Krea 2 Encode Rebalance

This is the node that actually earns the pack its "IP-Adapter-like" pitch. It's a text encoder for Krea 2 that also takes up to four reference images and folds them into the same conditioning - so instead of just describing what you want in words, you can show it a face, a style, an object, and let the model borrow from the picture. On a model with no native image-edit path, that's the whole appeal.

Krea 2's encoder is Qwen3-VL - the VL part meaning it's vision-language, it can look at images. This node leans on that: your reference images get encoded alongside your prompt and handed to the sampler as one blended conditioning. It's not a trained adapter like real IP-Adapter (that was a 22M-parameter network with its own cross-attention path); it's a lighter, hackier route to the same "prompt with a picture" outcome. Lighter and hackier, but it needs no extra weights and no key.

Inputs and outputs

The two required inputs are the basics:

• text - your prompt, multiline.
• clip - Krea 2's CLIP/encoder (the Qwen3-VL loader output).

Then the interesting part, all optional: image1 through image4, each with a matching image1_tokens … image4_tokens dropdown set to low, normal, high, or max (default normal). The token setting is your reference-strength dial. More tokens means the image gets more room in the conditioning, so it pushes harder on the result - max makes the reference dominate, low makes it a faint nudge. That's the closest thing this pack has to IP-Adapter's weight slider, one per image.

The output is a single conditioning. Wire it into your sampler's positive input. Under the hood it's already rebalanced (that's the "Rebalance" in the name), so you don't strictly need to chain the standalone rebalance node after it - this one bakes the encode and the layer-weighting together.

Installing it

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

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

Restart ComfyUI. No pip deps, no bundled models - but you need Krea 2 itself loaded: the model, the Qwen3-VL text encoder (~8GB), and the Qwen-Image VAE (~250MB). The encode node feeds off that encoder, so it's a hard prerequisite.

Common issues

Start reference-heavy prompts at normal tokens and only climb to high/max if the image isn't showing through - cranking every slot to max tends to fight your text prompt and flatten the result. If you're feeding multiple references, LoadImages (also in this pack) will batch a folder into the four slots for you.

The CFG-1 rule bites here too. Krea 2 Turbo runs at CFG 1, and the community's clear finding is you still want a negative-prompt slot wired to your sampler even if it's empty; ConditioningZeroOut misbehaves on Krea's recommended settings, so leave it out. Grainy output at CFG 1 is usually this, not the encode node.

Bigger picture: there's no official Krea 2 Edit model yet - the team's said one is "in the works," and until it lands, this encode-with-references trick is a community workaround. It can do style transfer and loose edits, but it's the "kinda works with a bit of fiddling" tier, not Kontext-grade instruction editing. Save your seed, move one token setting at a time, and judge by comparison.