Donut Prompt Injection

Every serious prompt-writer keeps a scratchpad of the phrases that reliably work - "Kodak Portra 400," "Rembrandt lighting," "golden hour morning." DonutPromptInjection is that scratchpad turned into a node: instead of remembering the exact wording, you pick from dropdowns organized like a photography production sheet - main category, subcategory, style, camera, lighting, time of day, weather, color grade, climate, pose, expression, hair, skin tone, outfit, location, vibe, and about a dozen more. It assembles the final prompt, in your chosen order, and hands it to your CLIP encode. It's prompt engineering as a form you fill out.

The name "injection" is doing a lot of work: it's about injecting style into a base prompt, not the security meaning. And it's chainable - feed its output into another copy and you can layer styles.

How it works

Each category is an enum where the value is either a specific choice, Random, or None (skip). The choices map to curated prompt phrases stored in the pack - "Vintage Kodak" becomes a full "vintage Kodak Portra 400 film photograph, creamy skin tones, soft grain…" string, and so on through dozens of styles, 60+ expressions, 40+ camera shots. If you pick Random, it picks from that category using the seed input, so randomness is reproducible - set a seed, get the same random look every run.

The assembly knobs:

• order - Style First or Prompt First decides whether your base prompt goes at the front or the style blocks lead.
• separator (default ", ") - what joins the pieces.
• seed - drives all the Random picks.

Two outputs: prompt (the assembled string for your encode) and style_preview (a separate string showing just the injected style block, so you can see what was added and tweak it without reading the whole prompt).

Install

cd ComfyUI/custom_nodes
git clone https://github.com/DonutsDelivery/ComfyUI-DonutNodes.git donutnodes
cd donutnodes
python -m pip install -r requirements.txt

or ComfyUI Manager → "DonutNodes," restart.

Using it well - and the honest caveat

The real value is the Random + seed combination: it's a cheap exploration tool - randomize style/lighting/camera on a fixed image-to-image pass and sweep a bunch of looks deterministically. But there's a 2026-model trap the KB calls out hard: on LLM-encoded models (Krea 2, Z-Image, Klein), your prompt is an instruction to a language model, not a token bag. These curated photographic phrases were written the SDXL way and mostly transfer as natural-language description - that's fine, even good - but don't expect Random to be tasteful, and if you're on an SDXL-lineage model, the comma-separated style lists this produces are exactly in their wheelhouse. Which model you're on decides whether this node's output reads as magic or as decoration.