———————————————————————————2025.11.21<h3 id="this-is-a-summary-of-the-prompt-refinement-i-did-using-gemini-3">This is a summary of the prompt refinement I did using Gemini 3</h3>Introduction I’ve been heavily testing Wan2.2’s I2V (Image-to-Video) capabilities, specifically focusing on surreal anatomy, non-standard physiological structures, and complex physical interactions (e.g., body morphing, penetration of non-standard orifices, etc.).After testing multiple prompting styles—Literary/Narrative, Slang/Erotic, and Clinical/Action—I’ve summarized my findings on how to precisely control the AI's motion generation. If you are struggling to get the AI to animate weird or surreal concepts correctly, this guide is for you.1. The Comparison: Three Prompt Styles TestedWhen describing an interaction involving "non-realistic body parts," I tested three distinct approaches:Type A: Descriptive &amp; Narrative (Storytelling)Example: "The camera follows her passion..." or "She is experiencing..."Result: ❌ Poor. The model gets lost in the "atmosphere." The motion is often subtle or barely moves because the AI focuses on the mood rather than the physics. It struggles to identify the active subject.Type B: Explicit / Slang StyleExample: Using common erotic slang or emotional words.Result: ⚠️ Unstable. While it generates motion, slang words heavily influence the art style (often making it look rougher or cartoonish). Furthermore, the model’s understanding of physics via slang is less precise than anatomical terms.Type C: Direct Action (Clinical/Mechanical)Example: "Subject A thrusts into Subject B," "Reciprocating motion," "Insert and withdraw."Result: ✅ Best. Video models (like Wan2.2) operate on pixel displacement. They respond best to dry, mechanical instructions about where objects are and how they move relative to each other.2. Why Wan2.2 Prefers "Direct Action"Wan2.2 (and similar Diffusion Transformer video models) understands spatial relationships and temporal consistency, not literature.When you write "Passionate," the AI cannot quantify the vector of that emotion.When you write "Thrusting back and forth" or "Rhythmic pumping," you are giving the AI a clear Motion Vector.Crucially for Surreal Anatomy: If you are generating something biologically impossible (e.g., an orifice in a location it shouldn't be), the AI has a strong bias to "fix" it back to a normal human body. If your prompt is vague, the AI will heal the hole or morph the object into a normal limb. Direct Action prompts force the AI to respect the physics of your surreal concept.3. The "Golden Formula" for Wan2.2For surreal or complex NSFW interactions, stop writing stories. Use this structure:[Visual Definition] + [Direct Motion] + [Texture/Physics] + [Camera Work]Case Study (Surreal Anatomy):❌ Bad Prompt: "Use a toy to play with the strange hole on her chest." (Too vague, AI will likely fail).✅ Optimized Prompt:"Close-up shot, surreal anatomy, a penis thrusting in and out of a vaginal opening located on a woman's nipple, wet texture, detailed flesh, continuous rhythmic motion."Why this works:"Surreal Anatomy": Tells the AI, "Do not fix this; it is supposed to look weird.""Located on...": Defines precise spatial coordinates."Continuous rhythmic motion": This is a key trigger for video models to maintain loop consistency and prevent the video from freezing or flickering.ConclusionWhen using Wan2.2 for I2V, treat the AI like a physics engine, not a novelist.The more bizarre or surreal the scene, the more clinical and mechanical your description needs to be. Don't be subtle—explicitly describe how Object A interacts with Object B, the frequency of the movement, and the texture involved.———————————————————————————2025.11.17<h2 id="my-insights-on-using-the-wan2.2-i2v-model">⚡️ My Insights on Using the wan2.2 i2v Model</h2><h2 id="kijai">（Kijai）</h2>Here is my breakdown and practical tips on optimizing the wan2.2 i2v model for better results, especially regarding motion and LoRA usage.<h3 id="1.-gpu-overclocking">1. GPU Overclocking</h3>A quick but crucial note: It is strongly advised against overclocking your GPU when running this model. Stability is key for reliable video generation.<h3 id="2.-lora-management">2. LoRA Management</h3>A. Combining LoRAsWhen stacking multiple LoRAs, be mindful that certain combinations can cause structural conflicts (e.g., in body anatomy) in the generated image. This often leads to a significant loss of dynamic movement or unexpected facial changes.