簡報圖案生成器

Point PresentationGenerator at a floor plan and it returns a renovation presentation's worth of content: slide titles, Chinese copy for each page, and one English image-generation prompt per slide. It's the "brain" half of the pack's PowerPoint pipeline - its sibling PresentationMaker does the actual assembly later. This node is where the ideas come from.

What it is

It's a vision-language model bolted into the graph, the same family as the prompt enhancers the ComfyUI community has been wiring in for years - except instead of rewriting a prompt, its job is structured: read the floor plan, and return a JSON document that describes a complete interior-design proposal deck. The editorially interesting part is that the LLM isn't generating pixels; it's generating the content and the prompts that your sampler then turns into pixels.

How it works

The node sends your floor plan image plus a long, very specific instruction to Gemini (the default), Ollama, or OpenAI - whichever your LLMManager configured. The prompt demands a strict JSON array of slides:

• a title page with a one-line style description,
• a "提案風格概述" page carrying four topics - 動線規劃, 設計主軸, 色彩與材質, 空間機能 - each with a ~35-character Chinese summary,
• one entry per room the model can see in the plan (客廳, 主臥室, ...), each with a Chinese description and an English image prompt,
• and a closing page.

Every entry's image field is an English prompt capped around 80 tokens, written to be dropped straight into a diffusion model. The node strips off any ```json code fences, parses the response, and splits it into the two outputs: image_prompts (a list - one prompt per slide) and json_data (the full JSON, for the maker node).

The inputs that matter

• image - your floor plan. Feed it a high-res PNG of the actual plan; the README's advice that better plans get better results is not idle talk.
• design_style - 11 canned styles (現代簡約風 default, 工業風, 北歐風, 日式和風, ...). Pick 自定風格 and fill in custom_style if you want your own look.
• custom_scene - optional. Tell it the space is a restaurant or classroom when the plan doesn't make it obvious; leave blank and it guesses.
• llm_type - Gemini by default, and honestly the one to keep unless you have a strong reason to go local.

How it slots into a workflow

floor plan → PresentationGenerator → image_prompts → CLIPTextEncode → KSampler → renders
                                  ↘  json_data ──────────────────────→ PresentationMaker

Each prompt from image_prompts goes into a CLIPTextEncode node and gets rendered. Collect the resulting images into a batch and hand that batch plus json_data to PresentationMaker. One hard constraint: the number of images you produce must equal the number of slides in the JSON, or the maker refuses to build.

Getting it installed and running

The node ships in ComfyUI-DesignPack - install via ComfyUI Manager (search "ComfyUI-DesignPack") or git clone https://github.com/dseditor/ComfyUI-DesignPack into custom_nodes and pip install -r requirements.txt. No model downloads; the heavy lifting is a cloud LLM. Before this node will do anything, run LLMManager once with your Gemini key - if you skip that, PresentationGenerator returns error strings telling you no key is configured.

Where it's fragile

The failure mode is the LLM refusing to return clean JSON. The node strips code fences, but a model that wraps its answer in prose preamble will still blow up json.loads - and the result is an "錯誤: ..." string flowing into your outputs instead of prompts. Local 12B Ollama vision models are the flakiest at this; Gemini is the default for a reason. Also know what you're getting: the output is Chinese-only and hard-structured around an interior-renovation proposal, so this is a specialist tool, not a general "make me slides" node. If you're producing interior design pitches in Traditional Chinese, it's genuinely clever. If not, it's not aimed at you.