ModelMergeKrea2

ModelMergeKrea2 is the block-merge node for Krea 2, Krea's 12B diffusion transformer that took the open ecosystem away from Ideogram 4 in under a week in June 2026. It's the newest node in this whole family - and it arrived fast, because the community needed it fast. Within days of Krea 2's open release, LoRAs, fine-tunes, and checkpoints appeared, and the moment you have multiple Krea 2 fine-tunes, you want to merge them.

There's a second, sharper reason this node matters, and it's Krea 2's defining defect. The KB panel is upfront: the open weights went through alignment training the hosted version never got, and the refusal behavior doesn't stop at porn - it flattens facial expression, drops prompt clauses, leaves "craters" in the model. The community's fixes are merges. People blend Krea 2 with weights that don't have the alignment baked in, or merge fine-tunes over the refusal-prone areas block-by-block. This node is the surgical instrument for that, and it's why a model-merge node for a model this young got real usage instead of being a placeholder.

How it works

Standard family mechanism - a ModelMergeBlocks subclass, per-tensor interpolation result = ratio · model1 + (1 − ratio) · model2 with ratio selected by longest matching key prefix. Krea 2's unusual architecture means the key map is genuinely different from the MMDiT nodes:

• first., last. - the first and last projection layers.
• tmlp., txtmlp., tproj. - the timestep MLP, text MLP, and text projection - Krea 2's text-conditioning path.
• txtfusion.layerwise_blocks.0–1, txtfusion.projector., txtfusion.refiner_blocks.0–1 - the text-fusion stack: two layerwise blocks, a projector, two refiner blocks. This is Krea 2's text/image fusion machinery, and it's the part you'd reach for when a merge model isn't obeying prompts the way you want.
• blocks.0–27 - the 28 main transformer blocks. The meat of the model.

Every slider defaults to 1.0 (all model1).

The inputs that matter

If you're doing the "de-alignment" merge, the practical instinct is to hold blocks high where you like model1's behavior and pull specific ranges down toward a cleaner parent - but there's no published formula, and honestly the community was still mapping Krea 2's block semantics weeks after release. Expect to experiment. The txtfusion.* inputs are the distinctive ones to learn: they control how the model reads text, and fiddling with them changes prompt adherence in ways the plain blocks don't.

Output is one MODEL - sample it directly, or send it to Save Checkpoint to share.

Where people get burned

Krea 2 is young, so the usual "this is a stable, known technique" comfort doesn't apply. You're merging a 12B model, which means memory pressure - both parents plus the merged result in VRAM. And the license deserves a look before you publish anything: Krea 2 ships under its own community license (there's a whole KB panel on it), and whatever rights you have to share a merge depend on that license, not on the merge node.

Also note the merge caveat that applies to the whole family: blending two fine-tunes averages their defects. With Krea 2, that includes inheriting the alignment crater if either parent carries it.

How you get it

Core ComfyUI, model/merging/model specific, added June 2026 right alongside Krea 2 support. It's about as fresh as a node gets - if you're on an older ComfyUI build, update before looking for it.