<ul><li>Best Practice: Start by testing one specialized LoRA at a time. Once you confirm the base effect, slowly combine it with others.</li><li>Effective Combination Example: I found that pairing <code>LoRA (Tentacle Play) + LoRA (2D Animation Effects)</code> yielded some fantastic results.</li></ul><a target="_blank" rel="ugc" href="https://civarchive.com/models/1814871/wriggling-motion-lora-wan-22-release-tentacles?modelVersionId=2161329">Wriggling motion LoRA (WAN 2.2 Release) [Tentacles!] - Wan 2.2 I2V low (v1) | Wan Video LoRA | Civitai</a><a target="_blank" rel="ugc" href="https://civarchive.com/models/1920897/wan22-2d-animation-effects-2d?modelVersionId=2174159">wan2.2 2D animation effects 2D动效 - v1.0 | Wan Video LoRA | Civitai</a>B. LoRA Strength (Weight)Using an overly high weight for certain LoRAs can overpower your regular text prompts, making them ineffective. I define this "high range" as 0.7 to 1.0.<ul><li>Tip: For powerful effect LoRAs like <code>LoRA (Foot Job)</code> or <code>LoRA (2D Animation Effects)</code>, slightly reducing the strength (e.g., to the 0.5-0.6 range) often produces a much better, more balanced outcome.</li><li>A Side Note: I did not use the high-noise acceleration LoRA, <code>Lightx2v</code>.</li></ul><a target="_blank" rel="ugc" href="https://civarchive.com/models/1861926/wan-22-i2v-footjob">Wan 2.2 I2V Footjob - v1.0_WAN22_HIGHNOISE | Wan Video LoRA | Civitai</a><h3 id="3.-prompting-for-superior-dynamics-(motion)">3. Prompting for Superior Dynamics (Motion)</h3>To achieve excellent and lively dynamic motion, I've had great success with a specific format for describing the camera movement:<ul><li>For General Fluid Motion:<ul><li>Prompt Segment: <code>(the camera follows the woman's movement and continuously sways left and right)</code></li></ul></li><li>To Enhance NSFW Dynamics:<ul><li>The Key: Describe the intense action immediately after the dynamic camera prompt.</li><li>Example: <code>(the camera follows the woman's movement and continuously sways left and right) + (the woman and man are having intense sexual intercourse)</code></li></ul></li><li>My Theory: I suspect the superior dynamics come from explicitly telling the model that the main subject for the camera's focus and movement is the woman's action.</li><li>Negative Prompt:</li></ul><pre><code> 色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，多余的脚趾，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走，马赛克，卡通崩坏，绘制错误，线条错乱，涂鸦感，CG感过强，塑料质感，AI感太强，动作僵硬，木偶般的动作，呆板的姿势，缺乏动态感，假笑，面部僵硬，表情呆滞，动态丢失，动画感缺失，低帧率感，动作撕裂，残影异常，身体卡顿，头身比例怪异，眼睛模糊，五官错位，重复的身体部位，肢体缺失，塌陷的皮肤，透视背景错误，光影不合理，场景违和，衣服凭空消失，不连贯的衣服，畸形嘴唇，突变嘴唇，画得不好的嘴唇，畸形嘴巴，突变嘴巴，画得不好的嘴巴，多余的牙齿，缺失的牙齿，不对称嘴唇，不自然的口红，口红涂抹模糊，不连贯的嘴唇，画得不好的脚，突变的脚，畸形的脚，深色嘴唇，深色口红，红色口红，红色嘴唇，畸形的脸，突变的脸，画得不好的脸，脸部变形，多张脸，脸型不一致，人物身份变化，脸部模糊，面部闪烁，畸变手部，突变手部，多余的手部，肢体穿模，畸变眼睛，突变眼睛，画得不好的眼睛，眼睛不一致，像素化, 锯齿, 粗糙边缘, 锯齿边缘, 伪影, 噪点, 模糊线条, 扭曲细节, 锐边, 线条抖动, 失焦, 低分辨率, 错误纹理，真实睫毛，现实睫毛</code></pre>———————————————————————————<a target="_blank" rel="ugc" href="https://civarchive.com/models/1920897/wan22-2d-animation-effects-2d?modelVersionId=2174159">wan2.2 2D animation effects 2D动效 - v1.0 | Wan Video LoRA | Civitai</a><span style="color:rgb(140, 233, 154);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:13px"> 👆This LoRA provides excellent control over facial details in 2D anime videos. It also effectively prevents the face swapping/drifting issue that arises when using too many LoRAs, and it can make NSFW effects smoother.———————————————————————————NSFW<s>（https://civarchive.com/models/1802070/wan-21-nsfw-clip-vision-h）</s>The model doesn’t work with WAN2.2. Thank you for the reminder.👆無駄無駄無駄無駄無駄無駄ァー！！！（https://huggingface.co/NSFW-API/NSFW-Wan-UMT5-XXL/tree/main?not-for-all-audiences=true）———————————————————————————Stomach Bulge LoRA V1.0 （Re-uploaded.）This version of the Stomach Bulge LoRA has been fused with the following models:<ul><li>Stomach Bulge Wan 2.1 T2V 14B LoRA (by bonetrousers)</li><li>Wan NSFW Posing Nude (by The_Cook)</li></ul>prompt：<code>st0m4chBulg3</code>（During my recent tests, I noticed a very pronounced nipple motion effect beneath the clothing in the generated videos. Could this possibly be due to the Reward LoRA?）Based on my extensive testing, this LoRA works better with images where the character’s lower abdomen is noticeably raised.