<h1 id="all-examples-are-t2v.">All examples are t2v.</h1>This is not only i2v Lora. i2v or using the Ref model would be an "easy" job. The idea here is to teach the correct male anatomy for the base model.It should perform even better with i2v if the anatomy is already present, but hopefully with this lora it should not be needed anymore.<h3 id="v2:start-with-strength-0.9-works-from-0.4-1.5">V2: Start with strength 0.9 – Works from 0.4-1.5</h3>This was trained A LOT more steps at 2048 resolution, read the model details. I decided to change the model name because it doesn't make sense anymore. This should be a "fix" for the whole male anatomy.This might not be the final version, but I don't consider this an experimental version anymore.I like to use the <a target="_blank" rel="ugc" href="https://huggingface.co/fal/MiniMax-H3-Realism-People-LoRA">Realism people lora</a> from 0.2 - 0.6 to get rid of the plastic look the model has, specially at higher resolution. But the Realism LoRA can make weird anatomy so you will need to experiment.Limitations:<ul><li>It still cannot control cut/uncut and barely show foreskin.</li><li>not that many penis diversities. Glans are very well converged now, but the penises can all look the same. Size difference works.</li><li>frenulum is barely there.</li><li>depending on the pose the anus can still look misplaced, but changing strength and seed helps</li><li>motion: it was not trained on motion. If you want sex and cum perfect motion, even though this lora can do it as you can see in the examples, you better try to use it with another lora.</li><li>woman: untested. It might work but it was not trained on it. Since MM H3 is a really good prompt follower and learns pretty well.</li></ul><hr /><h3 id="v1:this-is-a-first-experimental-version.-trained-with-my-experimental-implementation-of-musubi-in-my-gui.-https:github.comdiodiogodmusubi-tuner_simple_gui">V1: This is a first experimental version. Trained with my experimental implementation of <a target="_blank" rel="ugc" href="https://github.com/diodiogod/musubi-tuner_simple_GUI">Musubi in my GUI</a>. <a target="_blank" rel="ugc" href="https://github.com/diodiogod/musubi-tuner_simple_GUI">https://github.com/diodiogod/musubi-tuner_simple_GUI</a></h3>After proper quality control technique (inspired by Ostris AiToolkit), it maintains quality very well. I intend to keep training it. It was trained at first with just still images at 256x. Why? Since Klein training first on 256 has been positively good. I'm doing a continuation training on 1024 for details now and maybe I'll try 2048. I trained 10 epochs, each is 1.629 steps. But I choose epoch 8 as it had the best asshole positioning. Penis shape is still shit. But it's getting there. If you use a i2v for first frame with a penis it should be great. No trigger word, just reference a man from behind or showing assSome supporting trigger words used on the dataset:<pre><code>glans, large penis, small penis, medium-size penis, wide open anus gap, closed tight anus gap, penis upside-down, scrotum, testicles, low-hanging testicles, hairy, shaved </code></pre>

Male Anatomy – Ass, Penis, Testicles and Glans  – H3

This is a long session of refinements with penis anatomy focus + some other concepts: oral, fuck, cum, penis tip urethra opening, ass touching, anal gap, special "sports". The idea was to refine penis but keep the concept of anus still there. The result IMO is a good male anatomy, finally. That is why I'm changing the name. This is not just ass anatomy and penis from behind anymore. It still sometimes misses, the opening and glans will sometimes come out wrong. The frenulum is still not there yet, and large foreskin neither. But this is miles better than anything we got from the past loras. There was a problem along the way. After so many epochs and so many steps (over 50k) the quality of the image started to suffer. I think this is mainly my dataset fault. I fixed it (hopefully) by doing block analysis and chopping some blocks. And as a consequence the lora is also smaller in size. If anyone wants to play with blocks and save their lora try my Realtime Lora fork here for ComfyUI: <a target="_blank" rel="ugc" href="https://github.com/shootthesound/comfyUI-Realtime-Lora">https://github.com/shootthesound/comfyUI-Realtime-Lora</a> For reference the blocks preset remerged was: 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 1.30, 1.30, 1.30, 1.00, 1.00, 1.00, 0.50, 0.50, 0.50, 0.50, 0.50, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 1.00, 1.00, 1.00, 1.00 Training lineage: Male_Ass-MinMax-v2.02Cumulative saved training: 42.807 steps · 33 effective epochs1. Male_Ass-1024-MinMax-v1.03epoch 8 · 13.032 local steps2. Male_Ass-1024-MinMax-v1.04epoch 2 · 3.258 local steps3. Male_Ass-2048-MinMax-v1.04-contepoch 2 · 3.258 local steps4. Male_Ass-MinMax-v2epoch 9 · 6.147 local steps5. Male_Ass-MinMax-v2.01epoch 7 · 9.982 local steps6. Male_Ass-MinMax-v2.02final saved checkpoint: epoch 5 · 7.130 local steps

Files with SHA256: e639c295694309801173a7e6818be970b5535411ac0821ad0ed11578beade276