked, then tilts her head upward, staring toward the source of the cum.
A sudden jolt crosses her face. She flinches in shock, slightly recoiling as she slightly turns her head to the right.
She reacts immediately. Her eyes widen in shock, then she turns toward the source of the cum, looking very angry.
There is an instant shift in her expression. She flinches in shock, her eyes wide with disbelief.
She blinks slowly, without changing her expression.
A wave of shock washes over her face. She flinches in shock, slightly recoiling.
A sudden emotional shift appears on her face. She looks shocked, then turns toward the source of the cum, looking very angry.
A sudden change crosses her face. She flinches in shock, her eyes wide with disbelief.
There is a subtle shift in her expression. She blinks slowly, frowning faintly, her expression calm but clearly displeased.
A moment of disbelief flashes across her face. She looks incredulous, then tilts her head upward, staring toward the source of the cum.
Her expression remains unchanged.
Her face tenses briefly. She tilts her head upward, eyes following the source of the cum.
A quiet reaction passes over her face. She closes her eyes for a moment, frowning faintly. Her lips press together tightly, her face calm yet clearly displeased.
A quick flash of shock crosses her face. She flinches in surprise, her eyes wide with disbelief.
Initially, she is momentarily stunned. Then, flinching in shock, her eyes squeeze shut, and she instinctively recoils, noticeably turning her head to the right.
Her expression changes instantly. She furrows her brow, her mouth agape in disbelief and rage. She then looks up and glares toward the source of the cum, clearly furious.
As the fluid hits, she flinches in surprise, blinking a few times, then noticeably turns her head to the right.
A sudden wave of emotion strikes her face. She flinches in shock, her eyes wide with disbelief.
Her face is flushed. She stares ahead with hazy eyes, seemingly dazed yet still conscious, her expression calm and unresisting. Her left eye had briefly closed before reopening.
Her features soften. She smiles faintly and keeps her gaze on his penis without looking away.
Her brows twitch faintly. She closes her eyes for a moment, then reopens them with a slight smile.
A flash of shock flickers across her face. She flinches, her eyes wide with disbelief.
Her expression shifts suddenly. She flinches, then turns toward the source of the cum, her eyes wide with disbelief.
Her features soften slightly. With a soft smile, she lifts her head and looks up at him, letting the cum splash across her face and chest.
A sharp reaction crosses her face. She flinches in shock, her eyes wide open, then turns toward the source of the cum.
Her expression remains unchanged, her eyes simply closing.
She shifts her head slightly and watches him quietly, expression calm but aware.
A quiet disapproval flickers across her face. She blinks slowly, frowning faintly, then raises her head, her expression calm but clearly displeased.
As the fluid hits, her eyes widen in shock, and her mouth slightly opens, as she noticeably turns her head to the left.I usually put these supporting prompts below the main prompts I've written. This is a personal habit I find useful; you can decide if you want to use them yourself.
White translucent cum.
anime style, open eyes, eyelids match skin tone, Natural eyelids, anime eyelashes, simple eyelashesHere are my negative prompts for your reference.
blurry, out of focus, low detail, bad anatomy, ugly, overexposed, underexposed, distorted face, extra limbs, cartoonish, 3d render artifacts, duplicate people, unnatural lighting, bad composition, missing shadows, low resolution, poorly textured, glitch, noise, grain, static, motionless, still frame, overall grayish, worst quality, low quality, JPEG compression artifacts, subtitles, stylized, artwork, painting, illustration, cluttered background, many people in background, three legs, walking backward, zoom out, zoom in, mouth speaking, moving mouth, talking, speaking, mute speaking, unnatural skin tone, discolored eyelid, red eyelids, red upper eyelids, no red eyeshadow, closed eyes, no wide-open innocent eyes, poorly drawn hands, extra fingers, fused fingers, poorly drawn face, deformed, disfigured, malformed limbs, realistic penis, thighs, fog, mist, eyelashs, realistic eyelashes, detailed eyelashes, voluminous eyelashes, blush,If something like fog or mist appears, please add it to the negative prompts.
FAQ
Comments (80)
Emotion Response Version Update
I've updated the emotion response version, and this is still a test version. Currently, I can only add material for Surprise, Disbelief, Calm/Still, Anger, Soft Smile, and Displeasure.
I have the most material for Surprise and Disbelief. This is due to my personal preference and because they're the easiest for me to create. Therefore, the emotion responses for these two should theoretically be the most stable.
For Calm/Still, I don't have much—only about 3-4 pieces. My main goal with these is to increase generalization and neutralization, so that no single reaction overpowers the others. Of course, this is just my theory, and I'm not sure if it will work.
