Visual Regional Prompting for ComfyUI

This post contains separate, ready-to-use regional-prompting workflows for

multiple model families:

- Illustrious / Pony / SDXL

- Anima with optional LLLite layout control and regional Attention LoRAs

- Z-Image Turbo

- FLUX.2 Klein 9B

- Krea 2 Attention and regional LoRAs (Experimental)

Update v0.15.0 – Spatially Aware Prompt Enhancer

Adds spatially aware regional prompt enhancement using canvas geometry, region relationships, overlap and object ownership.

Supports ComfyUI CLIP, Ollama and OpenAI-compatible providers, with persistent response caching and separately stored API keys. Tested model recommendations and an Anima comparison are included in the README.

Update v0.14.0 — Impact Detailer Integration

Added the new BV Regional Detailer Mask node for integration with ComfyUI Impact Pack.

It allows you to:

• Select any BV region as the detailer mask

• Create an Impact-compatible BASIC_PIPE directly

• Combine Global, Background, Primary and additional Context Region prompts

• Adjust each prompt scope with individual influence values

• Create detailer-only regions that do not affect the initial generation

• Preserve all region selections and influence values inside the workflow

Connect the node’s image, mask and basic_pipe outputs directly to MaskDetailer (pipe).

Update v0.12.0 — Single-Pass Regional LoRAs

BV Regional Krea 2 Attention now uses token-gated single-pass execution for regional LoRAs and LoKr adapters by default. Regional adapter effects are applied directly to their assigned text and image tokens inside one model pass, resulting in significantly faster generation and, in my testing, better scene coherence with softer regional transitions.
The previous multi-pass implementation remains available as multipass_legacy for compatibility and reproduction of older results. Existing workflows without an explicit mode will use the new single-pass default after updating BV Node Pack.

WARNING

> **Temporary Krea 2 FP8 limitation:** Regional LoRA hooks may fail with quantized Krea 2 models due to an upstream ComfyUI issue:
> `AttributeError: 'Linear' object has no attribute 'weight_scale'`.
> See [ComfyUI #14382](https://github.com/Comfy-Org/ComfyUI/issues/14382).
> Until this is fixed upstream, use a non-quantized Krea 2 model or apply the temporary ComfyUI core workaround described in the Krea 2 workflow documentation.

Choose the downloadable workflow version that matches your model. The visual

editor and regional-document format are shared, while each workflow uses the

compiler and model loader required by its architecture.

The workflows showcase BV Regional Editor, a visual authoring system for

regional prompting in ComfyUI. Instead of assembling large prompt, mask and

conditioning chains, you can create and edit all regions inside one scalable

editor.

Each region can have its own positive and negative prompt, strength, feathering,

priority, display color and geometry.

Main features

- Any number of named prompt regions

- Rectangle and freehand brush geometry

- Multiple shapes inside one region

- Additive and subtractive brush layers

- Overlapping regions with joint-attention semantics

- Region strength, feathering and priority controls

- Global and Background prompts

- Layer locking, visibility and reordering

- Merge and split layer operations

- Quick Prompt Editor for fast text changes

- Floating and workspace editor modes

- Background-image preview with adjustable opacity

- Built-in local prompt autocomplete

- Import and export of reusable .bv-regional.json documents

- Named regional LoRA assignments for Anima and Krea 2 Attention

- Empty and temporarily disabled LoRA stacks as valid no-ops

- Standard KSampler and SamplerCustom compatibility

Included workflow versions

Illustrious / Pony / SDXL

This workflow uses BV Regional SDXL Attention to route independently encoded

Global, Background and regional prompt contexts through SDXL cross-attention.

The backend has been interactively verified with WAI Illustrious SDXL and

Pony Diffusion V6 XL. Other compatible SDXL checkpoints may use the same

architecture, but remain unverified until tested.

The included example separates a crimson knight and an azure mage in a shared

conservatory scene while keeping their hair, eyes, clothing and accessories bound

to their authored regions.

Recommended starting values:

- Attention Strength: 1.0

- Start: 0.0

- End: 0.5

Anima

This workflow uses BV Regional Anima Conditioning to route regional prompts

through Anima's attention layers.

The optional BV Regional Anima LLLite node converts the same editor document

into a deterministic color-control image and adds learned layout guidance through

ComfyUI's model-patch system.

Anima model
    -> BV Regional Anima Conditioning
    -> BV Regional Anima LLLite (optional)
    -> KSampler

The LLLite adapter is distributed separately and is not bundled with BV Node Pack:

https://huggingface.co/Sen-sou/Anima-LLLite-Regional-Controlnet

The new Anima Attention LoRA workflow assigns Nyamena to the left region and

Myaley to the right, both at strength 0.8. Approximately 13 percent overlap

around the hands lets both characters face one another, maintain eye contact and

hold hands while their yellow/orange and purple identities remain locally distinct.

