I am deeply grateful to everyone who supports the model through their comments, word-of-mouth, downloads, and any other actions (it's really very kind of you, thank you so much). Thank you so much for your kindness, support, and positivity!
▒▒▒ model released DRAFT V4 Release date: 09/12/2025 ▒▒▒
Information: the model operates in I2V as well as T2V
Read before use and downloading
☣️ LIBIDINOUS INSIDIOUS DRAFT V4 - Wan 2.2 MoE Architecture
"It is not just a model. It is a complete neural restructuring."
📄 Introduction
LIBIDINOUS INSIDIOUS DRAFT V4 represents a technological paradigm shift within the Wan 2.2 ecosystem. Developed through a unique experimental process, this model does not merely generate video: it simulates motion physics and texture through a radically optimized Mixture of Experts (MoE) architecture.
This model is the result of a close collaboration between a Human Architect and the Gemini 3.0 PRO artificial intelligence, acting as technical supervisor.
🧬 The Process: A Training Revolution
Unlike standard models, V4 has undergone a surgical, multi-phasic training protocol. We did not simply "teach" images to the model; we redefined how its neurons communicate.
Here are the technical pillars of this training:
Spectral Training (Spectral Balancing via FFT):
The model was trained to dissociate low frequencies (structure/motion) from high frequencies (texture/details). Using Fast Fourier Transform (FFT) analysis, we injected "High-Fidelity" detail density without ever compromising the video's structural coherence.Intelligent Temporal Locking (Optical Flow Learning):
Motion vector analysis was integrated at the core of the learning process. The model "understands" pixel directionality. If a motion inconsistency is detected, it is rejected during generation. Result: organic fluidity, eliminating jittery motions or unexplained reversals.MoE Router Optimization (Top-K Consistency):
The Wan 2.2 architecture relies on "Experts". We forced the training to maximize Router Sharpening. The model no longer hesitates. It selects the perfect expert for skin texture or fluid dynamics with surgical precision, eliminating ghosting or blurring effects.Neural Sanitization (Sparse Annealing):
A "selective annealing" process was applied to eliminate residual digital noise in the deep layers. This ensures a clean image free from saturation artifacts, even with complex prompts.
⚠️ Warning: This is not a classic "Plug & Play" model
LIBIDINOUS INSIDIOUS V4 is a competition engine. It is sensitive, powerful, and does not tolerate approximation. It was trained on a massive hybrid dataset (Video/Image) to grasp concepts that other models ignore.
1. Prompting (Crucial)
Forget your linear prompting habits. This model reacts to temporality and scenography.
The use of precise Positive Prompts and strict Negative Prompts is mandatory to channel its power.
Supported and Recommended Structures:
Temporal Format (Most Precise):
(at 0-2: description of the beginning)
(at 2-4: description of the action)Scenic Format:
[Scene: detailed description of the atmosphere and action]Hybrid Format:
**[Scene: close-up on face (at 0-5)]Natural Description (Supported but less directive):
A cinematic video of...
⛔ Interference Triggers (To Avoid):
As the model is based on a sensitive MoE architecture, avoid contradictory keywords within the same time segment (e.g., asking for "static" and "running" without separating the timecodes). Avoid endless "danbooru" tag lists without syntax; prefer structured natural language.
2. Required Technical Parameters
To exploit the full potential of the V4 training, please adhere to these settings:
Text Encoder: umt5_xxl_fp16.safetensors (Mandatory for semantic understanding).
Sampling: Usage of 3 chained KSamplers is recommended (Advanced Workflow) to decompose noise, structure, and detail.
CFG Settings (Guidance):
The model is highly obedient. Do not force the CFG.
Recommended: 2.5 / 1 / 1 (Depending on your sampling workflow).
Steps (Compute Steps):
"Flux-like" Mode (Fast): 4 to 6 Steps (if using adapted schedulers like Turbo/Lightning).
Classic Mode (Max Quality): 20 Steps.
