Upload 2839680/3205326/README.md with huggingface_hub

2839680/3205326/README.md

@@ -0,0 +1,66 @@
+---
+license: other
+tags:
+- concept
+---
+
+# H3 Footjob - v0 - 3205326
+
+**Model Type**: LORA
+
+**Base Model**: MiniMax H3
+
+**Trigger Words**: h3fj., the woman's feet move up and down. her toes surround his penis. the woman is performing a footjob.
+
+**Tags**: concept
+
+## Gallery
+
+<table>
+  <tr>
+    <td><video src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/4690c47f-e969-41ef-8e65-ad7ecbff79f1/original=true/138982497.mp4" width="200" controls muted autoplay loop></video></td>
+    <td><video src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/cb6c3f35-5a4e-4652-a947-45da1ac301f9/original=true/138980927.mp4" width="200" controls muted autoplay loop></video></td>
+  </tr>
+</table>
+
+## Description
+
+Same disclaimer as everyone else: This is version zero. My first attempt at training H3.
+
+This is for the FL2VA model.
+
+Strength at 1.0 seems best.
+
+I tested it mostly with I2V. It works with T2V but the cocks look pretty grisly. Thus, no T2V previews on this model card.
+
+Experimenting with H3 I found it will actually do a passable footjob with prompting alone on I2V, as long as feet and cock are already in the image.
+
+H3 has a very good understanding of feet to begin with so the lora can just steer that knowledge, as it should.
+
+The foot/toe position is pretty simple in V0- basically just toes gripping the cock on either side, and feet moving up and down. Not much variation in the lora at this point.
+
+You can however add some variation with prompting, which seems to steer it back toward vanilla H3 footjob performance but with better toe contact.
+
+Trained with DiffSynth-Studio. Not super pleased with it. I see Musubi just added support today but only for the BF16, which my 5090 likely cannot handle even with block swapping.
+
+This published version was trained with the NF4, using a combination of the DiffSynth full and NF4 H3 example scripts.
+
+Trained on only 5 clips at 512x512, with audio (just sloppy lube noises, didn't really seem to come through in the lora)
+
+This is only 1000 steps. Movement was looking a little overfit beyond that, probably because of the shit dataset.
+
+I have plenty of additional training data ready for next versions.
+
+Previews are low res for the same reason as everyone else.
+
+All in all this is a super promising first result. H3 looks like it's gonna be a fantastic model for the footboys. Definitely training more right now.
+
+---
+
+Author: [icelouse](https://civitai.com/user/icelouse)
+
+Model: [CivitAI Model Page](https://civitai.com/models/2839680?modelVersionId=3205326)
+
+Archive: [CivArchive Page](https://civarchive.com/models/2839680?modelVersionId=3205326)
+
+<!-- Version: 20260807_upload -->