Ideogram 4 Image Edit Rebalance

This is the full edit node for Ideogram 4 - the Ideogram sibling of Krea 2 Image Edit Rebalance. Same deal: Ideogram 4 has no official image-edit model, so this node bolts reference images and edit-style controls onto its conditioning to fake one. If you want to nudge an existing Ideogram render toward "change this, keep that" using a reference picture, this is the node.

Ideogram 4 is a genuinely capable renderer - it was the closed benchmark for in-image text and has strong character/IP recall - but it opened its weights in June 2026 as a text-to-image model, full stop. No editor. This node is the community's attempt to get edit-like behavior out of it by reshaping what the encoder feeds the sampler.

Inputs and outputs

Required:

• text - your edit prompt, multiline, negative phrasing included if you want to strip an effect.
• clip - Ideogram 4's encoder output.
• steering - default 1, range -2 to 2. The main dial: how hard the edit pushes. Negative values pull away from the change. This is the one you'll actually tune.
• layer_multiplier - default 1. Global scale on the internal per-layer rebalance; leave it unless the whole edit is too weak or too strong.
• enable_step - default on. Toggles the step-scheduled application (the output carries per-step conditioning). Off gives you a flat, simpler edit.

Optional reference slots: image1 through image4, each with an image1_tokens … image4_tokens dropdown (low/normal/high/max, default normal). Tokens set each reference's strength - more tokens, more influence.

The output conditioning is a list because of the step scheduling, so feed it straight into your sampler's positive input as-is.

Installing it

ComfyUI Manager: search ComfyUI-Conditioning-Rebalance, install, restart. Manual route:

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

Restart. No pip deps, no bundled models. You do need a functioning Ideogram 4 graph - model plus its encoder - for this node to have conditioning to work on.

Common issues

Steering overcooks fast. Above roughly ±1.5 the edit tends to blow out; start at 1 and move in small steps on a fixed seed so you can see the actual delta. If a reference isn't landing, raise its tokens before you crank steering.

The README documents none of this - it's screenshots and a clone line - so grab an example workflow.json from the repo and study the connections rather than guessing which socket goes where.

Be honest about the tier. People compare reference editing to Flux Kontext or Qwen-Image-Edit; this isn't that. Those are real in-context instruction editors doing clean, targeted changes. This reshapes conditioning to approximate an edit, and it's fiddlier and less predictable - good for style transfer and loose reference work, weak on precise "only change the hat" edits. It's a workaround until Ideogram or the community ships a proper editor, and the Ideogram side of this pack has seen less testing than the Krea side, so budget time for trial and error.