TS Ideogram Designer

Ideogram is the image model that's genuinely good at text - which means its prompts are structured JSON, and writing that JSON by hand is misery. TS Ideogram Designer is the visual editor that spares you: a full-screen canvas where you drag and resize text and object blocks over an artboard, pick fonts and styles, and it assembles a valid Ideogram 4 structured-JSON caption for you, plus the width/height for an empty latent. What you see in the editor's WYSIWYG preview is what Ideogram is asked to draw.

It's the pack's flagship single-purpose node, and it's a big one - the README reads like a feature list for a design tool. For the Ideogram model (a real open-weights player in the KB's Ideogram panel, from the ex-Google-Brain team), precise text placement is the differentiator, and this node exists to make that placement something you do with your eyes instead of your parser.

How it works

Design mode is the headline: open the editor, add text/object blocks, and the caption is built from the layout. Two-level presets help - 10 layout templates (what you're making) and 10 styles (palette + fonts + look) - in a RU/EN interface. Each text block gets a single Text style dropdown (fonts are described, not named - Ideogram has no typeface selector), a weight (Thin/Regular/Bold), and optional outline/solid plate, each with its own color. Text size comes from how big you draw the block. The editor rectangles become normalized bounding boxes in the Ideogram schema (integers 0–1000, top-left origin), and the caption is assembled to the exact key order the schema requires.

There's also an Auto mode: give auto_prompt a plain-text idea, and the pack's Qwen engine (via the Generate Prompt button) turns it into the structured caption. That's the LLM-in-ComfyUI pattern from the KB's llm doc - a small local model doing structured rewriting.

The inputs and outputs

• design_json - serialized editor state, managed by the UI. You don't edit this by hand.
• mode - designer (build from the editor) or auto (generate from auto_prompt).
• auto_prompt / auto_caption / auto_seed - the Auto-mode path.
• image (optional) - a reference underlay so you can trace text over an image; doesn't affect the caption.

Outputs: json_prompt (the Ideogram 4 caption STRING), width and height (INTs, derived from aspect ratio × a 0.5–2 MP slider, always rounded to multiples of 32) - wire those straight into an empty-latent node.

Installing it

Part of comfyui-timesaver (ComfyUI Manager → "Timesaver", or manual clone + pip install -r requirements.txt + restart). No model downloads for the designer itself; Auto mode uses the pack's shared Qwen engine, which downloads weights on first use (into models/LLM/).

Where people get burned

The Auto mode's model isn't downloaded silently - the node shows a dialog first with the model, its real size, and the target folder, and Not now means no. That's the design, not a bug. The other thing to respect is the difference between the two outputs: json_prompt is what Ideogram gets; width/height are what your empty latent uses. They must agree, and the node keeps them in sync - but if you resize the latent elsewhere, the mismatch is on you. And a fair warning: this is a designer, so it has a learning curve proportional to its power. The inspector organizes it into steps (what you're making → how it should look → what's in the scene) and every control has a tooltip - worth spending ten minutes in before you need it.