💡 Architect's Note
"This model was designed to push the boundaries of open-source video generation. We implemented mathematical safeguards (Soft-Tanh Limiting) to prevent the color saturation often seen elsewhere. What you hold is a precision tool. Treat it with respect, and it will produce impossible visuals."
Architecture & Training: XT-404 Solo Dev & Gemini 3.0 PRO Core.
Version: V4 (Stable Production Release).
Thanks to the beta tester: Toto4767, WaifuSynthLabs, Feoh, Aragon4000, FrogenGT.
Remember to share your creations. It was XT-404
▒▒▒ model DRAFT V4 Sortie à ce jour le 09/12/2025 ▒▒▒
Information le model fonctionne en I2V comme T2V
A lire avant Utilisation et téléchargement
# ☣️ LIBIDINOUS INSIDIOUS DRAFT V4 - Wan 2.2 MoE Architecture
"Ce n'est pas juste un modèle. C'est une restructuration neurale complète."
📄 Introduction
LIBIDINOUS INSIDIOUS DRAFT V4 représente une rupture technologique dans l'écosystème Wan 2.2. Conçu via un processus expérimental unique, ce modèle ne se contente pas de générer de la vidéo : il simule une physique de mouvement et une texture via une architecture Mixture of Experts (MoE) radicalement optimisée.
Ce modèle est le fruit d'une collaboration étroite entre un Architecte Humain et l'intelligence artificielle Gemini 3.0 PRO, agissant en tant que superviseur technique.
🧬 Le Processus : Une Révolution de l'Entraînement
Contrairement aux modèles standards, la V4 a subi un protocole d'entraînement chirurgical et multi-phasique. Nous n'avons pas simplement "appris" des images au modèle, nous avons redéfini la manière dont ses neurones communiquent.
Voici les piliers techniques de cet entraînement :
1. Entraînement Spectral (Spectral Balancing via FFT) :
Le modèle a été entraîné à dissocier les basses fréquences (structure/mouvement) des hautes fréquences (texture/détails). Grâce à une analyse par Transformée de Fourier Rapide (FFT), nous avons injecté une densité de détails "High-Fidelity" sans jamais compromettre la cohérence structurelle de la vidéo.
2. Verrouillage Temporel Intelligent (Optical Flow Learning) :
Une analyse des vecteurs de mouvement a été intégrée au cœur de l'apprentissage. Le modèle "comprend" la direction des pixels. Si une incohérence de mouvement est détectée, elle est rejetée. Résultat : une fluidité organique, fini les mouvements qui tremblent ou s'inversent sans raison.
3. Optimisation des Routeurs MoE (Top-K Consistency) :
L'architecture Wan 2.2 repose sur des "Experts". Nous avons forcé l'entraînement pour maximiser la "Netteté Décisionnelle" (Router Sharpening). Le modèle ne doute plus. Il choisit l'expert parfait pour la texture de peau ou le mouvement de fluide avec une précision chirurgicale, évitant l'effet de "ghosting" ou de flou.
4. Sanitisation Neurale (Sparse Annealing) :
Un processus de "recuit sélectif" a été appliqué pour éliminer le bruit numérique résiduel dans les couches profondes. Cela garantit une image propre, sans artefacts de saturation, même avec des prompts complexes.
⚠️ Avertissement : Ce n'est pas un modèle "Plug & Play" classique
LIBIDINOUS INSIDIOUS V4 est un moteur de compétition. Il est sensible, puissant, et ne tolère pas l'approximation. Il a été entraîné sur un dataset hybride massif (Vidéo/Image) pour comprendre des concepts que d'autres modèles ignorent.
#### 1. Le Prompting (Crucial)
Oubliez vos habitudes de prompting linéaire. Ce modèle réagit à la temporalité et à la scénographie.
L'utilisation de Positive Prompts précis et de Negative Prompts stricts est obligatoire pour canaliser sa puissance.
Structures supportées et recommandées :
* Format Temporel (Le plus précis) :
(at 0-2: description du début)
(at 2-4: description de l'action)
* Format Scénique :
[Scène : description détaillée de l'ambiance et de l'action]
* Format Hybride :
**[Scène: gros plan sur le visage (at 0-5)]
* Description Naturelle (Supportée mais moins directive) :
Une vidéo cinématique de...