Anger has been the most challenging for me because the expressions tend to distort too easily, and I don't like that. For me, it's about finding the smallest facial movement that can still convey a clear sense of anger with a noticeable furrowed brow. I haven't been able to produce the kind of "exploding with rage" material I want, so please don't have high expectations for that.
I only use Soft Smile when it's for a woman giving a man a handjob or in a lying-down pose. You can try it out yourselves.
I have a fair amount of material for Displeasure, but I feel like it could be categorized under Anger. However, sometimes it's not very obvious.
Facial expressions are really difficult to get right. Just completing all the different angles and trying to increase generalization is tough, so I can't guarantee this version is perfect. However, I personally feel this version has been successfully trained. At least for me, I can usually get the desired output on the first try.
The main difficulty is that I can't control the intensity of the emotions. That's about it for now. I've been working on this for the past eight days, and it's been exhausting!
I feel like a lot of people would enjoy an aroused response. I know I would.
I have no idea who in history came up with the idea that a p3nis should be shaped like a mushroom,
turning it into a trend.
Wherever you are, may you be damned!
You've ruined years of training.
ah. thanks for sharing
This isn't actually what I wanted; the AI made it this way. I never intended for it to look like a mushroom. It's the base model's recognition capabilities that caused this. All my original images were of real penises, but after the WAN dynamic video process, it turned out like this. You can't blame me. You could try adding a PENIS LORA to see if it improves. I've done my best. I don't want it to be strangely shaped either. Otherwise, all my long hours of training would be for nothing, and I have no solution.
@FLOW0308 nah brother it wasnt a message for you but to who invented that shape 🤣 there are tons of anime loras and models that make the penis looks like that. is not you again thanks for sharing
@p_p My bad, I must not have read carefully. I misunderstood what you meant.
2025/06/25 UPDATE
Hey everyone, I want to clarify something regarding the eyelash density issue that some of you have noticed.
This isn't actually a problem with my LoRA. I believe it's because WAN itself isn't very friendly to anime styles. It tends to automatically convert eyelashes to look more like realistic human eyelashes. That's why I've added the following auxiliary prompts to the top of my prompt suggestions for everyone to see:
White translucent cum.
anime style, open eyes, eyelids match skin tone, Natural eyelids, anime eyelashes, simple eyelashes
I also strongly recommend checking the negative prompts I've included at the very bottom of the description. I've specifically added negative prompts related to eyelashes.
The eyelashes becoming denser or the upper eyelids getting darker are things that WAN automatically perceives as normal, but in an anime style, these issues theoretically shouldn't appear. So, I genuinely can't prevent it. The only thing I can do is add relevant keywords about eyelashes and upper eyelids to both the positive and negative prompts.
If anyone has a better solution or can help me train to improve this, I'm open to it! Thanks.
This lora has been absolutely amazing for me so far. But there is one problem and I don't know if this lora is causing it but... The video seems to gradually get more and more underexposed as the video goes on. I do about 81 frames at 16 fps. Any idea why?
Which version? emotion? no emotion?
If you're referring to the 'No Emotions' version, that's my earliest iteration and the one with the most flaws, as it was my first attempt at training. My latest 'Emotion Edition' should have already fixed that problem. If you insist on using the 'No Emotions' version, please try pairing it with WAN CUMSHOT to resolve the issue. Of course, the 'Emotion Edition' can also be used with WAN CUMSHOT; in theory, my latest version should have already fixed the darkening problem.
@FLOW0308 I am using the latest version. I am using 3 loras: Your lora, cumshot lora, and general nsfw lora at weights of 1.0, 0.4, and 0.9, respectively. I'm using the Q4 480p gguf model, and I also tried putting prompts such as "even lighting, bright ambient lighting, consistent illumination, evenly lit." And although it does work most the time, about 1 in 4 videos get that super dark look to it.
@jaykob If video become fuzzy as goes on, either your denoise strength is <1.0 or shift is low or both,
If not fuzzy but low contrast your cfg can be low <2.
If multi sampler, total steps can be low, or first sampler "add denoise" is disabled.
@jaykob @spittreyways Professionally speaking, it's likely as you said: my latest version should have already resolved the screen darkening issue.
My workflow only uses one Sampler, with CFG at 6 and Shift at 5. I've tested a few 5-second generations without issues, but it's possible I haven't tested enough. However, in theory, pairing it with WAN CUMSHOT has basically solved all problems for me. I haven't really tested what happens when a third LoRA is added, so I don't know the impact there.
