🌀 BV Regional Krea 2 Attention (Experimental)

Krea 2 has the strongest prompt comprehension of any open model that shipped in a while - its Qwen3-VL text encoder does real bounding-box-style control for free - but it's not a segmentation model, and putting two characters in one frame still tends to blend them. BV Regional Krea 2 Attention is the pack's experimental attempt to fix exactly that: joint-attention regional routing through Krea 2's DiT, flagged EXPERIMENTAL right in the node. The flag is the point. Read it as "this works, verified by the author, but it's new enough that your mileage may vary."

How it works - and the honest limit

Krea 2's transformer is a single-stream joint-attention design with a specific topology: there are 28 main blocks plus four upstream text-fusion/refiner blocks that execute before the public attention-patch seam. This node routes Global, Background, and regional Qwen3-VL contexts through all 28 main blocks - and leaves those four upstream blocks global. The consequence is stated plainly in the node's own description: this improves spatial prompt-to-region binding, but it is not strict end-to-end isolation. The four upstream blocks still mix text globally before routing takes over.

That's a rare thing to see an author admit in a node description, and it's the reason this backend earns its experimental label honestly. The verified tests showed stable character and outfit separation, region reassignment, and a wide empty center when prompts and masks supported that composition. Good routing, not magic.

Raw vs. Turbo, and the negative-prompt wrinkle

The node supports both Krea 2 Raw and Krea 2 Turbo, and they differ in exactly the way you'd expect from the ecosystem docs. Raw retains conventional CFG and negative-prompt semantics - you get real negatives. Turbo is normally used near CFG 1, where a separate negative branch has little or no practical effect - so the node's own description warns that Turbo negatives require a sampler CFG branch. That's the Krea-specific gotcha the KB flags too: on CFG-1 Krea paths people silently produce grainy output when the negative slot isn't handled right. If you're on Turbo and something looks off, check what your sampler is doing with the negative, not the regional node.

The knobs

Same family trio: attention_strength (1.0), start_percent (0.0), end_percent (0.5). Per-region strength stays independent, and strength-zero-hides-regional-slots rather than bypassing the patch - bypass the node for a true unpatched A/B. Outputs are patched_model, positive, negative for a standard KSampler.

Install

Search BV Node Pack in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/BlackVortexAI/bv_nodepack.git

Restart, hard-refresh with Ctrl + F5. No extra dependencies; bring your own Krea 2 model and Qwen3-VL text encoder.

The honest take

If you're already on Krea 2 and two-subject composition is your pain, this is worth a serious look - the 28-block routing is real and the separation tests are convincing. But because it's experimental and the four upstream blocks stay global, don't expect pixel-perfect isolation, and don't skip the negative-slot handling on Turbo. The pack's own research note on this backend is unusually candid, which is exactly the signal you want from an experimental patch that reaches into a model's attention.