This example also shows the relationship between prompting and routing: regional

Attention LoRA passes retain the separate identities, while deliberate interaction

prompts help Anima construct a coherent shared pose.

Both LoRAs come from the same series and creator and recommend similar strengths.

This is a favorable validation case, not a guarantee for arbitrary LoRA pairings.

Z-Image Turbo

This workflow uses BV Regional Z-Image Attention to route Global, Background

and regional prompts through Z-Image Turbo's unified text/image attention.

Because the distilled Turbo model has no conventional negative CFG branch, the

node emits zero negative conditioning for ComfyUI sampling nodes that require a

negative input.

The backend is architecture-gated and rejects SDXL, FLUX, Anima and unrelated

models instead of patching them speculatively.

FLUX.2 Klein 9B

This workflow uses BV Regional FLUX.2 Klein 9B Attention to route regional

Qwen3-8B text contexts through the model's double- and single-stream attention

blocks.

The node deliberately accepts only the exact FLUX.2 Klein 9B architecture. It

does not support Klein 4B, full FLUX.2 or FLUX.1 checkpoints.

The distilled profile has no conventional negative-prompt branch, so the workflow

provides the zero negative conditioning required by standard ComfyUI samplers.

Krea 2 (Experimental)

This workflow uses BV Regional Krea 2 Attention (Experimental) to route

Global, Background and regional Qwen3-VL contexts through the 28 main

single-stream attention blocks.

Four upstream text-fusion blocks execute before the public attention-patch seam

and remain global. The backend therefore provides useful spatial routing, but not

strict end-to-end isolation.

Krea 2 Raw supports conventional CFG and negative prompts. Krea 2 Turbo is

normally used near CFG 1, where a separate negative branch has little or no

practical effect.

Krea 2 regional LoRAs were validated in two scenarios. The character test assigns

MaluX and AiriX to overlapping regions and retains distinct appearance and

skin tone around their shared hand area. Both LoRAs were trained by the same creator,

so this is another favorable compatibility case.

The harder style test combines Detailcore Forever on the left and **CRT

Vaporwave** on the right in one continuous futuristic street. Dense mechanical

linework remains concentrated on the left, while cyan-magenta retro-futuristic

styling remains concentrated on the right. The road, perspective and lighting stay

continuous through the 10-percent overlap.

A control generated with lora_registry and lora_bindings disconnected preserves

the prompts, seed, geometry and sampler settings. It retains the prompt-driven scene

while substantially reducing both trained visual languages. The stack-swap result is

not used as primary evidence because style LoRAs are content- and training-dependent,

not neutral filters, and the prompts were tailored to their respective priors.

How the workflows work

Every version follows the same basic structure:

BV Regional Prompt
    -> matching BV Regional Attention/Conditioning node
    -> standard KSampler or SamplerCustom workflow

1. Open BV Regional Prompt in the Regional Editor.

2. Define Global and Background prompts.

3. Create named regions and draw rectangles or brush layers.

4. Connect the regional document to the backend matching your model family.

5. Connect the patched model and conditioning outputs to the sampler.

6. Optionally return generated images to the editor as a visual background.

The model-specific backend is important. SDXL cross-attention, Anima attention

and the unified text/image attention used by Z-Image, FLUX.2 and Krea 2 require

different routing implementations.

Regional LoRA connections

External LORA_STACK
    -> BV Named LoRA Stack
    -> lora_registry

BV Regional Prompt
    -> regional
    -> lora_bindings

MODEL + CLIP + regional + registry + bindings
    -> BV Regional Anima Conditioning
       or BV Regional Krea 2 Attention (Experimental)
    -> standard KSampler

The external stack node remains the owner of its LoRA entries and strengths. BV

Named LoRA Stack only provides a stable workflow-local name that can be selected

for Global, Background or an individual editor region. Empty and fully disabled

assigned stacks remain valid no-ops for controlled comparisons.

Installation

ComfyUI Nodes Manager

1. Open Nodes Manager or Extensions in ComfyUI.

2. Search for BV Node Pack.

3. Install the package.

4. Restart ComfyUI.

5. Perform a browser hard refresh.

The installation includes the complete BV Node Pack and Regional Editor. No

separate editor extension is required.

GitHub

Source code, documentation, examples and issue tracking:

https://github.com/BlackVortexAI/bv_nodepack

For manual installation, clone the repository into the ComfyUI custom_nodes

directory and restart ComfyUI.

The published regional LoRA examples use ComfyUI-Lora-Manager as their external

LORA_STACK producer. Other compatible stack producers can be substituted:

https://github.com/willmiao/ComfyUI-Lora-Manager

Required model files

BV Node Pack does not bundle or download model weights, text encoders, VAEs,

LoRAs or the optional Anima LLLite adapter. Install the files required by the

selected workflow and follow their respective licenses.

