心宝❤多模态LLM反推

This is the pack's "look at this picture and describe it" node. Feed it up to three images, pick which LLM channel and model should look, and it returns a Chinese description of the subject, background, and style - the reverse-captioning (反推) step that drives a lot of the author's workflows, where a product photo gets turned into the prompt text that regenerates a designed version of it.

It's an API node, like the Banana generators: no model runs locally. What's interesting is the choice - this is one of the few nodes in the pack that spans both channels. You can caption through the 心宝 Gemini channel, the old "Banana same-model" channel, or ModelScope's Qwen VLMs. Same node, same output, different bills.

The inputs that matter

• image, plus optional image_2 and image_3 - up to three images. The first is required; the extras let the model compare/merge content across shots.
• channel - 心宝❤新渠道, 香蕉同款渠道(旧渠道), or 魔搭社区 (ModelScope). Pick where the request goes. Each channel needs its own key.
• banana_api_key and modelscope_api_key - leave the one blank that doesn't apply to your channel; the pack's global key node can backfill them.
• system_prompt and user_prompt - both editable, with sensible Chinese defaults ("you are a professional image description assistant…"). This is where you'd tune what the caption emphasizes - style-heavy, product-spec-focused, whatever your workflow needs.
• model - eight choices spanning two ecosystems: gemini-3-pro-preview-c (较耗时, "slower"), gemini-3-flash-c, gpt-5.1-high-c, plus ModelScope's Qwen/Qwen3.5-397B-A17B, Qwen/Qwen3-VL-8B-Instruct, Qwen/Qwen3-VL-235B-A22B-Instruct, Qwen/Qwen3-VL-8B-Thinking, and moonshotai/Kimi-K2.5. The -c suffix marks the 心宝 channel's conversational models; the Qwen/... ones are ModelScope's. This is the classic speed-vs-cost-vs-quality trade: flash is fast and cheap, the 235B Qwen VLM is a bigger brain, and the Thinking variant reasons before answering (slower, often better).
• max_tokens (100–8192, default 8192), temperature (0.1–2, default 0.7), seed (-1 randomizes) - the standard generation knobs. Lower temperature for deterministic captions, higher for varied phrasing.
• 禁用SSL验证 - the pack's usual trust-network-only toggle.

The one output

description (STRING) - the caption. Wire it to a text preview, or straight into a prompt input on a generation node: caption → prompt is the whole loop this node enables.

Installing and gotchas

Same pack:

cd ComfyUI/custom_nodes
git clone https://github.com/98624017/comfyui-banana-li-linux

Restart ComfyUI. No local models, no downloads - the pack's requests handles everything.

Where people stumble: forgetting which channel needs which key. Pick 魔搭社区 and the ModelScope key is mandatory; pick the 心宝 channel and a Banana-family key is. The blank-key fallback in config.ini only covers one key per slot, so if you caption through ModelScope and generate through the Banana channel, set the modelscope_api_key on the node directly. And if captions come back too generic, the fix is in system_prompt/user_prompt, not in swapping to a bigger model - the default prompts are tuned for "subject, background, style," which is the right base but a thin brief for specialized work.