<h1 id="v2-update-10112026">V2 Update 10/11/2026</h1>My intention with this project was always the same from the beginning. Teach H3 the proper way to handle a deepthroat/facefuck/throatpie with the correct anatomical animations AND most importantly, high quality oral sounds and believable wet throat foleysTHR0AT H3 AV V2 — Rebuilt for Better Sound, Detail &amp; RealismThe motivation behind V2 was simple: to create one of the best-sounding oral sex and deepthroat LoRAs available on the Civitai Red marketplace. V1 demonstrated what was possible, but I wanted to push the model further, particularly in audio realism, synchronization, anatomical detail, and consistency. I believe V2 delivers on that ambition. Rather than simply extending the original training, I revisited the training configuration to extract more quality from the source material and produce a more refined audiovisual experience.For V2, I increased the training resolution from approximately 0.25 MP to 0.37 MP, giving the model roughly 48% more spatial information to learn from. I trained against MiniMax H3's native INT8 ConvRot base, avoiding the additional quantization error associated with NF4 conversion, and used the Qwen3-VL-32B text encoder for caption conditioning. The training employed a rank-16, alpha-16 LoRA with approximately 74.5 million trainable parameters, targeted transformer-block training, Automagic3 adaptive optimization, and EMA weight averaging to refine the learned representations. The complete run covered 108 epochs, and 1,512 training steps, with checkpoints preserved throughout for comparison.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/47df72ce-15bc-490d-88da-4ba96717b5bc/original=true/47df72ce-15bc-490d-88da-4ba96717b5bc.jpeg" />In the end, Epoch 100 was the checkpoint that converged the best and is the one I selected to upload to the community. Ironically, it was also the checkpoint that performed the best in terms of loss close to .40 which doesn't always mean better, but in this case, it turned out to be true. Have fun with it and create lots of wonderful smut.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/e9ff1a0f-1596-4407-8b1a-702e89f8e220/original=true/e9ff1a0f-1596-4407-8b1a-702e89f8e220.jpeg" />The goal wasn't simply to make V2 stronger; it was to make it more convincing. I focused on preserving the character of the original recordings while improving the model's ability to reproduce convincing vocal textures, nuanced sounds, synchronized movement, and finer visual details. This represents a substantial technical evolution from V1, and based on my assessment, I believe we've achieved what we set out to accomplish: a serious contender for one of the most immersive audiovisual LoRAs in its category on Civitai Red. Thank you to everyone who downloaded and supported V1. Your enthusiasm was a major motivation for taking V2 further.<h1 id="h3-throat-time-minimax-h3-lora">H3 Throat Time — MiniMax H3 LoRA</h1>H3 Throat Time is a specialized audiovisual LoRA trained on a highly curated data set of oral sex video clips over a 700-step, 50-epoch training process, designed to enhance throat-focused adult scenes, motion, and visual realism.Recommended Strength: 1.0<h3 id="included-versions">Included Versions</h3>I'm releasing two checkpoints:<ul><li>Epoch 36: The most versatile and balanced version overall. Recommended as the starting point.</li><li>Epoch 48: Slightly more visceral and intense in its output. The differences are subtle and largely stylistic rather than dramatic.</li></ul><h3 id="trigger-word-and-prompting">Trigger Word &amp; Prompting</h3>Trigger word: <code>thr0at</code>This LoRA responds particularly well to detailed, descriptive prompting. More comprehensive explanations of the intended action, movement, pacing, and visual characteristics generally produce better results than short, generic prompts.For best results, use the trigger word alongside clear, natural-language descriptions of what you want the model to accomplish. The more precisely you communicate the intended scene and motion, the better the LoRA tends to interpret your instructions.So far testing has confirmed it handles the following scenarios quite well.<ul><li>Solo Wet Slobbery Blowjobs</li><li>Assisted Wet Slobbery Blowjobs (character guides another character to blow a character)</li><li>Alternating 2 person blowjob (prompt language must be strict to avoid them both trying to blow the person at the same time)</li><li>Fast Paced Visceral FaceFuck (with or without cum and slobber)</li><li>Coughing while deepthroating penis in mouth</li><li>Coughing while coming off penis (with or without semen)</li></ul>Note on Semen: Model does better when semen is in source picture. H3 tends to default to pancake batter consistency without a visual cue. You can also prompt it away by asking for semi transparent seminal fluid or cloudy transparent semen.<h3 id="audio-a-major-motivation-behind-this-lora">Audio — A Major Motivation Behind This LoRA</h3>One of my primary motivations for creating H3 Throat Time was to address the shortcomings of MiniMax H3's native audio generation, particularly its tendency to produce unnatural, crunchy Foley effects and unconvincing wet mouth sounds.I wanted to push beyond those characteristics and improve the overall audiovisual experience rather than simply training another motion-focused LoRA.Standard 20-step generation (recommended):The LoRA performs best without Turbo acceleration, using a conventional 20-step sampling process. This provides the strongest overall balance of animation fidelity, audio characteristics, and the stylistic qualities learned during training.Turbo generation (6–8 steps):Turbo is supported as an alternative for faster generation. In my testing, it generally preserves most of the animation fidelity, making it a practical option when rendering speed is a priority.However, expect some regression toward H3's baseline audio characteristics when using Turbo. The original crunchy Foley effects and less convincing wet mouth sounds may become more apparent, as Turbo can diminish some of the audio improvements learned by the LoRA.For Turbo workflows, I recommend 6–8 sampling steps.Bottom line: If audio quality is important to your generation, I strongly recommend running the LoRA without Turbo at the full 20 steps.<h3 id="anatomy-and-reference-guidance">Anatomy &amp; Reference Guidance</h3>The training dataset included footage featuring male anatomy. When male anatomy is fully visible and properly referenced in the source shot, the LoRA can generally handle it without additional anatomy assistance.However, despite being trained on this material, I do not recommend relying on this LoRA alone to generate male anatomy from scratch. For best results, use either a complete visual reference or a dedicated anatomy-trained LoRA.<h3 id="compatibility-and-testing">Compatibility &amp; Testing</h3>Currently validated:<ul><li>I2V (Image-to-Video)</li><li>FL2VA</li></ul>Not yet validated:<ul><li>T2V (Text-to-Video)</li><li>REF2VA</li></ul>T2V and REF2VA compatibility remain unknown. I make no guarantees regarding performance or functionality in those modes at this time.<h3 id="recommended-starting-point">Recommended Starting Point</h3>Start with Epoch 36 for versatility, then experiment with Epoch 48 if you prefer a slightly more visceral aesthetic. For the most faithful reproduction of the LoRA's learned motion and audio characteristics, use detailed prompting and the standard 20-step process.<h3 id="credits">Credits</h3>Credit to @fatberg_slim for several really great Wan videos that were used in the training set for this lora. The remaining data set was from sourced via Pornhub - LOL