⛔ Ce qui cause des interférences (À éviter) :
Le modèle étant basé sur une architecture MoE sensible, évitez les mots-clés contradictoires dans le même segment temporel (ex: demander "statique" et "courir" sans séparer les timecodes). Évitez les listes de tags "danbooru" interminables sans syntaxe ; préférez le langage naturel structuré.
#### 2. Paramètres Techniques Requis
Pour exploiter le plein potentiel de l'entraînement V4, respectez ces réglages :
* Text Encoder : umt5_xxl_fp16.safetensors (Obligatoire pour la compréhension sémantique).
* Sampling : Utilisation de 3 KSamplers en chaîne recommandée (Workflow avancé) pour décomposer le bruit, la structure et le détail.
Réglages CFG (Guidance) :
* Le modèle est très obéissant. Ne forcez pas le CFG.
* Recommandé : 2.5 / 1 / 1 (Selon votre workflow de sampling).
Steps (Pas de calcul) :
* Mode "Flux-like" (Rapide) : 4 à 6 Steps (si vous utilisez des schedulers adaptés type Turbo/Lightning).
* Mode Classique (Qualité Max) : 20 Steps.
💡 Le Mot de l'Architecte
"Ce modèle a été conçu pour repousser les limites de la génération vidéo open-source. Nous avons implémenté des sécurités mathématiques (Soft-Tanh Limiting) pour éviter la saturation des couleurs souvent vue ailleurs. Ce que vous avez entre les mains est un outil de précision. Traitez-le avec respect, et il vous sortira des visuels impossibles."
Architecture & Entraînement : XT-404 Solo Dev & Gemini 3.0 PRO Core.
Version : V4 (Stable Production Release).
Remerciement au béta tester : Toto4767, WaifuSynthLabs, Feoh, Aragon4000, FrogenGT,
N'oubliez pas de partager vos créations. merci a vous c'était XT-404
Description
Custom Node Workflow Update
https://github.com/XT-404/XT-404_SKYNET
New Hyper Optimizer workflow for 8GB/12GB/16GB/24GB and higher graphics cards.
Note: Currently only works with Nvidia cards.
For GGUF, please read carefully: city96/ComfyUI-GGUF is required to use the new dedicated Safetensors & GGUF node.
FAQ
Comments (30)
hello frenchy guy, cool, et même un manuel en français. Merci. ton I2V est génial en suivi de prompt.
fyi: i2v workflow is named "xt404.exterminabeur.json"
"beur" is slang for arab in France.
exterminabeur = arab exterminator
wil the next versions be extarminajew or exterminablack ?
The new custom working process is very cool, although I struggled to install it using the tyka method) I did not see a particular difference in speed, but the process itself is very convenient.
Is there a GGUF version available
Was there any training done on males, particularly males with visible faces and full bodies? If so, what is the approximate ratio of female to male subjects in the training?
This seems to be the best model. The only issue I have is that it changes breasts to big ones. Is not possible to have different sizes of breasts.
I tried your t2v workflow. I'm not sure if im doing something wrong, but it doesn't seem to want to follow a simple prompt such as, "A man and woman are having doggystyle sex."
I got to say, my initial tests of this have produced amazing img2video results. Thank you so much for your hard work and the dedication you shown making this model for us all!
This checkpoint is absolutely amazing. Well done! 👏🏼
Just wanted to say incredible job. As someone below mentioned I find it a tiny bit "unruly" with motion sometimes but that's probably my prompting as well. Much happier with this rather than struggling with motion in other models. Tested I2V. Thank you!
I tried this model although my go-to diffusion model is one of the Enhanced Natural i2v versions. I have to say that this is at least as good, on initial testing. Maybe better - excellent fidelity to the input image. I'm using your workflow although I have no clue what all those samplers do but the results are what matters. Great job. Now I'll test the "uncensored" claim ;)
Am I correct to assume this is not a "general purpose" model? It seems to be only trained on female characters.
