Word Cloud

You want a wall of text where the words are sized by how often they show up, layered over an image, with a clean alpha channel so you can composite it - and you want it inside ComfyUI, not in some web tool you have to screenshot. That's this node. The name is not a lie: it's the word_cloud Python library by Andreas Mueller, wrapped into a graph node, plus jieba for splitting Chinese text that has no spaces to split on. Paste in a prompt, a tag list, or a whole transcript and get back a poster-shaped typography image.

It's from chflame163, the same author behind the big LayerStyle and DzNodes pack families. Expect Chinese-first documentation habits and a 😺dzNodes category prefix, but the code itself is a thin, readable wrapper - no model downloads, no network calls, nothing spooky.

Why you'd reach for it

The realistic uses: generate a background of keywords for a thumbnail or title card, make a tag-frequency cloud from a dataset's captions, or drop a word cloud into an img2img pass as a starting image where the text gets interpreted and re-rendered. The MASK output is the underrated part - because the node outputs image and mask, you can composite the cloud over anything with a normal Blender-style mask operation instead of fighting a hard background.

How it works

jieba chops Chinese into words, then WordCloud().process_text() counts frequencies for everything. keynote_words you set get their frequency artificially boosted by keynote_weight (default 60, added on top of the current max), so you can force "BREAKING" or "sale" to dominate without writing it a thousand times. stopwords are dropped - and the node also deletes them from the frequency dict afterward, a deliberate workaround for a quirk in the underlying library.

Layout is standard word_cloud: word size maps to frequency, prefer_horizontal (0.9 default) decides the horizontal/vertical mix, max_words caps the count, and random_state set to -1 means a fresh layout every run while any other number seeds it. Two inputs change the whole output: color_ref_image recolors the text using that image's palette (which overrides colormap), and mask_image turns the cloud into a silhouette - white areas are cut out, and it silently forces the output size to match the mask, ignoring width/height. transparent_background on (default) gives you RGBA plus the mask; turn it off and background_color (#FFFFFF) applies instead.

The inputs that actually matter

• text - your corpus. Leave it empty and you get a demo string, which is a nice touch.
• font_path - the dropdown is built from every .ttf/.otf in the directory named in font_dir.ini.
• mask_image and color_ref_image - the two shortcuts to a shaped, nicely-colored result.
• stopwords / keynote_words - comma (Chinese or English) or space separated.

Outputs: image (with alpha) and mask, both wired into any preview, Save Image, or compositing setup.

Installing it

ComfyUI Manager finds it by searching "WordCloud" or the author name. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/chflame163/ComfyUI_WordCloud
cd ComfyUI_WordCloud
# on the embedded Python build:
..\..\..\python_embeded\python.exe -m pip install -r requirements.txt

Then restart ComfyUI. The requirements (numpy, pillow, torch, matplotlib, wordcloud, jieba) are all things you almost certainly already have - torch ships with ComfyUI. There are no model files to hunt down; the only asset is the bundled Alibaba-PuHuiTi-Heavy.ttf font.

Troubleshooting

• CJK text renders as boxes (tofu). The font you picked has no Chinese glyphs. Stick with the bundled Alibaba font, or add a CJK font to your font_dir.
• On Linux/macOS, only one font shows up. font_dir.ini defaults to C:\Windows\fonts, which doesn't exist off Windows; the plugin falls back to its bundled font. Edit font_dir.ini to point at a real font dir (e.g. /usr/share/fonts) and restart.
• Old saved workflows. The README's own reminder: after a font change, re-select the font in font_path for nodes saved before the change, or the workflow may load with a stale path.
• "Why is my size setting ignored?" A connected mask_image owns the dimensions. That's by design.