I usually stick to 3-second generations; if it goes over 3 seconds, I use WAN CUMSHOT—for example, my LoRA at 1 and WAN CUMSHOT at 0.3-0.5. I haven't tried adding a third LoRA on top of that. It's odd, though; in theory, the darkening issue shouldn't reappear with my latest version. My training material is also balanced for screen brightness, which is much better than the first version, where darkening was guaranteed. This is strange
@jaykob I know there's another possible solution: I could change all the footage to 5 seconds. But I'm not sure if my computer can handle it ＞＜
@FLOW0308 Another important thing to mention is that I use a native workflow with tea cache and skip layer guidance. I get confused when people talk about the "shift". Is that something exclusive to Kijai's wrapper? I only use native since that is the only thing compatible with my measly 8gb vram. I've heard people say it's the node called ModelSamplingSD3 but i'm not sure if that's correct or where to put that in my workflow. What does it even do in the first place?
@jaykob ModelSamplingSD3
Currently, I can only assess that there might be a conflict between different LORAs, but I don't know the exact reason why. So far, I've only used my LORA in combination with cumshot-related LORAs and haven't encountered the darkening issue yet. It's possible I just haven't tested it for longer than 3 seconds.
I'm unsure if this problem arises when combining my LORA with LORAs related to other sex poses. I simply don't have the time to go through and test every single LORA combination to rule out the problem. Even if I were to find a specific conflict, I wouldn't know why it causes darkening, and I likely wouldn't be able to fix it.
At that point, my only suggestion would be to try a different LORA combination. If my LORA has this darkening problem when paired with every other LORA for over 5 seconds, then I'd be out of ideas. I'm just training LORAs, and I really don't know why this issue would occur.
@FLOW0308 I've done some experimentation and it doesn't seem like Loras are the problem at all. If anything it makes the problem worse. I think what's causing it is the length of my videos since they are about 5-6 seconds long. And after increasing the length to 8 seconds, I confirmed that it got even darker. But honestly that's ok since the results itself are really, really good. I can just simply take my output, drag it onto a video editing software (I'll probably just use shotcut) and slowly adjust the brightness overtime. Still looks clean and I get to use your lora for longer videos potentially. I would be even more happy if you manage to improve this lora even better overtime, but for now this is my best solution. 3 seconds just won't cut it for me lol.
@jaykob Honestly, I still can't figure out the reason. Even after asking around in DC (Discord) groups, no one has responded, so it's possible nobody knows the answer. In theory, I've already significantly increased the amount of my source material. The only thing left I can think of trying is extending the duration of each clip to 5 seconds to see if that resolves the issue.
However, that would be a huge undertaking since I'd have to re-create all my source material from scratch. My plan is to wait until I've implemented all my backup plans for the 3-second clips. Then, I'll use the saved backup data to expand on. This seems like the quicker way for me to achieve my goal, as I'm confident that my prompts and seeds are reproducible. If I were to spend time on 5-second clips from the very beginning, video creation would also take a lot of time, and there's no guarantee I'd even achieve the desired visuals.
@FLOW0308 I have this happen when working with some anime images. I use WAN on magespace, basically normal setting, and i do image to video mostly. So if youre speaking about using an anime image, this happens for me too. It will get really dark, or really bright as the video goes on.But real life style pictures never have this issue for me. I dont know if this helps at all but it's just my experience with wan video loras and models so far, not just specifically this one.
@mousespirit In my next version, I'll try a different training approach and incorporate more training material. If I still encounter issues, then I'll just have to accept it. I'll be powerless and lack the ability to do more.
@FLOW0308 Thank you for your work and time very much
@FLOW0308 This is just another question, but do you have any tips or advice on how to control the AMOUNT of liquid generated? My outputs seem to vary a lot with some having almost nothing coming out, while some have a good amount (my preferred output). Do you have any tips on prompting or anything else? Even when I put "large amount" in my prompt my model doesn't seem to listen. By the way I am using GGUF 14b 480p Q4 model. Again, thanks for your hard work!
@jaykob Have you tried using my LORA alone, or just with Wan Thiccum or Wan Cumshot? How many LORAs did you use simultaneously? If you only used one of my LORAs and this issue occurred, then it would be a problem with my LORA. Also, were your prompts based on mine, or did you write them yourself? Which LORAs did you use simultaneously? Could you please tell me?
@FLOW0308 I have only really used multiple Loras I haven’t tried using yours alone. Currently I use 4 Loras: general nsfw, wan cumshot, your lora, and masturbation cumshot. My prompts are general based on yours and other people’s posts. I just try to write “a large amount” wherever possible. I generate videos using tea cache, skip layer guidance, and modelsamplingSD3 (shift 3) at 25 steps 5.0 cfg.