- WAI Illustrious SDXL:

https://civarchive.com/models/827184/wai-illustrious-sdxl

- Pony Diffusion V6 XL:

https://civarchive.com/models/257749/pony-diffusion-v6-xl

- Z-Image Turbo:

https://huggingface.co/Tongyi-MAI/Z-Image-Turbo

- FLUX.2 Klein 9B:

https://huggingface.co/black-forest-labs/FLUX.2-klein-9B

- Krea 2 Raw:

https://huggingface.co/krea/Krea-2-Raw

- Krea 2 Turbo:

https://huggingface.co/krea/Krea-2-Turbo

- Anima Regional LLLite:

https://huggingface.co/Sen-sou/Anima-LLLite-Regional-Controlnet

- Nyamena LoRA:

https://civarchive.com/models/2749213/nyamena-around-40-otoko-no-isekai-tsuuhan-anima?modelVersionId=3092643

- Myaley LoRA:

https://civarchive.com/models/2749226/myaley-around-40-otoko-no-isekai-tsuuhan-anima?modelVersionId=3092658

- MaluX LoRA:

https://civarchive.com/models/2853443/bemyhero-malux

- AiriX LoRA:

https://civarchive.com/models/2851019/bemyhero-airix

- Detailcore Forever:

https://civarchive.com/models/2747452/detailcore-forever-by-stx?modelVersionId=3221683

- CRT Vaporwave:

https://civarchive.com/models/2856432/crt-vaporwave?modelVersionId=3226199

Local filenames and model directories may differ from those stored in the

downloaded workflow. After loading a workflow, select the corresponding files

installed in your own ComfyUI environment.

Important limitations

Regional attention and optional LLLite control provide strong generative guidance,

not pixel-perfect object placement or hard segmentation.

The model still interprets composition, anatomy, pose, scale and region

boundaries. Results also depend on the checkpoint, prompt vocabulary, seed,

sampler, scheduler, steps, guidance, region geometry, overlap and timing settings.

Regional LoRA support in the Anima and Krea 2 Attention nodes remains experimental.

Every distinct effective model-side LoRA stack adds another masked denoiser

evaluation. A global baseline plus two different regional stacks normally requires

three model evaluations per sampling step. Krea 2 is especially demanding; begin at

768 x 768 on constrained hardware and expect substantially longer sampling plus

high VRAM and system-memory use.

Krea 2 also contains four upstream text-fusion blocks outside the public attention

patch seam. Useful spatial routing is demonstrated, but strict end-to-end isolation

is not claimed.

For a meaningful A/B comparison, keep the model, seed, prompts and all sampling

settings fixed. Change only the regional backend or bypass only Anima LLLite when

testing its additional layout control.

Suggested experimentation

Try changing:

- region size, position and overlap

- regional prompt strength

- region feathering and priority

- attention strength

- attention start and end percentages

- Global and Background prompts

- sampler and scheduler settings appropriate for the selected model

- Anima LLLite strength and active interval

- LoRA stack assignment, strength and trigger compatibility

More than regional prompting

BV Node Pack also includes:

- Smart Pipes with incremental slot inheritance

- Configurable Subgraph interfaces

- Group- and Subgraph-aware Control Center operations

- Wireless workflow connections

- Structured prompt AST tools

- Prompt deconstruction and regional mask rendering

- Deterministic latent utilities

- Image preview and save-to-editor nodes

- Local CSV/TSV prompt autocomplete

Feedback wanted

Feedback is especially welcome about:

- workflows with more than two regions

- complex brush masks

- overlapping characters or objects

- useful defaults for each model family

- additional compatible SDXL checkpoints

- experimental Krea 2 behavior

- regional LoRA combinations and disabled-stack comparisons

- usability of the floating editor

- compatibility problems with current ComfyUI versions

Leave feedback on Civitai or open an issue on GitHub:

https://github.com/BlackVortexAI/bv_nodepack/issues

Bug reports, workflow examples and feature suggestions are very welcome.

Development disclosure

This project was developed with AI-assisted programming. Architecture, testing,

workflow design and final decisions were reviewed and directed by the project

author.

Credits

Special thanks to Sen-sou for publishing the Anima Regional LLLite model and

related regional-conditioning research.

Thanks to the creators of the Nyamena, Myaley, MaluX, AiriX, Detailcore Forever and

CRT Vaporwave LoRAs used in the validation workflows, and to willmiao for

ComfyUI-Lora-Manager.

Thanks also to the ComfyUI community and the authors of regional-prompting,

configurable-pipe and workflow-navigation projects that inspired parts of BV Node

Pack. Detailed acknowledgements and third-party notices are available in the

GitHub repository.