My first Minimax-H3 LoRA: Busty Women. Much like the earlier versions for <a rel="ugc" href="https://civarchive.com/models/1670104/busty-women-wan">Wan</a> and <a rel="ugc" href="https://civarchive.com/models/2736657/bustywomen-ltx">LTX</a> 2.3, this model will add cup sizes to the models in a text-to-video setting and will provide lots of good natural body motion in image- or ref-to-video scenes.<hr />Important:<span style="background-color:rgb(26, 27, 30);color:rgb(141, 141, 143);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:16px"> As always with my models, this LoRA does not represent any real person in any way. If this model produces a video with any resemblance at all to a real person, that is purely an accident of AI random generation.<hr />Recommended strength:<ul><li>text-to-video: 0.75</li><li>I2V or ref-to-video: 1.00 or 0.75, as needed</li></ul>Unlike the LTX version, H3 does well with nude bodies, so this LoRA includes topless scenes and motions. It should work well with or without clothing. It was tested with text-to-video, image-to-video (using the first-last-frame workflow) and reference to video using both still images and videos as reference. The sample scene above with the comparison montage of the topless woman walking in a shopping mall was created with ref-to-video using <a rel="ugc" href="https://civarchive.com/images/139368451">this scene</a>.Even though this model was not trained with any audio, I find that text-to-video at 1.00 strength sometimes gives odd audio results and the models rarely lip synch to match the speech. If you scale back the strength to 0.75 or so, it's much better at keeping audio and video in synch. Feel free to play with the strength setting to find what you like for your own scene.In image-to-video and ref-to-video, that tendency wasn't as strong and I was able to use the 1.00 strength more often, but just be aware you may need to adjust this in your workflows.<h3 id="a-note-about-workflows-and-comfyui">a note about workflows and comfyui</h3>Recent versions of ComfyUI have been making changes to how memory is managed with large models like H3. On my computer, it has gotten much worse and I am now likely to crash out of Comfy with an out-of-memory error in scenes that I ran only a week ago without any trouble. It's been frustrating finding a workflow setup that runs reliably just so I could make the sample videos. If you look at the embedded workflows from the samples, that's why you'll see some disabled nodes and some KJ debug nodes for clearing VRAM and so on. Sorry they're messy, I just did what I had to do to get the thing to run.My biggest frustration is that the Cache-DIT node no longer works as of the most recent Comfy update, and that was a huge speed-up that is now missing in my workflow. I'm hoping that'll get repaired soon.In the meantime, I hope you enjoy using this LoRA and don't forget to post your clips back to this model so I can enjoy them too.

Files with SHA256: ab8b1eecd0b9f932525541aca5972175f8901b68c4f418d110591f99b8046c0b