Why I'm asking is because, in cases where the starting image shows only a male characters face, and the prompt tells the camera to show more of him so to speak, it adds small female-looking breasts on him when the camera zooms out or pans down.
The face retains the male form. So does the arms and the the rest of the upper body, except for the chest area.
I can't get your workflow operating on comfyui. Comfyui doesn't recognise the RIFE node. And there's no easy-to-find single, obvious custom node for RIFE on huggingface. Help please
wow this is really good! and libidinousv2 is awesome. i think it’s one of the best base models for making videos with wan2.2 right now, like top 3 easy. because it's so good we should tell it what could be even better!
it works super well with girls. they look how you expect and stay looking that way. if you give it a picture of a girl laughing and ask her to look normal, it guesses her face really well!
for an uncensored model, i think it's cool to focus on things like making undressing look real, skin looking real, and girls having regular bodies with little flaws.
basically, if you give it a good picture and tell it what you want, it would be awesome if the girl could undress herself and turn around, and everything down there looks right like textures, colors, shapes, all of it! no loras!
First sampler produces the starting image? Is there a way to preview so we don't waste render time?
Where is the gguf version?
This is a really great model, I hope you keep up the good work!
I see the user frozenGT make Always the idem vidéo why why ???
Clicking download results in it trying to download a zip rather than safetensors or gguf. Seems like this is training data rather than the model?
Thank you for the model. The previous one was excellent, and I'm happy to test this version as well.
Regarding the release itself — it was a mistake to use a zip archive. Lora Manager doesn’t recognize it, which is not very convenient. The second issue is that metadata tags are unlikely to be loaded, so there’s a chance there won’t be any user examples on this page, or there will be far fewer than there could have been.
In any case, thank you for the release. I'm sure it took a lot of effort to make it happen.
Nah... it seems very noisy and lots of morphing... Loras make it even worse :(
And still the Pussy transitions into being a Penis xD Yeah... I try a tiny bit longer with some other Seeds and Pics, but it does not look promising for my usecase ^^, sadly.
After testing V3, I’m genuinely impressed.
The prompt adherence is on another level—extremely responsive, accurate, and able to follow detailed instructions far better than previous versions. The overall visual quality and consistency feel noticeably improved, making it one of the most obedient WAN-style models I’ve tried so far.
After some very limited testing, It absolutely turns to shit if you use 2 or more loras, but on it's own is brilliant! For undressing, simple motions, poses, it's perfect. The faces are very consistent!!!
It looks like your Model is incompatible with the WanVideoWrapper. It is working with the ClownsharkChainsampler but when I use the same lora and prompt with the Wrapper it creates horror.
Hey. awesome finetune. One question: Can I enchance skin textures? after i2v looks a little bit smooth/plastic
man, still having issues with lora's catching... reverse cowgirl for instance... gave the girl a pecker!
Well, after 50 tests and regardless of the configuration, with negative or positive prompts, there is still a problem with the penis appearing in the vagina. It's a shame because this model is the most accurate in terms of vagina quality; it's just incredible. Please fix this problem. It's impossible to zoom out + fingering. But I'm still impressed by the level of detail.
I'm trying the flow in here, https://github.com/XT-404/XT-404_SKYNET, the cyberdyne model hub node isn't looking at my diffuser models folder or my checkpoints folder...what folder is it looking for the models in?
Alors tu t'en ai sortit comment finalement tu as suivis certains de mes conseils ?
"LIBIDINOUS INSIDIOUS DRAFT V4 represents a technological paradigm shift within the Wan 2.2 ecosystem. Developed through a unique experimental process, this model does not merely generate video: it simulates motion physics and texture through a radically optimized Mixture of Experts (MoE) architecture"
.... So you're saying it does slightly more detailed nipples?
Looks like we don't have an active mirror for this file right now.
CivArchive is a community-maintained index — we catalog mirrors that volunteers upload to HuggingFace, torrents, and other public hosts. Looks like no one has uploaded a copy of this file yet.
Some files do get recovered over time through contributions. If you're looking for this one, feel free to ask in Discord, or help preserve it if you have a copy.