@jaykob Based on my past experience, using general nsfw alongside anything related to cum/ejaculation tends to either reduce the amount of cum or cause it to split and disperse when ejaculated. My early works also used general nsfw in combination, but the cum amount decreased, and the effect was very poor. While general nsfw works exceptionally well for everything else I've used it with, it consistently yields very bad results when applied to ejaculation.
This is just my experience. You might try not using general nsfw for your next attempts. You can certainly try combining it with three other additional elements, but for now, see if excluding general nsfw makes a difference.
@jaykob general nsfw works exceptionally well for various sexual positions, poses, or even oral sex – it's quite versatile in those applications. However, based on my extensive experience, its effect is consistently very poor specifically when it comes to ejaculation. My guess is that it either completely omits ejaculation or creates some kind of conflict with cum itself; I'm not sure of the exact reason why.
So, whenever the scene involves ejaculation, I simply avoid using general nsfw. For everything else, though, general nsfw is incredibly versatile and useful. Just give it a try and see for yourself.
@FLOW0308 It seems to work although the difference is very small. Thanks though I think it most likely resolved some inconsistencies.
Thank you for your hard work. Very nice Lora with good results. Any chance you might create a cumshot on breasts Lora someday? I'm too stupid to train a Lora myself. 😇
I will try to train it, that's for sure. The main thing is, I don't have training material specifically for cumshots only on the breasts. So, figuring out how to create material that exclusively shows cumshots on the breasts is my only challenge.
@FLOW0308 I've send you a message.
@FLOW0308 I imagine training a lora for i2v 480p wan is difficult. Is there any noob friendly step by step guides on how to do so? I'd maybe need to use a virtual service as well as my hardware probably isn't good enough.
@anonameguyman1234252 Honestly, I've heard some people say the footage can randomly darken, which tells me there's still an issue with my current training method. So, I won't be releasing any tutorials until I can definitively solve this problem. It's frustrating because the 3-second videos work fine, but some say that as soon as you get to around 5-second videos, there's a random dimming of the screen. This is a huge setback for me because I have no idea what the problem is. My very first version also had this screen darkening issue, meaning this problem has existed from the beginning. This indicates I need to consider slightly modifying my training method until I can pinpoint the problem.
Also, I'm just an amateur AI enthusiast who only speaks Chinese and has very limited English skills. My LoRA training ability still has its flaws. So, when I communicate with everyone, I rely on translation tools. Please bear with me if my English isn't perfect.
另外，我只是一個業餘的 AI 愛好者，只會說中文，英文能力非常有限。我在訓練 LoRA 的能力上還有不足之處。因此，我跟大家交流時，都是透過翻譯工具來溝通的。如果我的英文不夠好，還請大家多多包涵。
FLOW can you please share article or quick video on how is your lora creation process? As it is targeting directly Wan, more people will able to work on that, and contribute.
-
FLOW，您好，能分享一下您的 Lora 创作过程的文章或视频吗？由于它直接面向广域网，所以会有更多人参与其中并做出贡献。
@spittreyways 如果你懂中文的話我用中文回你比較快
老實說我是跟T8跟湯糰豬的影片學的 我是參照他們的影片學的
而且我自己還是找不到有關 影片速度變慢的問題 我不確定是不是整合包問題 還是我訓練素材問題 我分不太出來 我學習訓練LORA也10天左右吧 這是我的第一個LORA 所以我還是很菜的新手 我還在摸索 我不太敢亂教人 因為我不確定我學的是否正確 而且感覺我的LORA 問題還是很多
I'll occasionally pin some of the posts people make, and then just as casually unpin them. Please don't pay any mind to my actions; I'm just doing it on a whim.
Love this updated version <3
Thank you^^
Very, very good description. Thank you very much! A friendly question: Is it really enough to train video with some short gifs or is it mp4? I dont find a "how-To-make" here. Where can i find? Thank you very much for an answer.
As a newcomer, my first expressionless version was trained using 15 self-made videos. I managed to train it successfully, and the results for videos under 3 seconds were quite good. However, the screen sometimes lagged, and any video over 3 seconds definitely darkened.
So, I trained a second version with 33 self-made videos. I thought I'd finally solved the screen darkening issue, and I believed that combining it with "WAN CUMSHOT" could resolve my LORA problems. Yet, I'm still getting reports that the screen darkens. At this point, I have no idea what I'm doing wrong.
I'm also really unsure if my training method and tagging process are correct, since I've been experimenting entirely on my own. I've only been following online video tutorials. I only know how to use Musubi Tuner, specifically the all-in-one package version from China. I'm still trying to find out what the correct training method actually is.
Hi. Do I generate an i2v 480p video first, then use your upscale/smoothing workflow and drag the 480p video there?