H3 Throat Time

Version 2:My intention with this project was always the same from the beginning. Teach H3 the proper way to handle a deepthroat/facefuck/throatpie with the correct anatomical animations AND most importantly, high quality oral sounds and believable wet throat foleysTHR0AT H3 AV V2 — Rebuilt for Better Sound, Detail &amp; RealismThe motivation behind V2 was simple: to create one of the best-sounding oral sex and deepthroat LoRAs available on the Civitai Red marketplace. V1 demonstrated what was possible, but I wanted to push the model further, particularly in audio realism, synchronization, anatomical detail, and consistency. I believe V2 delivers on that ambition. Rather than simply extending the original training, I revisited the training configuration to extract more quality from the source material and produce a more refined audiovisual experience.For V2, I increased the training resolution from approximately 0.25 MP to 0.37 MP, giving the model roughly 48% more spatial information to learn from. I trained against MiniMax H3's native INT8 ConvRot base, avoiding the additional quantization error associated with NF4 conversion, and used the Qwen3-VL-32B text encoder for caption conditioning. The training employed a rank-16, alpha-16 LoRA with approximately 74.5 million trainable parameters, targeted transformer-block training, Automagic3 adaptive optimization, and EMA weight averaging to refine the learned representations. The complete run covered 108 epochs, and 1,512 training steps, with checkpoints preserved throughout for comparison.In the end, Epoch 100 was the checkpoint that converged the best and is the one I selected to upload to the community. Have fun with it and create lots of wonderful smut. The goal wasn't simply to make V2 stronger; it was to make it more convincing. I focused on preserving the character of the original recordings while improving the model's ability to reproduce convincing vocal textures, nuanced sounds, synchronized movement, and finer visual details. This represents a substantial technical evolution from V1, and based on my assessment, I believe we've achieved what we set out to accomplish: a serious contender for one of the most immersive audiovisual LoRAs in its category on Civitai Red. Thank you to everyone who downloaded and supported V1—your enthusiasm was a major motivation for taking V2 further.

Files with SHA256: af6a2069554297a18e4715d40ddbda23b5c22e9bb18296ce29b1ec38412efadf