ComfyUI Regional Prompting for Multi-Subject Compositions: Regional, Cutoff, and Conditioning Combine

You set up a 1024x768 image with a short-haired girl in red on the left and a boy in a blue jacket on the right. The result puts red on the jacket and blue in the girl’s hair, mixing both subjects’ clothing and hair colors. This is not a seed problem or a weak model. CLIP reads the full sentence as context: in “red hair, blue jacket,” red can drift toward jacket and blue toward hair. Regional conditioning nodes can restrict a prompt to an area in the latent, but the embedding still carries globally encoded context. There are three practical routes: native coordinate regions with ConditioningSetArea + Combine, attribute binding with Cutoff, and mask-based regional sampling with Impact Pack Regional. Each solves a different problem and has different limits.

Why multi-subject prompts bleed

CLIP encodes global context

Stable Diffusion’s CLIP text encoder does not encode every word independently. It reads the sentence as context. In “blue hair, yellow eyes,” blue can influence eyes because tokens interact while the embedding is generated. ConditioningSetArea limits conditioning to a latent region, but the embedding itself still comes from global encoding, so blue can still drift toward the eyes. Regional prompts reduce spatial bleed, such as red on the left affecting the right side, but they cannot prevent token-level attribute drift, such as blue affecting eyes.

Three types of bleed

Type of bleed	Typical symptom	Direction to take
Spatial bleed	Red clothing on the left affects the right; foreground blue affects the background	Regional conditioning (Set Area / RegionalPrompt)
Attribute bleed	In “blue hair, yellow eyes,” blue affects the eyes	Attribute binding (Cutoff)
Inconsistent identity	The same character has a different face with another seed	FaceID / LoRA (not covered here)

Selection table: problem → approach → tool → poor fit

Problem	Recommended approach	Tool nodes	Poor fit
Left-right character split	Native coordinate regions	ConditioningSetArea + ConditioningCombine	Complex masks, because coordinates must be calculated
Color or clothing bleed	Attribute binding	Cutoff BasePrompt + Set Region + Regions To Conditioning	Coordinate regions; SDXL requires testing because the README reports weak results with SDXL 0.9
Complex mask regions	Mask-based regional sampling	Impact Pack: RegionalPrompt + CombineRegionalPrompts + RegionalSampler	Requires Impact Pack and may produce missing-node issues
Simple color-block masks	Simplified regional nodes	Inspire Pack: Regional Conditioning By Color Mask	Depends on Impact Pack or IPAdapter Plus
Incorrect poses	Pose control	ControlNet (OpenPose / Canny / Depth)	Does not solve attribute bleed
Inconsistent identity	Identity control	FaceID / ReActor / LoRA	Does not solve regional prompting

A useful combination is: regional prompts set the position → OpenPose sets the pose → low-strength global conditioning at 0.3-0.5 unifies the atmosphere. You must align OpenPose, ControlNet, or IPAdapter positions with the regions manually. The nodes do not automatically know that the left OpenPose subject belongs to the left prompt region.

Native route: ConditioningSetArea + Combine

Coordinate parameters

Parameter	Meaning	Unit	Range	Notes
x / y	Top-left coordinate of the region	Pixels	0 - MAX_RESOLUTION	The coordinate origin is at the top left
width / height	Region width and height	Pixels	64 - MAX_RESOLUTION, step 8	The UI steps by 8 because the source divides values by 8 when writing to the latent
strength	Regional conditioning weight	Float	0 - 10, default 1.0	Overlapping-region weights are normalized during sampling

The step size is 8 because ConditioningSetArea.append() writes the latent area as (height // 8, width // 8, y // 8, x // 8). A 512x512 image becomes 64x64 in latent space, so coordinates should align to multiples of 8. ConditioningSetAreaPercentage accepts values from 0 to 1 if you prefer percentages over manual pixel calculations.

Wiring from prompts to the sampler

This example splits two subjects across a 1024x768 canvas:

Step 1: region for the left-subject prompt

• Left-subject prompt → CLIP Text Encode (conditioning output)
• Conditioning → ConditioningSetArea (x=0, y=0, width=512, height=768, strength=1.0)
• Output the area-tagged conditioning

Step 2: region for the right-subject prompt

• Right-subject prompt → CLIP Text Encode (conditioning output)
• Conditioning → ConditioningSetArea (x=512, y=0, width=512, height=768, strength=1.0)
• Output the area-tagged conditioning

Step 3: combine conditioning

• Left Set Area output → ConditioningCombine (input 1)
• Right Set Area output → ConditioningCombine (input 2)
• Combine output → KSampler positive conditioning

Optional step: global background conditioning

• Background-scene prompt → CLIP Text Encode
• Conditioning → ConditioningSetArea (x=0, y=0, width=1024, height=768, strength=0.3-0.5)
• Combine it with the regional conditioning at a lower strength to unify the atmosphere

The key point is that strength is not “opacity.” It is the weight of the noise prediction. When regions overlap, multiple conditioning weights are normalized before sampling, so you do not need to calculate the total manually.