YES
Love this LORA. I've found that it makes a weird angle for the shot rather than adding height, for an arcing shot, good place to improve. I could show you examples. It basically comes out of the p3nis at a strange direction.
I've recently tried to incorporate more training material featuring long-distance ejaculations with an arc, hoping to improve the LORA's performance in this regard.
However, I've found this task to be quite challenging, mainly for a few reasons:
Firstly, the angle of the images used by different individuals varies significantly. More critically, the ejaculation angle is heavily dependent on the angle at which the p3nis itself is pointed towards the face. If the p3nis isn't directly facing the subject, my LORA currently forces the ejaculated fluid onto the face, which is a problem I'm unable to resolve at the moment.
Since there's such a wide variety of angles in users' generated images, I can only try my best to include various perspectives. But this also needs to be coordinated with facial expressions, making it a very substantial amount of work. Furthermore, even with all this effort, I can't guarantee that the model will generalize well across all these angles.
Nevertheless, in the next version, I will do my best to add more materials with different angles and distances to try and improve this issue. If it still can't be effectively improved, I truly might be at a loss, as creating just one satisfactory piece of material often requires multiple attempts.
Moreover, when creating my own training data, I personally tend to use angles that I consider normal and natural. Therefore, when I see user feedback presenting various unexpected angles, it makes me realize, "Oh, so there are angles like that too!" Unfortunately, for every new angle I need to incorporate, I might have to create approximately 5 additional pieces of material for the model to generalize successfully.
@FLOW0308 Well, good to hear you're working on it. And in the meantime, it's still a great LORA
I've trained numerous versions recently, but the results haven't been very satisfying. The main issues have been either an excessive amount of liquid, or eyelashes that frequently feel distracting and negatively impact the overall aesthetic.
While a recent change in my training method has made the workflow much smoother, I've now run into a significant bottleneck with the source material itself.
I don't mind an excessive amount of liquid if you know what I mean 😏. But yes, I've also noticed that whenever a character's face has a lot of blush, it tends carry over that blush to the eyelids when the eyes close, making it seem like my characters has just a ton of makeup all over their face. Maybe include material that has no blush? Just a thought I don't have any experience in Lora training
jaykob so good to be me - I got every time an escessive amount of liquid, no matter what weights or lora I use. :(
forfreelsd368 Interesting. Care to share your prompt? And are you using the v1 or v1.1? Wan 2.1 or 2.2?
jaykob nah, forget about it, now it's no liquid at all with all the prompts I try, especially from gallery here and other loras.
Every works literally except cum.
Wan2.2, ver1.1. But I tried loras wor wan2.2 of this plan, also those which compatible from wan2.1, as they wrote in comments. Even workflows, based on cum. Not even close results, as on pictures.
Hi there! This may be a stupid question, but is there a good recommended workflow on how to get to this to work well?
I'm still stupidly new to ComfyUI and its stuff, and I see that there's a picture of a workflow in the description here, but it's a bit too low quality for me to see and learn from it :(
I really wanna try this out more, but I feel very lost and the workflow templates I did find don't help much either with getting it right just yet.
Since the WAN 2.2 version has already been released, I might consider switching my training to it in the future. However, it will likely take some time for the WAN 2.2 ecosystem to fully mature, including updates to the trainers.
So, for now, I'll continue training with WAN 2.1. I've already trained one version, but I've noticed occasional pupil color changes, which I'm still trying to fix. I'll temporarily stick with WAN 2.1 for LORA training.
I'm also unsure when WAN 2.2 will be widely accessible, as its high VRAM usage might prevent most users from adopting it. I'll only transition to training with WAN 2.2 once it has matured.
Nice Lora ! I've tried WAN2.2 here. Surprisingly it takes a little bit less VRAM (FP8 version) than 2.1 FP8 in my setup as it is split in two model (low noise and high noise), at maximum my workflow consumes 25GB VRAM in comfyUI. And it's faster (on RTX5090) for 49 frames 720p (720x1280) is 10 min for 2.1 an 7 min for 2.2 I would say. The success rate of generation is way better with 2.2, it's more reliable, maybe it would fix your eye color problem? (on 20 gen with 2.1 it was a 50% rate succes and with 2.2 it is 80% for NSFW, but I do not have a lot of Lora enabled as there only a few compatible). Have a nice day and thank for sharing your lora !
Mortyh I've noticed that too, but I'm not sure if there's an error in my settings. I can still load LoRAs, and they do have an effect, but the console displays messages like "LORA KEY NOT LOADED". Despite that, the effect is still there, though I feel the strength is only about 50% of the original.
