fix: 移除过度压缩,VLM 图片质量提升至近乎无损

- 压缩参数改为 2048px/Q92(肉眼无法区分)
- VLM 模型锁定 qwen3-vl-plus(中餐识别准确)
- 上传限制 20MB,支持 jpg/png/heic
- VLM 只负责识图返回结构化 JSON,饮食分析交 DeepSeek
This commit is contained in:
MingNian
2026-06-02 14:29:43 +08:00
parent d095832a10
commit b3ea5cc8ea
3 changed files with 1 additions and 1 deletions