Krea 2 Image Edit Rebalance

If you're here because you googled "Krea 2 image edit," this is the node you were looking for - and the first thing to know is that Krea 2 doesn't have an official edit model yet. The team's said one is coming; it isn't out. So this node is the community's stopgap: it wires reference images into Krea 2's conditioning and adds edit-flavored controls on top, so you can push the model toward "change this part of this picture" without a real editor behind it. It's the most complete node in the pack, and the one the README's showcase workflows are built around.

Think of it as Krea 2 Encode Rebalance with the extra machinery an edit needs - steering, a layer multiplier, and step-based scheduling - folded in. Same Qwen3-VL vision-language encoder doing the heavy lifting; this node just gives you more handles on how hard the edit lands.

Inputs and outputs

Required:

• text - your edit instruction / prompt, multiline. As of the pack's June update this supports negative phrasing to strip unwanted effects, and later updates added multi-character reference support for posing.
• clip - Krea 2's encoder output.
• steering - default 1, range -2 to 2. The overall push of the edit. Higher drives the change harder; go negative to pull away from it. This is the one you'll reach for most.
• layer_multiplier - default 1. A global scale on the internal per-layer rebalance. Leave it at 1 unless the edit is too weak or too aggressive across the board.
• enable_step - default on. Toggles the step-scheduled application (the node hands back per-step conditioning). Turn it off for a simpler, flat edit.

Optional, and this is the reference-editing core: image1 through image4, each with an image1_tokens … image4_tokens dropdown (low/normal/high/max, default normal). Tokens = how much weight each reference image gets, same as the encode node. More tokens, stronger reference.

The output conditioning is a list (it carries the per-step scheduling), so wire it straight into your sampler's positive input - it's built to feed the sampler as-is, not to be chopped up.

Installing it

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

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

Restart. No pip deps or bundled files. The prerequisite is a working Krea 2 setup - model, Qwen3-VL text encoder (~8GB), Qwen-Image VAE (~250MB). Grab the linked Krea 2 Edit (Full Context).json from the repo; it's the fastest way to see the intended wiring, since the README doesn't document parameters.

Common issues

Steering is powerful and easy to overcook. Past about ±1.5 the edit tends to blow out - start at the default 1, move in 0.1–0.2 steps, and keep a fixed seed so you can actually see what changed. If the reference image isn't showing through, raise its tokens before you touch steering.

CFG 1 caveat, again: Krea 2 Turbo runs at CFG 1, keep a negative slot on your sampler (empty is fine), and skip ConditioningZeroOut - it doesn't behave for Krea's recommended settings, and the symptom is grainy output people blame on the wrong node.

Calibrate the ceiling. This is compared, fairly, to instruction editors like Flux Kontext or Qwen-Image-Edit - and it's not that. Those consume references in-context and do clean, targeted edits; this reshapes conditioning to approximate an edit. It's genuinely useful for style transfer and loose reference posing, and genuinely finicky. When the real Krea 2 Edit ships, expect this to become the fallback rather than the tool.