by chengyansen-ai · Codex Skill · ★ 21
H3 Multimodal LoRA Training 面向 MiniMax H3 的多模态 LoRA 训练、审计与交付技能。它把视频、立体声音频、参考图像/视频/音频、真 Ref2VA 条件训练、低显存后端、检查点选择和推理验证组织为一条可执行、可追溯、失败即停止的十步证据链。 项目不是参数合集,也不把“LoRA 能加载”当成“参考条件参与过训练”。每项能力必须对应固定源码、解析后的配置、真实运行证据和推理 round-trip;缺少证据时会明确拒绝或降级。 独立社区项目,非 MiniMax、ModelScope、AI Toolkit 或 SimpleTuner 官方产品。模型、训练器、数据集和衍生权重分别受其自身许可证约束。 为什么需要这个项目 H3 是联合音视频模型,训练风险不止是显存不足:目标媒体与参考媒体可能混用,音频可能被静默丢弃,CFG 蒸馏能力可能漂移,profile 字段可能没有真正进入后端,最终检查点也可能弱于中间检查点。 本项目将这些问题变成可检查的工程契约: 、、 分开记录; manifest、授权台账、审计报告、训练 profile、后端配置和启动计划逐层哈希绑定; 未识别字段、后端不兼容值或重复配置来源直接失败,不做静默 fallback; 训练前重新检查后端 Git SHA 和关键源码文件哈希; 生成样片必须通过 ffprobe...
| Stars | 21 |
| Forks | 1 |
| Language | Python |
| Category | Codex Skill |
| License | MIT |
| Quality Score | 55.2133174534131/100 |
| Last Updated | 2026-08-26 |
| Created | 2026-08-24 |
| Platforms | codex, python |
| Est. Tokens | ~9k |
These tools work well together with h3-multimodal-lora-training for enhanced workflows:
Explore other popular codex skill tools:
h3-multimodal-lora-training is MiniMax H3 多模态 LoRA 训练、审计与交付技能:Ref2VA、联合音视频、低显存后端与十步证据链。. It is categorized as a Codex Skill with 21 GitHub stars.
h3-multimodal-lora-training is primarily written in Python. It covers topics such as agent-skill, ai-toolkit, audio-video.
You can find installation instructions and usage details in the h3-multimodal-lora-training GitHub repository at github.com/chengyansen-ai/h3-multimodal-lora-training. The project has 21 stars and 1 forks, indicating an active community.
h3-multimodal-lora-training is released under the MIT license, making it free to use and modify according to the license terms.