Upload 2364658/3432352/README.md with huggingface_hub

2364658/3432352/README.md

@@ -32,6 +32,10 @@ tags:
     <td><video src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/54b61d9e-9396-4889-bcf7-ef8075039f0e/original=true/146959948.mp4" width="200" controls muted autoplay loop></video></td>
     <td><video src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/0d9ec297-ec6b-40ff-b6fb-fe12cdcbde2d/original=true/146959583.mp4" width="200" controls muted autoplay loop></video></td>
   </tr>
+  <tr>
+    <td><video src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/c8567109-b525-45f9-a712-0ff13f88037d/original=true/147070665.mp4" width="200" controls muted autoplay loop></video></td>
+    <td><video src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/724710bb-de07-443f-9545-8ecd7a55642e/original=true/147080316.mp4" width="200" controls muted autoplay loop></video></td>
+  </tr>
 </table>
 
 ## Description
@@ -40,8 +44,8 @@ tags:
 
 ### Feel free to mirror any of my loras to HuggingFace
 
-**H3 V2: 10Eros Max Beta6 Highly recommended! Updated to a new dataset with much more data and a higher resolution. Thrusting and squeeze should be better than V1. For sampler I use euler with beta and 20 steps on 10Eros Max Beta6 and use comfy kitchen attention. Sparse attention will lower the quality noticeably same with the turbo loras. Audio can get out of sync sometimes but is primarily a prompting issue when it happens. Prompts are provided in the example videos. Prompting is very important for H3 bad prompts will give bad results on H3.**
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
+**H3 V2: 10Eros Max Beta6 Highly recommended! Updated to a new dataset with much more data and a higher resolution. Thrusting and squeeze should be better than V1. For sampler I use euler with normal or beta and 20 steps on 10Eros Max Beta6 and use comfy kitchen attention. Normal works better than beta in my testing and gives smoother motion. Sparse attention will lower the quality noticeably same with the turbo loras. Audio can get out of sync sometimes but is primarily a prompting issue when it happens. Prompts are provided in the example videos. Prompting is very important for H3 bad prompts will give bad results on H3.**
+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
 
 H3 V1: H3 version works on both fl2va and ref2va and stacks well with Eros MAX: post-trained with GoonRL (I named it that because its funny also put a shitpost in the lora metadata) my RL pass built on DiffusionOPSD (on-policy self-distillation, arXiv 2608.24646), run in rounds on the model's own generations. The objective is motion completion and coherence; there is no aesthetic scorer, no learned preference model, and no human ratings anywhere in the loop. Not sure if I will release the RL method since its experimental and takes a fucking shitton of time and experimenting to get right and I really don't have the time to maintain a musubi tuner fork for it. I will post better gens with a higher resolution later they take forever to do over 0.4mp 5sec. Not trained on audio so audio may be mid.
 ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
@@ -111,4 +115,4 @@ Model: [CivitAI Model Page](https://civitai.com/models/2364658?modelVersionId=34
 
 Archive: [CivArchive Page](https://civarchive.com/models/2364658?modelVersionId=3432352)
 
-<!-- Version: 20261027_upload -->
+<!-- Version: 20261102_upload -->