🌀 BV Regional Z-Image Attention

Z-Image Turbo is a fantastic model with one structural quirk that makes regional prompting weird: it has no conventional negative prompt branch. It's guidance-distilled, runs near CFG 1, and its "negative" is essentially a no-op. So when you wire a regional system into it, you can't lean on negative conditioning to carve out "not here." BV Regional Z-Image Attention is the pack's answer - it routes your regions through the model's own joint attention (the unified text/image attention matrix) and hands you sampler-ready positive plus a zero negative, because that's the honest thing to emit.

How it works

Z-Image uses the Lumina2/S3-DiT-style joint-attention architecture. This node patches a clone of the model and routes Global, Background, and regional text slots through that unified matrix: Global text stays available throughout the image, Background is limited to uncovered pixels, and each regional slot is exposed inside its rendered mask. joint overlaps expose every participating regional context.

Because Z-Image Turbo doesn't use a conventional negative CFG branch, the node emits zero negative conditioning - but ComfyUI's KSampler still demands a negative input, so this node gives you one that does nothing rather than one that actively fights you. That's the mechanism detail worth internalizing: an empty negative isn't a bug in this node, it's the correct behavior for the model.

The backend is architecture-gated. It rejects FLUX, SDXL, Anima, and unrelated models rather than guessing. Wrong model, clean error, no mystery.

The knobs

Same trio as the other attention backends: attention_strength (1.0), start_percent (0.0), end_percent (0.5). The README's verified recipe was 1024×1024 with res_multistep, eight steps, CFG 1.0, the simple scheduler, and a ModelSamplingAuraFlow shift of 3.0 - worth copying as a starting point, not as gospel. The pack stresses these are test values, not universal quality recommendations. And remember the standard BV caveat: Attention Strength 0 hides regional slots rather than bypassing the patch; bypass the node for a true unpatched run.

What the evidence showed

The pack's comparison against untuned Native Conditioning - same seed, document, and settings - found both approaches preserved broad left/right placement, but joint-attention routing held specific color attributes better: white/cyan and black/orange stayed on their characters, while the native run let the right-side suit drift toward beige and gold. That's a clean demonstration of what attention routing buys you on a model whose text encoder is already strong: it's about spatial binding, not about understanding the prompt.

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 deps. You bring the Z-Image Turbo checkpoint yourself (the verified family is on HuggingFace as Tongyi-MAI/Z-Image-Turbo).

The honest take

If you run Z-Image Turbo and you want two subjects to stop merging, this is the node - there's no other backend in the pack that speaks this model's attention architecture. Just don't expect negative-prompt magic you can't have. And one more thing: because the model runs at CFG 1, your regional quality is almost entirely prompt-and-geometry driven. Spend the time in the editor, not the strength sliders.