PIP Artistic Text Generator

Diffusion models are famously bad at rendering text - they'll happily give you a sign that says "LONDON" when you asked for "HELLO", and the longer the phrase, the worse the spelling. This node takes the other route: instead of begging the model, it draws your text onto the image with Pillow, applying styled effects defined in SVG files. It's the flagship node of the small PIP_ArtisticWords pack (chenpipi0807's "艺术文字生成器"), and it's the reason people install the pack at all.

What it does

You feed it an image, type the text you want, pick a style, and it returns a new image with that text rendered on top - auto-sized, word-wrapped, and placed inside a "safe area" so glow and shadow effects don't get chopped off at the frame edges. Think posters, thumbnails, watermark-style titles, meme captions that actually read correctly.

How it works

Every style is an SVG file in the pack's SVG/ folder. On load, StyleManager parses each one and converts it into a style dict (gradient fill, outline, drop shadow, inner shadow, outer glow). TextRenderer then fits your text into the safe area - it picks a font size and wraps on word boundaries so multi-word text stays on one tidy block. EffectsProcessor layers on the effects, and the result is alpha-composited over your background image. Nothing here touches a model: it's Pillow + OpenCV + numpy, so it runs in well under a second even on CPU.

The inputs that matter

• image - your background. It's required.
• text - the string to draw. Multiline, so paste in whatever.
• seed - 0 means random every run; set it to reproduce a specific random style/font combo.
• style - one of the 14 bundled styles (new-cyberpunk, new-luxury-gold, new-matrix-code, new-pastel-dream…) or random.
• color_match - enable only does anything when style is random: it samples the image's dominant color and picks a matching style from config/style_color_mapping.json. Nice trick, but it overrides your random choice.

The optional margin floats (margin_top 0.25, margin_bottom 0.15, margin_left/margin_right 0.1) define the safe area as fractions of image size - the one thing you'll actually tune when text gets clipped. opacity scales the whole overlay, and debug_info (none/basic/detailed) dumps style internals to the console when you're hunting a weird render.

Output: a single IMAGE tensor, ready for Save Image or to chain into whatever post-processing you like.

Installing it

Use ComfyUI Manager (search "PIP Artistic Words") or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/chenpipi0807/PIP_ArtisticWords.git
cd PIP_ArtisticWords
pip install -r requirements.txt

Then restart ComfyUI. Dependencies are light - Pillow, numpy, opencv-python, and webcolors; torch is already in ComfyUI. No model downloads, no GPU requirement for these nodes.

Where people get burned

• New styles need a restart. Drop an SVG into SVG/ and the dropdown won't see it until ComfyUI reloads. That's by design - styles are scanned at startup.
• Font licensing is your problem. The repo ships 方正风雅宋 (a 17 MB CJK font) plus four Latin fonts, and the README says it outright: "字体版权问题请自行了解后使用" - check the license before you use the pack's fonts in commercial work.
• The node prints on every run. Even with debug_info set to none, there's an unconditional console line about the input image shape. Harmless, just noisy.
• Chinese UI labels. The pack's author writes in Chinese; several sibling nodes use Chinese parameter names. This node is the exception - its labels are English.

If text lands partially outside the frame, raise the margins. That's the single most common fix, and it's the one the author calls out too.