多模态｜Gemini 3 Flash 图片/视频分析

This is the "look at my content and tell me what's wrong with it" node. TikpanGemini3FlashPreviewAnalystNode wraps gemini-3-flash-preview as a specialized analyst: product/ad diagnosis, video shot-by-shot breakdown, prompt reverse-engineering, and compliance checks. It's the deeper cousin of the pack's generic Gemini 3.5 Flash node - fewer raw capabilities, but purpose-built presets and, crucially, a frame-extraction engine so you can feed it videos without paying to upload whole files.

How it works

Media goes in, a structured report comes out. The node accepts up to 12 images, image URL lists, a 视频帧_IMAGE input (video frames from a LoadVideo-style node - it auto-samples and compresses them), local video under ~18MB (inlined), or a public/YouTube URL. It then calls Gemini with your 分析任务 preset and returns 分析报告, 反推提示词 (reverse-engineered generation prompt), 结构化JSON, 用量, and 状态日志.

The inputs that actually matter

• API_密钥 - sk- from tikpan.com.
• 分析任务 - seven presets: 通用分析, 视频分镜拆解, 商品卖点分析, 广告素材诊断, 画面提示词反推, 安全与合规检查, 自定义. Pick the job, the node loads the template.
• 分析要求 - your focus areas on top of the preset (default asks for subject, scene, motion, lens, light, color, text, issues, and a reusable prompt).
• 抽帧策略 - how video becomes frames: 均匀覆盖, 按秒抽帧, 首尾加密, 运动变化优先, 混合智能 (default). For action analysis, 运动变化优先 is the one you want.
• 最大抽帧数 - up to 48 (default 24); more frames = more accurate, more expensive.
• 媒体解析度 - default / low / medium / high. Resolution is a token-cost lever; high detail on a long video is spendy.
• 视频输入策略 - 自动优先抽帧 (default, cheapest), 只用抽帧, 只用视频原件, or both (highest cost). Frames are the budget move.

Optional: 本地视频 (≤18MB inline; bigger → use URL or frames), 视频URL (accepts YouTube links), and the 高级自定义JSON payload deep-merge.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/htrert/ComfyUI-Tikpan-Pro

restart (or Manager → "Tikpan"). No models to download; this is relay + frame handling.

Where people get burned

• Frame strategy is a cost dial. 混合智能 at 48 frames on a long video is the expensive combo. Start with 均匀覆盖 at a modest frame count, then raise detail where it matters.
• Video inputs have size rules. Local video inline is capped around 18MB per the tooltip; past that it falls back to URL or pre-extracted frames. A 1080p 60fps file will blow the cap instantly - extract frames first.
• URL errors abort by default. URL错误处理 is 严格报错; switch to skip-bad-links for batch runs.
• Cache on. 复用本地缓存 defaults on and saves bandwidth/time on repeat analysis of the same file. Leave it.

Honest take: if you're doing "watch this ad and tell me why it converts / what to change," the preset templates plus the motion-aware frame strategy make this the pack's best content-diagnosis tool. It's a specialist - use it for analysis, not general reasoning, where the 3.5 Flash node is the better default.