文本扩写 (SillyDream)

"Expand this prompt" is a one-liner that sounds trivial and isn't. SillyDreamTextExpand (文本扩写, "text expand") is the second node in the qianchi7/ComfyUI-SillyDream-PromptReverse pack: you feed it a short description, it sends the text to an OpenAI-compatible chat API, and returns a longer, more detailed prompt you can actually run. Same gateway, same key, same philosophy as its image-reverse sibling - just text in, text out.

Same honesty as that node, too: "SillyDream" isn't a model. The node is a courier. Whatever chat model your api_url gateway exposes is what does the writing, and for this one it doesn't even need vision - a plain chat LLM handles it fine, so you can point it at a cheap small model and not feel the per-call cost.

Why reach for it? The blank-page problem, again, but text-side. You have a vague idea - "a rainy cyberpunk street at night" - and you want a structured prompt that names materials, lighting, environment, lens and style. If your checkpoint has an LLM encoder (the Z-Image / Flux Klein generation), a second LLM writing the instruction is a translation between two things that speak the same language. It's also the natural second stage after the pack's image node: reverse the image into a prompt, then expand that prompt, then generate.

How it works

Plaintext in, chat-completions out:

• Your text goes into a system_prompt + user message pair, posted to <api_url>/v1/chat/completions (the node appends the path for you - http://IP:3000 is enough).
• temperature, top_p, seed and max_tokens ride along. seed is only sent if it's above 0, so a 0 just means "let the gateway decide."
• The reply parser is the same forgiving one as the image node - it strips chat scaffolding and handles gateways that return content as arrays or from reasoning fields.

The one input worth reading before you touch anything else is the default system_prompt. It tells the model it's a prompt-polish expert, told to fill in material, light, environment, lens and style - but not to change the meaning or add subjects the original text never implied. That's the author pre-answering the classic prompt-enhancer failure: subject drift, the thing that makes an "enhancer" invent content you never asked for. Leave the default in; it's doing real work.

The inputs that matter

• text - your starting description, multiline. Empty input errors immediately, so there's no silent no-op.
• api_url / model_name / api_key - same trio as the image node: your gateway, the model name exactly as the gateway lists it, and a key (or blank if the gateway needs none).
• system_prompt - the quality lever. Edit it to control output language (there's no language dropdown here) or to target a specific model's prompt style.
• temperature / top_p / max_tokens - defaults are 0.7 / 1.0 / 2048. Lower temperature for more faithful rewrites; raise max_tokens if you want genuinely long output.

The single output is expanded_text, a STRING - same wiring as the image node: into a text encoder or any text-socket node.

Installing it

One pack, both nodes. ComfyUI Manager → Install via Git URL → https://github.com/qianchi7/ComfyUI-SillyDream-PromptReverse.git, then Restart (not Reload - nodes only scan at startup). Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/qianchi7/ComfyUI-SillyDream-PromptReverse.git
pip install requests Pillow numpy

requests, Pillow, numpy - that's all. No model downloads, no heavy dependencies. Search 文本扩写 or SillyDream after restarting.

Troubleshooting

Same gateway-shaped failures as its sibling, because it's the same client code:

• 404 - model_name doesn't match your gateway's configured name.
• 401/403 - key wrong or no access to that model.
• 429 - rate limited; the node retries, then gives up.
• "response isn't valid JSON" - api_url points at a web page, not the API.

Two things that trip people specifically on this node: if you point it at a hosted, censored model, don't expect it to expand anything spicy - that's the API path's filter, not a node bug. And if the output comes back in Chinese no matter what you do, it's the default system_prompt doing its job; swap in your own English one. There's also a bundled template (工作流 → 浏览模板 → Image-Prompt-Reverse) that shows the reverse-then-expand chain wired up, worth loading once to see the pattern.