Krea2Edit Normalized Attention Guidance

This is the image-edit cousin of the pack's text-to-image NAG node, and it answers a specific frustration: you're doing a Krea 2 edit ("change the bird's feathers to blue"), the model keeps leaving the old color in, and the normal negative-prompt fix is dead because Krea 2 runs at CFG 1. Krea2Edit Normalized Attention Guidance fuses the Krea2Edit forward pass with Normalized Attention Guidance so you can say "brown feathers" as a negative during the edit - without destroying the reference identity you're trying to preserve.

Why a combined node

Krea2Edit steers by source-reference attention: the model pays attention to your source image while regenerating a target, and that's what holds identity and pose. The problem is that a naive NAG patch would guide all image tokens, which would smear your reference as much as your output. This node instead runs NAG only on the target-image tokens - the code literally keeps the [text | source references] prefix of the attention output and guides just the noisy tail. Source-reference attention stays on the positive path; the negative only subtracts from what's being edited.

There's a second, practical reason it exists as one node: stacking Krea2EditModelPatch with the T2I NAG node doesn't work. The combined node raises a ValueError if it sees either wrapper already on the model and tells you to connect the base/LoRA model directly.

Inputs that matter

Required: model, nag_negative (the conditioning to suppress - put the concept that keeps sneaking in, e.g. "brown feathers"), and source_latent (your reference image, encoded). The knobs are shared with the T2I node: phi (4.0, how hard to push away), tau (2.5, normalization cap), alpha (0.25, blend back toward the plain positive). The output is one patched model - sampler gets the same model you'd use without NAG.

The optional block is where Krea2Edit's power lives: source_image + vae re-encode the reference at the target's resolution on the fly (with fit_mode "fit" vs the legacy "crop"), source_latent_b / ref_boost_a add a second reference image with its own strength, ref_boost sets overall reference influence (1.0 = neutral), and ref_boost_mask restricts where the reference binds. Beginners: leave all of these alone. Defaults are sensible.

Install

Same pack as the T2I node, but you also need the edit framework it plugs into:

cd ComfyUI/custom_nodes
git clone https://github.com/iljung1106/ComfyUI-Krea2-NAG
git clone https://github.com/lbouaraba/comfyui-krea2edit   # v1.2.5 or newer

Restart ComfyUI and load workflows/Krea2Edit NAG - Raw FP8 Turbo.json. Expect the same Krea 2 file set as the T2I workflow: the FP8 raw checkpoint, the Turbo LoRA, the Qwen3-VL encoder (~8GB), and the Qwen-Image VAE. This pack itself has no pip dependencies.

Gotchas

• CFG must be 1.0, hard. If the sampler runs any branch other than the positive one, the node throws at runtime. This isn't optional - CFG > 1 isn't "less optimal," it's an error.
• Don't use CLIPTextEncode for the edit or the NAG negative. Krea2Edit was trained with image-grounded (Qwen3-VL) encoding; the README is explicit that stock text encode isn't appropriate for edit instructions. Use Krea2 Edit (grounded encode) for both prompts, wired to the same source image - that's the training path the model expects.
• Don't stack a Krea2EditModelPatch or the T2I NAG node ahead of this one; it errors on purpose, because the combined forward already includes both.
• Keep the negative specific. NAG is strong medicine, and a vague negative during an edit can drag the whole composition.