[2026/10/11]Upddate for MiniMax H3 EROS v5 CheckPoint int8 with base model to train lora ver.I don't know why, but the Lora I trained with the offical checkpoint has very subtle results on EROS.更新基於MiniMax H3 EROS v5 CheckPoint int8 為基礎模型訓練的Lora版本不知道為什麼，我用官方checkpoint訓練的Lora在EROS 效果很微妙。The data on the real people is still being collected slowly....真人的資料集還在緩慢收集中....===========================================================Because testing MiniMax H3 shows that it often cannot clearly distinguish the positions of the anus and vagina, and does not understand anus-related NSFW actions. (PS: It frequently confuses the anus and vagina, but in most cases it mistakes the anus for the vagina.)Therefore I decided to train this LoRA myself. This is my first experimental LoRA.All image materials are AI-generated. A total of 266 static images were used.Generated by Stable Diffusion:<ul><li>Illustrious / Anima (about 90%)</li><li>Partial 3D images (about 5%)</li><li>Real photos (about 5%)</li></ul>I2VA-related results are quite good. Ref2VA and T2V have not been tried yet. Current tested weight range of approximately 0.9–1.2 works well.The main materials are:<ol><li>Anal fisting</li><li>Penis insertion (anal sex)</li><li>Anal beads inserted into the anus</li><li>Anal plug inserted into the anus</li><li>Anal stretching / gape (self-spreading or being spread by another person)</li><li>Anal creampie / semen flowing out, etc.</li></ol>All material images were generated through a self-developed WebUI that integrates Stable Diffusion. It automatically combines random character LoRAs + concept LoRAs + template prompts using Qwen3.8 Flash Next, then performs AI image review (only keeping images scored 80 or above), followed by final manual selection.Captioning was also done through another self-developed WebUI, where Qwen3.8 Flash Next automatically generated the captions, which were then manually reviewed and confirmed.Because various tests showed that it is very difficult to make the model correctly understand the positions, three trigger words were used to train the relative relationships.<ul><li>Female character is referred to as zxqmei</li><li>Anus is referred to as zxqanus</li><li>Vagina is referred to as zxqvagina</li></ul>===========================================================由於測試MiniMax H3似乎分不清楚肛門與陰道的位置，以及不理解肛門相關的NSFW行為。(PS:經常把肛門與陰道搞錯邊，但大多是把肛門當成陰道)。因此自己訓練這個lora。 這是我第一個嘗試性Lora。圖片素材全部也是AI生成。總共使用了266張圖片素材，全部使用靜態圖。由stable diffusion :illustrious / anima 動漫生成(90%)3D圖片生成(5%)真人照片(大概5%) 。I2VA相關的部分，效果不錯，Ref2VA與T2v沒有嘗試過。權重目前測試約0.9~1.2效果良好。主要的素材都是1.肛門拳交2.陰莖插入肛交3.肛門珠插入肛門4.肛塞插入肛門5.肛門擴張(自行撐開或他人撐開)6.肛門射精流出等照片。全部素材圖由自行開發WebUI整合SD，由qwen3.8 flash next 自行隨機人物lora+概念lora+樣板Prompt自動生成素材，先AI審圖評分80分以上後再最後人工篩選。而標註部分也是由另一個自行開發的Webui，由qwen3.8 flash next自行標記，再最後人工確認。因為各種測試發現很難讓模型知道位置，因此用3個提示詞訓練相對關係。女性角色 用 zxqmei 代稱肛門 用 zxqanus 代稱陰道 用 zxqvagina 代稱

AnusPussy

My original LoRA seems to work well with the official checkpoint, but its performance appears to be noticeably worse with the EROS v5 checkpoint.I'm not sure why this happens, so I decided to train a new LoRA specifically for MiniMax H3 EROS v5 INT8.The original EROS model also seems to have alignment issues when strong image guidance is not applied.我原本的Lora 貌似在官方的CheckPoint能有效，但是在EROS v5 CheckPoint似乎效果不好。 我不清楚原因，所以我所幸針對MiniMax H3 EROS v5 int8訓練一個。 原始的EROS似乎沒有圖片強引導還是會錯位問題。

Files with SHA256: 0238cc40a424dcf6a3bcb680e43837539fa3a4676e90e6a179314e8b7df2b6c7