<h2 id="gynecology-chair!"><span style="color:rgb(121, 80, 242)">Gynecology Chair</span>!?</h2><p></p><h3 id="minimax-h3"><span style="color:rgb(64, 192, 87)">MiniMax H3</span></h3><p></p><p>I'm baaaaack! Apologies for the delay, I got incredibly obsessed with extending WebCap, the captioning/training repo I've been vibe coding. I added Inference for real testing suite, and then extended it to do casual generations and then extended it for a feature I call Storyboard, for which I built in LLM integration to give it a premise and have it write an entire story with scenes and then render those using ComfyUI. Things just ballooned from there.... LLM models, prompting guides and strategies, context, output, censorship, oh my.</p><p>What is this about again? <span style="color:rgb(64, 192, 87)">Gynecology chair!!!!</span> Works well, as usual pretty strict in terms of perspective and positioning, which is what I do on purpose to get real stability.<br />I have like 6 other concepts already trained, just need to test them and release them, they'll all be free from the getgo as a little thank you to those who've donated and supported me, didn't mean to leave you all hanging so long!</p><p>Now that WebCap is a nuclear beast of a training tool and more, I should have less trouble finding the perfect epochs and releasing more quickly, assuming I don't find something else to play with...</p><p></p><h3 id="wan2.2"><span style="color:rgb(253, 126, 20)">WAN2.2</span></h3><p></p><p>Yes indeed.</p><p></p><pre><code>A beautiful {Russian|French|Swedish|Swiss|African|Latina|Austrian|Dutch|English|American|Siberian|Chinese|Ukrainian|Thai|Czech|Japanese|Indian|Caucasian|Spanish|Portuguese} woman is sitting in a gynecology chair, ankles bound, wrists bound, legs in stirrups,

She has {very long|long|short} {blonde|brown|dirty-blonde|black|auburn} hair and is wearing a bikini,
{
A doctor is {putting a blindfold over her eyes|{sitting|standing} near her|securing her wrist to a support}||
}

indoor lighting, doctor's office, {high-angle|three-quarter|front} view</code></pre><p></p><h3 id="training-notes"><span style="color:rgb(250, 176, 5)">Training Notes</span></h3><p>WAN2.2 is a lot easier to train. I was really concerned about the time required to train two models and test two models but I am having lots of luck with shorter clips and the resulting models seem so much stabler than what I was used to with HunyuanVideo that epoch selection is taking me way less time.</p><p>I was exploring a single LORA file with a fellow user but it did come with slightly less quality, so two models for full quality it is!</p><p>My captioning style matches my wildcard prompts pretty closely in case you are curious.</p><p></p><p>A note about LTX2.3, as I've been getting requests for it. Everytime I read up on whether or not it's worth trying that out, I keep seeing that WAN2.2 is better, so... not planning on doing that anytime soon.</p><p></p><h3 id="disclaimer"><span style="color:rgb(253, 126, 20)">Disclaimer</span></h3><p>Ask your doctor!</p>

Files with SHA256: ba3b91384ddcef28ef7ef4167dfaf0336451204411eeab9e728079e632e615f7