Are you able to use them perfectly/fully? I've completely switched to WAN 2.2 workflows, trying both the official and KJ versions. There is an effect, but the console still shows LoRA not loaded, yet the effect persists. However, I feel the naturalness of the fluid output/volume isn't as good as with 2.1. But on the positive side, the overall actions and emotions are stronger/better. So, the fluid generation/output isn't as good as before, but the actions and reactions are much better. Also, the eye problem is gone now! Perhaps 2.2 fixed my LoRA issue.
I'm still unable to figure out if it's fully compatible or universally usable. Are you currently able to use it completely without any problems?
FLOW0308 I have exactly the same console messages (LORA KEY NOT LOADED), wich I do not fully understand except the numer of layer is maybe different between 2.1 and 2.2 as it is between WAN 2.1 14B and WAN 2.1 1.3B. So it can't load all the lora layer or not the full match of it. Wich maybe results in a strenght loss of the lora on WAN2.2 .
I noticed the effect of movements are less marked (from other lora) but still there, as you said it looks like a 50% strength. I've not fully tested your lora with the fluid yet, I'm doing right now. I'll tell you if a see a difference in fluid quantity/output between 2.1 and 2.2.
Mortyh Thank you for your answer and for helping me test.
Hey there, I've been using this LORA for a couple weeks now and it's been absolutely amazing. I want your advice on this one problem I just cannot seem to fix. The shape of the penis (especially the tip) is always a bit deformed. It also seems to have trouble understanding what the back and side of the penis looks like so it just morphs it at the beginning. I tried various combinations and confirmed that this is the lora that's causing the issue. Is this because of the lack of material? I am using image to video, wan 2.1 GGUF, this lora, kijair's self-forcing lora, and cumshot lora. Any help would be appreciated, thanks!
From my own experience and observing other users' samples, there generally haven't been any issues.
However, for scenarios where a man enters the frame from off-screen before ejaculation, you'll need to manually add a separate penis LoRA for enhancement. I haven't specifically trained on such source material; I've only included a few basic elements to help with general composition.
Given that the WAN CUMSHOT LoRA is already available, if the penis is already present in your initial image, you should generally face no problems. At least from my own testing and reviewing contributions from others, I haven't encountered any issues. If the penis is already in your input image, I'm unsure what LoRA weight you're applying or if it might be affecting the outcome.
FLOW0308 Currently using every lora at 1, except cumshot at 0.5. I've tried experimenting with weights, but nothing seems to work. Do you mind if I send you a couple of frames of the first second of my video to see if you can figure out what's happening?
fulove1010 You could try using WAN 2.2 instead; that might solve your issue. Alternatively, search online for a "PENIS LORA" and add that in—it should help you out. Otherwise, feel free to upload your work to your files, and I'll take a look when I have a moment. I'm currently out and about, so I can't look into the issue for you right now.
FLOW0308 Got it. I think I will wait for you to finish training your lora for wan 2.2. I'm having trouble with vram, but hopefully I can figure that out soon. Thanks for your help.
I've updated and pinned a few new videos for everyone to check out. The fluidity offered by WAN 2.2 this time is truly astonishing, and I've further enhanced its smoothness by integrating my own upscaling, smoothing, and frame interpolation workflow.
These videos were generated using the latest version of WAN 2.2, alongside a new LoRA version I've trained (which isn't publicly released yet as it's still undergoing internal testing). My latest training focused on a wider variety of dynamic actions. Emotional expressions were achieved by combining my existing trained data with WAN 2.2's robust computational power, specifically in conjunction with WANCUMSHOT. These samples were created using a total of four LoRAs: my newly trained LoRA, WANCUMSHOT, and two others (LIGHTX2V and PUSAV1) for acceleration and enhanced motion.
I'm still exploring and haven't yet tested if WANCUMSHOT is strictly necessary for optimal results, as WAN 2.2 is proving to be very powerful on its own. While I'm still fine-tuning the optimal weights and ratios for various elements, the current settings I've used for these samples feel the smoothest I've achieved so far.
I'd greatly appreciate your feedback, especially on whether the quantity and color of the cumshot in this latest training version are acceptable. Honestly, it turned out quite different from my original expectations, and I didn't specifically make changes for this during training, which I find rather strange. I've also run tests on real-person scenarios, and I personally find those results to be even better. Any feedback you can offer would be highly valuable. Thank you.
😄 really looking forward to that new lora. any chance the 2.1 version works with wan2.2? or will it give strange results?
jridge1993 All the videos I currently have pinned in the post were generated using my previously trained 2.1 LoRA, which was refined with my latest training methods. I haven't officially released it yet because I'm still troubleshooting an issue where the cumshot sometimes appears off-center in the middle of the frame; I haven't figured out the cause.
