Ideogram 4 Encode Rebalance

The Ideogram 4 counterpart to the pack's Krea encode node. It encodes your prompt for Ideogram 4 and lets you fold in up to four reference images, producing one rebalanced conditioning that behaves a bit like an IP-Adapter without being one. If you're running Ideogram locally and want to prompt with a picture instead of only words, this is the entry point.

Quick context on the model: Ideogram 4 opened its weights in June 2026 as the closed-benchmark king of legible in-image text, and it turned out to also have unusually strong character and IP recall. It's a strong renderer that shipped with no native edit model, quantized-only and non-commercial. This node exists to squeeze reference-image control out of it anyway, by working on the conditioning.

Inputs and outputs

Required:

• text - your prompt, multiline.
• clip - Ideogram 4's encoder output.

Optional, and this is the point of the node: image1 through image4, each paired with image1_tokens … image4_tokens (low/normal/high/max, default normal). The tokens dropdown is the reference-strength dial - more tokens gives that image more room in the conditioning, so it influences the output harder. max lets the reference dominate; low is a gentle hint. One dial per image, which is how you balance several references against each other.

The output is a single conditioning, already rebalanced, that wires into your sampler's positive input. You don't need to chain the standalone Ideogram rebalance node after it - the encode and the layer-weighting are combined here.

Installing it

ComfyUI Manager: search ComfyUI-Conditioning-Rebalance, install, restart. Or clone it:

cd ComfyUI/custom_nodes
git clone https://github.com/nova452/ComfyUI-ConditioningKrea2Rebalance.git

Restart ComfyUI. No pip dependencies, no bundled models - but you need a working Ideogram 4 graph loaded, since this node feeds off its encoder. The nodes are just conditioning math; the model is the weight on disk.

Common issues

Don't max every token slot. Start references at normal and only climb if the image isn't showing through - overloading tokens fights your text prompt and tends to mush the result. If you're juggling multiple references, the pack's LoadImages node can batch a folder straight into the four slots.

The README is close to useless for parameters - it's screenshots and a git clone line. Pull an example workflow from the repo and copy the wiring.

Set expectations low-ish. This is a community workaround, not an official Ideogram editing pipeline, and Ideogram 4 settled into a narrower graphic-design niche fairly quickly while attention moved to Krea 2 - so there's less community road-testing on the Ideogram side of this pack (the page's search traffic is basically zero). It works, but you're pioneering. Fix your seed, change one token setting at a time, and compare outputs instead of trusting any single setting.