vision-skilllisted

国产大模型生态的视觉层 / Visual layer for Chinese LLM ecosystem. 为纯文本 AI 模型提供深度图像理解能力 / Deep visual understanding for text-only models. 通过硅基流动 Qwen/Qwen3.5-4B 视觉多模态模型桥接(可切换智谱 GLM-4.6V-Flash),支持场景图、空间推理、上下文分析、情感解读、因果推理链、追问查询 / Bridges text-only LLMs (GLM-5.2, DeepSeek, Qwen, GPT) with visual understanding via SiliconFlow Qwen/Qwen3.5-4B (or Zhipu GLM-4.6V-Flash): scene graphs, spatial reasoning, contextual analysis, emotional interpretation, causal reasoning chains, follow-up visual queries. 当用户提供图片(路径或 URL)并要求理解、分析、描述、识别、推理、提取文字、回答关于图片的问题时触发 / Activate when the user provides an image (file path or URL) and asks to understand, analyze, describe, recognize, reason about, or extract text from it. 兼容任何遵循 Agent Skills 开放标准的框架 / Works with any Agent framework following the Agent Skills open standard.
ZYQIHUI/vision-skill · ★ 0 · AI & Automation · score 64

Install: claude install-skill ZYQIHUI/vision-skill

# Vision Skill 为纯文本语言模型提供深度图像理解能力,默认通过硅基流动 Qwen/Qwen3.5-4B 桥接(可切智谱 GLM-4.6V-Flash)。 Provides deep image understanding for text-only language models via SiliconFlow Qwen/Qwen3.5-4B by default (Zhipu GLM-4.6V-Flash as backup). ## When to Use / 何时使用当以下任一情况发生时激活本 skill / Activate when: - 用户提供图片路径或 URL,询问其内容 / User provides an image path or URL and asks about its content - 用户问"发生了什么"、"为什么"、"这是什么意思" / User asks "what's happening", "why", "what does it mean" - 用户需要空间推理(左/右、上/下、近/远) / Spatial reasoning (left/right, above/below, near/far) - 用户需要上下文、情感或因果分析 / Contextual, emotional, or causal analysis - 用户对先前分析过的图片追问 / Follow-up questions about a previously analyzed image - 当前模型无法直接处理图像 / The current model cannot process images directly ## Execution / 执行 ### Phase 1: Deep Understanding / 深度理解运行理解脚本,提取结构化表示 / Run the understanding script to extract a structured representation: ```bash python scripts/vision.py understand --image "IMAGE_PATH_OR_URL" [--mode MODE] ``` Modes / 模式: - `fast` (default / 默认): 单次综合调用。一次返回场景图 + 描述 + 文字 + 推理的 JSON。快速,约 5-10 秒。 / Single comprehensive call. Returns scene graph + description + text + reasoning in one JSON. ~5-10s. - `deep`: 串行多次分析。分别提取场景图、空间布局、OCR、上下文、情感、推理链。更慢但更详细,约 30-60 秒。 / Sequential multi-call analysis. More detailed but slower. Extracts scene graph, spatial layout, OCR, context, emotion, reasoning chain separately. ~30-60s. 读取 JSON 输出,将其作��你的视觉上下文内化 / Read the JSON output and internalize it as your visual context. ### Phase 2: