⭐ Star Qwen-Rebalance-Prompter

Modern image models are LLM-encoded: your prompt is an instruction fed to a language model, not a bag of tags. And one of the first things you notice when you switch to that mental model is how often the output looks like the model read your prompt and only paid attention to the subject. Background comes out as mush, foreground as an afterthought, because the prompt itself was subject-shaped. StarQwenRebalancePrompter is a text-to-image prompt builder that attacks exactly that - hence the name.

It builds a structured generation prompt from separately-described depth planes: subject, foreground, midground, and background. Each is a multiline field, so you give the model an instruction for every layer of the image instead of one sentence where the subject eats the paragraph.

What it assembles

• composition_preset - the framing engine. Seven presets, from the default "rule of thirds, off-center subject, vertical orientation, depth layering" through "golden ratio, diagonal leading lines," "centered subject, radial composition," and others, down to Custom (which routes to custom_composition). This is where the structure comes from - the preset turns your layer descriptions into a composed scene rather than a list.
• color_tone and lighting_mood - the two fields people forget and then complain the image is flat. Filling them in is what turns "a room" into "a warm, amber-lit room."
• auto_generate_visual_guidance (on by default) - appends a general visual-guidance sentence to the output so the model gets explicit instructions about framing and depth. Turn it off and write custom_visual_guidance when you have specific camera language (and it's worth learning the camera-phrase habit - naming a lens or angle is among the highest-leverage additions you can make to an instruction prompt).
• auto_generate_caption - same idea for a caption/description of the scene, with custom_caption as the manual override.
• output_format - JSON or Plain Text. JSON is for when you're feeding the result into a structured pipeline or an LLM; Plain Text goes straight into your text encoder.

The single prompt output is the finished instruction - wire it into your CLIP encoder.

Why "rebalance" matters

Qwen-Image is a 20B model with genuinely strong prompt adherence. That cuts both ways: when you write a subject-heavy prompt, it obediently draws a subject-heavy image. Structuring the prompt in blocks - foreground, subject, background, color, light - is the documented shift for instruction-encoded models, and this node makes you do it by construction. You can't forget the background if the background has its own text box.

Install

Pack standard: ComfyUI_StarBetaNodes is the retired beta repo; the nodes live in ComfyUI_StarNodes now. Install via ComfyUI Manager (search Starnodes) or git clone https://github.com/Starnodes2024/ComfyUI_StarNodes into custom_nodes, restart. No models, no deps - pure prompt string assembly. If you're generating with the Qwen-Image family, this is the prompt node you'll actually reach for; if you're on an SDXL-lineage model, its tag-based habits may serve you better, so keep that in mind before rewiring your whole graph.