———————————————————————————Penis LoRA V2.2 （The previous v2.1 version of this LoRA has been removed.）[v2.2 UPDATE] LoRA Re-MergeThis version features a complete re-fusion of the core LoRA. The new merge now incorporates the following models :<ul><li>Wan 2.1 - Erect Penis, Cock, Dick, Lora, I2V （by <span style="background-color:rgb(37, 38, 43);color:rgb(230, 148, 135);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:20px">aipinups69）</li><li>POV Insertion （by <span style="background-color:rgb(37, 38, 43);color:rgb(230, 148, 135);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:20px">LocalOptima）</li></ul>Unlike previous versions that suffered from p*nis folding or p*nis displacement, this version of the LoRA provides a much better understanding of p*nile structure, allowing for proper shrinking and erection representation.Tip: If you wish to generate a fully erect p*nis, you may want to try using the prompt: <code>penis becomes erect</code>.（The file name includes “v1,” but please don’t mind that — I just forgot to rename it when uploading.）———————————————————————————FluidCore LoRA V1.0The following LoRA files were merged to create this model:<ul><li>Wan NSFW Posing Nude (Base LoRA)<ul><li>Author: The_Cook</li></ul></li><li>creampie (nakadashi/cum inside) concept WAN2.1 I2V<ul><li>Author: shura490</li></ul></li><li>Breast Squeeze and Lactation (Milk spray, milking) WAN 2.1 i2v/t2v model<ul><li>Author: shura490</li></ul></li></ul>prompt：（She is riding on a penis inside her vagina at the bottom of the screen. The penis pulls away from the vagina and some white translucent cum drips out from vagina in a creampie action.）（<span style="background-color:rgb(44, 46, 51);color:rgb(250, 220, 167);font-family:ui-monospace, SFMono-Regular, Menlo, Monaco, Consolas, &quot;Liberation Mono&quot;, &quot;Courier New&quot;, monospace;font-size:12px">white cum drips out of vagina）（One person in the middle of the screen, Her hands cradle the breasts and she milk her breasts by squeezing her breast nipples. loads of jets of milk shoot from her nipples.）————————————————————————————(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 penis rapidly moves out/slides out from the woman's vagina. The man fully exposes his penis to the camera, and the woman exposes her vagina. The man's penis then rapidly inserts into the woman's vagina again. The penis gradually disappears from tip to base in front of the camera. The man's penis continuously and repetitively thrusts back and forth, from inside to outside, within the woman's vagina. A woman is straddling a penis, engaging in sexual intercourse. The woman is grinding on top of the penis. The woman is rubbing her pussy on the penis., she rides a penis, moving her body up and down as the penis slides in and out of her vagina，Viewing angle, nsfwsks,penis goes in and out of vagina，The penis rapidly thrusts in and out of the vagina.,————————————————————————————Reward LoRA V1.0The fused LoRA consists of the following components:Wan2.2 I2V Reward LoRAWan NSFW Posing Nude (Author: The_Cook)Wan 2.1 - Erect Penis, Cock, Dick, LoRA, I2V (Author: aipinups69)prompt：2d anime（The prompt '2d anime' helps stabilize the face consistency.）Setting the Reward LoRA strength to 1 or higher might be better.(This reward LoRA provides structures related to penises and nudity, so please use it in combination with other LoRAs.)——————————————————————————————*Most version history removed for cleaner documentation.*————————————————————————————*Extensive testing has revealed that the v2.1 LoRA performs poorly in certain scenarios. We strongly advise against downloading it for now. Once optimized, we will delete the v2.1 file and release a new, improved version. We will also consider deleting the Beta 2.0 version following the new release.*

wan2.2 NSFW Motion Enhancer

Fused_Triple_LN_Blend.safetensors

wan2.2 ???? Motion Enhancer

Reward_LoRA_low_v1.0

Stomach_Bulge_high_v1

Stomach_Bulge_low_v1

Penis_low_v2.2

Penis_high_v2.2

FluidCore_high_v1.0

FluidCore_low_v1.0

Reward_LoRA_high_v1.0

❌low_v2.1

❌high_v2.1

hunyuan_foley_Workflow

L_v3.0

H_v3.0

❔beta_low_v2.0

❔beta_high_v2.0

low_v1.1

high_v1.1

beta_v1.0

Files with SHA256: 84403f502180e8d64b93083ce037d1ab18a08ec4b82b83f39282f81885e0927b