Shake harder ref h3

Faster! Harder! Shake Harder! | H3 Motion Booster 🚀 Run It on RunningHub 🎁 New users on RunningHub International can get 1000 RH Coins after registration. Tested hardware: RTX 4070 Ti 12GB VRAM. 实测硬件：RTX 4070 Ti 12GB 显存。 A 10-second H3 video at around 1080P / 2.0MP takes approximately 15 minutes locally on 12GB VRAM. 10秒 H3 视频在约1080P / 2.0MP下，本地12GB显存实测生成时间约 15分钟。 The same 10-second test takes approximately 7 minutes on RunningHub. 同样的10秒测试，在 RunningHub 上大约 7分钟完成 V2 / Anime — Sound Tip This applies to both the live-action V2 and Anime versions. Put the the moaning in the action as spoken lines. That is what drives the mouth. Leave overall_soundscape for cloth, wet slaps, hip hits, and room tone.If you Write “she moans passionately” in the soundscape with no dialogue, and you get silent-film porn plus a haunted audio track. Example She snaps: "What the hell are you doing?! Get off me!" The next thrusts break the line: "Don't— wait— ah—" Then the lines turn into moans and keep changing: "Nnh..." "Ah— slow down—" "Hah... hah... there—" "Mm— wait, wait—" "Ahh... nnh... I can't—" "Ah-ah-ah—" "Fuck... right there..." To be completely transparent: rendering and training large video LoRAs like H3 on cloud GPUs is an absolute money burner 💸. High-tier cloud compute costs add up so quickly with every experiment and epoch iteration. Every single bit of support helps cover those heavy cloud GPU rental fees, and it directly serves as the motivation and fuel to keep this Bunny training and bringing more high-quality models to you all! 🐰⚡ NEW UPDATE — FASTER! ANIME Edition FASTER! ANIME is a newly trained 2D animation-specific version. It is not a variant of the old V1, and it is not simply FASTER! V2 with an anime style added on top. this version was trained entirely on anime material and captioned with animation-native motion language rather than realistic human physics. It focuses more on anime timing, large pose changes, sharp reversals, strong bounce, exaggerated follow-through, delayed secondary motion, expressive facial animation, and brief anime smear. Because of this, it generally keeps a much stronger 2D anime look even at higher LoRA weights. However, prompt wording matters a lot. If you describe the motion using too much realistic body physics, weight transfer, grounded force, or live-action movement language, the result can still drift toward a more 3D look. Don’t prompt it like live action. Prompt it like animation. SIMPLE PROMPT STRUCTURE Full speed from the first cycle. Large pose-to-pose travel with tightly spaced cycles. Compressed pose → short transition → open pose → snap into the next cycle. Strong bounce. Chest follows a beat late, hair follows after the head, secondary parts react last. Brief anime smear on the fastest reversals. Facial expressions also change with the motion instead of staying fixed. All example videos include their prompts. You can also paste this template into Grok, ChatGPT, or another AI, then tell it your character, pose, camera angle, and desired motion and let it rewrite the prompt for you. You guys will probably make much better videos with it than I do. :) RECOMMENDED USAGE • Best for I2V • T2V works, but usually needs more rerolls • • Prompt style has a major impact on the performance ceiling of this LoRA Ref2VA Edition A few users have reported that V0.2 already works very well with the FL2VA/Ref2VA Hybrid model, so this version is not intended to replace it. This is a native Ref2VA-trained variant. In my testing, it provides better reference character consistency, with some trade-offs in image quality and overall stability. Since it's already finished, I'm releasing it so you can try both and choose whichever works better for your workflow. Use at the start: is the adult woman. The woman from is etc. ⬇Alternate prompt structure for REFV2A⬇ Reference: is the adult woman for Subject 1 (S1) (from [Shot 1]). is the adult male for Subject 2 (S2) (from [Shot 1]). subject_definitions: Woman with a slender physique, big breasts, is naked Muscular male, large penis, is naked Environment: Bedroom with natural lighting coming in from a window integrated_multimodal_description: dynv2, [Shot 1] (S1) is etc. on top of (S2). - Please let me know your testing results. The two versions were trained with different training methods and captioning logic, so I'm very interested to see how they behave across different workflows, prompts, and reference images. This LoRA is provided as a motion-generation tool. Users are responsible for how they use their reference images and generated outputs. H3 Motion Booster V2 V2 is a major upgrade over the original Motion Booster — broader motion coverage, stronger impact response, more natural secondary body movement, and added audio support. What’s New - Expanded motion coverage – more movement patterns beyond simple forward-backward / vertical motion - Stronger impact & rebound – better compression, separation, recoil, and visible body bounce - Improved secondary motion – hips, thighs, torso and chest now react with natural delayed follow-through - Audio support – trained with audio for better contact sounds, breathing, and motion-linked effects - Better rhythm control – handles steady, fast, compact, and heavy impact-driven motion more reliably Usage Trigger: dynv2 Recommended weight: 0.6–0.8 (start at 0.7) Works best with I2V. Start from a correctly posed reference image, then describe motion direction, rhythm, contact, rebound, and audio. I2V performs significantly better than T2V (T2V needs more rerolls). When paired with NSFW ALLINONE, it can also handle fast oral motion well. Recommended prompt structure: integrated_multimodal_description: overall_soundscape: non_diegetic_music: I2V Example integrated_multimodal_description: dynv2. The two subjects from preserve the exact starting position, body alignment, and interaction shown in the reference image. His cock is