Create Flux Regional Cond

CreateRegionalCond is how you tell Flux "this area of the image should obey this prompt" - one node per region, chained together. It's the pack's implementation of InstantX's Regional Prompting for Flux: training-free masking of the attention so each area's text tokens attend only to that area's image tokens. The whole reason it exists is the same one that's been driving regional prompting since 2023: a single global prompt smears two subjects' attributes together, and no amount of prompt engineering reliably fixes that.

How it works

You give the node a conditioning and a mask, and it stores the pair as a region. The mask defines where the region lives on the canvas; the conditioning is the prompt's text tokens that should be confined to it. Feed multiple CreateRegionalCond nodes into one chain via prev_regions and they stack into a single REGION_COND list - one for the background, one for the left character, one for the right character, and so on.

Inputs:

• cond (CONDITIONING) - a CLIPTextEncode output for this region's prompt.
• mask (MASK) - any mask. The pack's example uses CreateShapeMask, but a mask from a paint node or segmentation works fine.
• prev_regions (REGION_COND, optional) - the chain link. Connect the previous node's output here to stack regions.

Output:

• REGION_COND - the accumulated region list, which feeds ApplyRegionalConds.

Where it fits

The regional workflow is: ConfigureModifiedFlux → ApplyRegionalConds (with the latent), and each region made by CreateRegionalCond feeding its REGION_COND in. ApplyRegionalConds also takes an optional FluxAttnOverride to pick which attention layers enforce the masks. The README notes you can combine regional prompting with Redux, but that Redux is so powerful it tends to dominate the generation - so don't expect the two to play nice without tuning strength way down.

Installation

Same as every node here: ComfyUI Manager → search "ComfyUI-Fluxtapoz" → install → restart, or git clone https://github.com/logtd/ComfyUI-Fluxtapoz into custom_nodes. einshape only, no models to download.

Common issues

• Regions bleeding into each other - the attributes from one area leak into its neighbor. The masks in regional prompting are soft, not hard, so some bleed is expected; the fix is usually a cleaner, tighter mask and making sure the regions' prompts don't describe overlapping concepts.
• Missing the base/background region. ApplyRegionalConds always treats the first item in the list as the background (it internally prepends an all-ones region), so don't be surprised that your first CreateRegionalCond acts as the backdrop rather than a foreground area.
• Chained nodes must connect through prev_regions - if you feed ApplyRegionalConds only the last node's output, you've still got the whole chain, but only if each node passed its predecessor through.