However, I can confirm that this 2.1 LoRA works perfectly with WAN 2.2, and it effectively resolves all previous issues. For example, problems like the screen getting darker, eye color changes, and eyelash distortions are significantly less severe than with WAN 2.1. I've attempted to generate at least 30 five-second videos so far, and I haven't encountered any issues such as the screen dimming, eyes changing color, or sudden addition of unexpected makeup to the face or eyes. At least, I personally haven't experienced them.
Please note that the workflow is currently a bit complex, and I'm still experimenting with various setting values as I continue to explore WAN 2.2's capabilities. Nevertheless, I wanted to share some initial results so you can review the current samples and provide feedback on their acceptability.
FLOW0308 Thanks so much for your hard work. I posted my own experiment with WAN2.2 + your 2.1 LoRA: https://civitai.com/images/91666896
To compare to a video made using WAN2.1: https://civitai.com/images/91094540
Good:
as you said, it works really well after some tuning. A lot of the previous issues seemed to magically disappear. Eyelashes look good, and there's no longer weird penis distortions at the beginning of the clip. WAN2.2 also gives much more natural movement/expression to anime characters out of the box which is great.
Bad:
The amount of cum seems much smaller (boo 😭) than with WAN2.1, and the consistency is pretty different. It is much more solid-white and doesn't "run" as smoothly as it did before. It kinda looks like elmer's glue 😜. It is easily worth it for the smoother and more dynamic motion that WAN2.2 gives. It also seems like some of my other favorite LoRAs don't mix as well with this one as they used to (namely https://civitai.com/models/1606766/pov-handjob?modelVersionId=1871755 and https://civitai.com/models/1796502/better-anime-sex-faces-wan-14b-i2v?modelVersionId=2033053) so I had to disable them and use only the 4 you listed.
jridge1993 You could try setting the high noise strength to 2. Also, in the prompt section, you can try to write the details a bit more specifically. Personally, I feel that this WAN 2.2 can help you fill in a lot of details.
Honestly, regarding the semen color, I haven't specifically changed it at all, so I don't know why it's become whiter. I'll see if I can try to replace the source materials/data even more. I actually prefer my previous semen color; the current version is too white for me, which is strange since I definitely didn't specifically change the color. Perhaps the assets need further adjustment.
As you saw, my setup just uses WAN CUMSHOT + my own trained LORA + an acceleration LORA, and that's it. With that simple setup, I can achieve the results and effects you saw in my video – in terms of speed, fluidity, and fluid volume.
jridge1993 I'm currently unsure about the maximum number of LoRAs that can be loaded, as LIGHTX2V is already essential for WAN 2.2. Beyond the high-noise and low-noise LoRAs, I'm currently limiting myself to four. I haven't tested if more yield better results, as I find WAN 2.2 to be very strong already in terms of facial expressions, actions, camera work, and lighting. Therefore, I personally believe many LoRAs are redundant. If you do add them, just set their strength to 2.
My workflow is unique: besides the high-noise and low-noise samplers, I've also integrated an intermediate WAN 2.1 sampler. This allows me to better reproduce the full characteristics of my original 2.1 setup, meaning I'm using three samplers this time. It's an experimental setup, but yields decent results. This might also be the source of the color issue, which I'm still troubleshooting. Honestly, I'm not an expert in this field.
FLOW0308 By
"Set the high noise strength to 2"
you mean set LoRA strength for high noise model differently than the low noise? Currently I'm using the same LoRA setup for both. Let me try that and see the results. Your posts definitely seem to have more cum than mine 😄
I've also integrated an intermediate WAN 2.1 sampler.
that's a really neat setup, I'll see if I can achieve something similar. Based on the output it looks like it's getting the best of both worlds. Do you need to keep swapping the 2.2 and 2.1 models back and forth from VRAM? I imagine that takes a while.
jridge1993 This is a "three-stage sampler" workflow I've been experimenting with, primarily for generating content that requires intricate dynamic effects, especially when dealing with volumetric changes like fluids (e.g., ejaculation). Please note, these are my personal experimental settings and are not guaranteed to work in all scenarios.
My Setup Logic:
CFG is consistently set to 2, and the total sampling steps are 10.
Each stage consistently utilizes the "most basic acceleration LORA model: LIGHTX2V" along with a sampler. Any additional LORA model combinations are added based on individual needs.
High Noise Stage: Shaping Dynamic Actions
Range: Sampling steps 0∼3
LORA Strength: I set the LIGHTX2V LORA strength to 2. This value is based on my trials, but you can adjust it according to your specific generation results, as there's no absolute fixed number.
