Text2Image _O

Diffusion models are famously bad at spelling. The trick that actually works is to stop asking the model to write text and instead give it text to look at: render the word on a canvas, feed that as a ControlNet condition, and let the model regenerate the scene while honoring the letter shapes. Text2Image _O is the renderer half of that trick - it takes a string and draws it onto a canvas, outputting a proper IMAGE tensor.

It's under O/text in the Quality of Life Suit. The ControlNet knowledge base essay calls out this exact pattern - embedding patterns like text, logos, and barcodes via spatial conditioning - and this node is one of the easiest ways to produce the text plate for it. Beyond ControlNet, it's handy for quick watermarks, title cards, or feeding a text mask into an inpainting workflow.

How it works

Under the hood it's PIL: the text is drawn with ImageFont.truetype and ImageDraw, then converted to a normalized float tensor (/255) and shaped into an IMAGE. The canvas dimensions are enforced to a multiple of 8, which keeps the output latent-friendly if you route it toward a sampler.

The inputs that matter

• text - the string to draw (multiline).
• font - a dropdown of the fonts in the pack's fonts/ folder. Ships with Calibri, Comic Sans (regular/italic/bold), and Alkatra. To add your own, drop a .ttf, .otf, or .ttc into custom_nodes/ComfyUI-QualityOfLifeSuit_Omar92/fonts/ and restart ComfyUI - it appears in the dropdown automatically.
• size - font size in points (0–255, default 36).
• font_R/G/B/A and background_R/G/B/A - RGBA colors, each channel 0–255. Default black text on white.
• width / height - canvas size in pixels (defaults 128×128).
• expand - true (default) auto-grows the canvas to fit the text; false keeps your canvas and clips.
• x / y - where the text lands. The node centers the text on (x, y) and clamps it to stay on-canvas, so these behave like a "focus point" rather than a raw top-left coordinate.

Output: one IMAGE.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/omar92/ComfyUI-QualityOfLifeSuit_Omar92.git

or ComfyUI Manager → search ComfyUI-QualityOfLifeSuit_Omar92, then restart. The fonts ship with the pack; no model downloads.

Gotchas and honest advice

• Keep expand on unless you have a specific canvas in mind; a text string longer than the canvas silently gets clamped (the draw position is pushed back on-canvas, so the tail can get cut off).
• The output is a flat text plate - it works best as a ControlNet/inpaint seed, not as "generated text." For truly embedded text in a scene, pair it with a ControlNet and let img2img/denoise do the blending. If you skip ControlNet and just feed the plate to a sampler at high denoise, the model will mostly ignore it. That's the one mistake people make here: they expect the plate itself to be the final image, and it isn't.