Function: I treat the "high noise" stage as a "dynamic action" feature. Its primary purpose is to sculpt the overall form and motion of objects, particularly for expressing large movements or deformations.
Application: In this stage, I utilize the LIGHTX2V LORA model in conjunction with the sampler.
Mid-Noise Stage: Core Function Restoration and Refinement
Range: Sampling steps 3∼7
LORA Strength: The LIGHTX2V LORA strength is typically set to 1. However, I occasionally adjust it to 0.5 or 0.7 based on specific needs.
Function: In this stage, I introduce an additional WAN2.1 model. I consider this stage crucial for "restoring the original LORA functionality," allowing for refinement and solidification of core details based on the initial form.
Application: Also combined with the LIGHTX2V LORA model and the sampler.
Low Noise Stage: Final Denoising and Fine-tuning
Range: Sampling steps 7∼ final step (10000 for official versions, −1 for KJ versions).
LORA Strength: The LIGHTX2V LORA strength usually mirrors the mid-noise stage, adjusted as needed. I view the "low noise" stage as a "fine-tuning version" of the WAN2.1 model.
Function: Its main role is to perform final denoising and precise detail adjustments.
Application: Combined with the LIGHTX2V LORA model and the sampler.
jridge1993 To ensure continuity and consistency in the generated output, all LORA models employed in the "High Noise Stage" are to be applied identically and consistently across both the "Mid-Noise Stage" and the "Low Noise Stage."
In simpler terms, whatever LORA models you select for the High Noise Stage (e.g., three specific LORA models), that exact same set of LORA models will be continuously utilized throughout the entire sampling process, across every single processing stage. The purpose of this approach is to maintain stability and coherence in the model's behavior across different denoising steps.
FLOW0308 thanks for the detailed writeup. I set up a flow based on your post and spent a lot of time tweaking LoRA strength and prompting. The results were significantly better than before 🥰 This is the best I could achieve thus far: https://civitai.com/images/91778587
Here's the LoRA strengths I used, but I'm sure they can still be improved on.:
Stage 1
* Anime cum 1.5
* LightX2V 2
* pusa_v1 0.7
* wan-thiccum 0.6
Stage 2
* Anime cum 1
* Light X2V 1
* pusa_v1 0.7
* wan-thiccum 0.3
Stage 3
* Anime cum 1.5
* LightX2V 2
* pusa_v1 0.7
* wan-thiccum 0.6
I'd also like to keep experimenting with different numbers of steps, cfg values, and possibly sampler methods. After testing with some live action images (not using your LoRA), I think the solid-white cum issue might be inherent to WAN2.2, as it took a lot of prompting and tweaking to get something that didn't resemble glue.
Pleaaaase make this for Wan 2.2 ? it needs a High and Low LoRa for that.
I've been using it as the high and low. Might not be ideal but it seems to work.
I have updated the WAN 2.2 version. Here is the URL. Please test it and see how it performs.
WAN 2.2 Anime Cumshot Aesthetics – Precision Load I2V (TEST) - V1 Low-Anime Cumshot | Wan Video LoRA | Civitai
I have updated the WAN 2.2 version. Here is the URL. Please test it and see how it performs.
WAN 2.2 Anime Cumshot Aesthetics – Precision Load I2V (TEST) - V1 Low-Anime Cumshot | Wan Video LoRA | Civitai
Anybody managed to use this in combination with vaginal sex? Like "he fucks her and then he pull s out his penis and cums in her face"`?
This always leads to the dick "glitching" out of the pussy or not going out at all...
This is a bit beyond my expertise, so you'll have to ask someone more experienced about this part. I haven't trained for this kind of result, so honestly, I'm not sure how I can help you.
However, I have seen someone succeed with WAN2.2, but I don't know how they mixed or used it either. This model was one of my early projects that I trained specifically for cumshots.
I hope you're successful in finding a solution.
Hey, thanks for the lora. I noticed a funny behavior, where it seems to be that the lora adds the handjob-motion if it detects any hands. I guess you added a ton of handjob data ? :D
Any solution for that?
yes.
handjob has been added automatically on my workflows too
Details
Files
Anime CumShot-Emotion.safetensors
Mirrors
Anime CumShot-Emotion.safetensors
Anime CumShot-Emotion.safetensors
Anime CumShot-Emotion.safetensors
Anime CumShot-Emotion.safetensors
acf-vid.safetensors
Cum - Emotion Edition - Anime Cumshot Aesthetics v1.1 Emotion.safetensors
Anime CumShot-Emotion.safetensors
Anime CumShot-Emotion-動漫射精美學 – 精準裝填（真人版相容）.safetensors