Overlapping regions and strength

When regions overlap, strength controls the mixture of their conditioning. Sampling normalizes the weights automatically. A sensible starting point is strength=1.0 for foreground subjects and strength=0.3-0.5 for global background conditioning. Lowering background strength reduces interference with the subject, but it cannot remove bleed entirely. Conflicting colors in both character prompts can still interact through CLIP’s global context.

Attribute-binding route: Cutoff

Where it fits and how it works

The classic case is blue affecting eyes in “blue hair, yellow eyes” because CLIP encodes the whole sentence as context. Cutoff aims to make an attribute affect only a specified subset of the prompt. It masks target tokens, regenerates the embedding, and uses the resulting direction to isolate attribute relationships.

Cutoff is not a coordinate-based regional prompt. It cares about token relationships, not positions. Its nodes are Cutoff BasePrompt, Cutoff Set Region, Cutoff Regions To Conditioning, and Cutoff Regions To Conditioning (ADV), under conditioning > cutoff.

SDXL warning: the README reports poor results with SDXL 0.9. Results with current SDXL or FLUX versions require testing, so Cutoff is not a reliable universal solution.

Installation and use

Installation:

• Search for ComfyUI_Cutoff in ComfyUI-Manager
• Or clone it manually with git clone https://github.com/BlenderNeko/ComfyUI_Cutoff into custom_nodes
• Restart ComfyUI after installation

Basic flow:

• Enter the complete prompt in BasePrompt, such as “blue hair, yellow eyes”
• Define the attribute binding in Set Region, such as binding “blue” to “hair”
• Generate conditioning with Regions To Conditioning
• Optionally pass that conditioning through Set Area to restrict its position

Limits and cautions

Cutoff is a poor fit for coordinate splits. It binds token attributes but does not place “red clothing on the left and blue clothing on the right.” To address both attribute and spatial bleed, bind attributes with Cutoff first and then restrict the result with Set Area.

Its SDXL results are uncertain. The README specifically reports weak results with SDXL 0.9, and current versions need testing. If SDXL is your main model, test Cutoff on a simple workflow before depending on it.

Mask-based regional sampling: Impact Pack Regional

RegionalPrompt and RegionalSampler nodes

Node	Input	Output	Purpose
RegionalPrompt	mask + advanced_sampler + variation_seed/strength/method	REGIONAL_PROMPTS	Binds a mask to a sampler and creates a regional prompt
CombineRegionalPrompts	Multiple REGIONAL_PROMPTS	REGIONAL_PROMPTS	Combines multiple regional prompts
RegionalSampler	base_sampler + REGIONAL_PROMPTS + overlap_factor + restore_latent	latent	Applies sampling to regions at every step
RegionalSamplerAdvanced	RegionalSampler inputs plus step control	latent	Uses steps instead of denoise for control

Parameter meanings:

• overlap_factor: controls the transition range outside the region
• restore_latent: restores the base latent outside the mask during regional sampling to reduce extra noise there

RegionalSampler applies regional sampling at every step. TwoSamplersForMask, by contrast, performs global sampling and then combines by mask. Per-step sampling gives the former finer control over each region.

Installation and troubleshooting

Installation:

• Search for ComfyUI Impact Pack in ComfyUI-Manager
• Restart ComfyUI and refresh the browser after installation

Common troubleshooting:

• Red or missing nodes → see importing workflows and installing missing nodes
• Python dependency errors → see ComfyUI troubleshooting and maintenance, covered in a later article

Simplified route: Inspire Pack Regional Conditioning (optional)

If you do not want to calculate Set Area coordinates, a color-block mask can simplify the workflow. Inspire Pack provides these nodes:

Node	Purpose
Regional Prompt Simple (Inspire)	Simplified regional prompt input
Regional Prompt By Color Mask (Inspire)	Defines regions with a color-block mask
Regional Conditioning Simple (Inspire)	Simplified regional conditioning
Regional Conditioning By Color Mask (Inspire)	Defines conditioning regions with a color-block mask
Regional IPAdapter Mask / By Color Mask / Encoded Mask	Regional nodes related to IPAdapter

Some nodes depend on Impact Pack or IPAdapter Plus. Check the README dependencies before using them.

This route fits workflows that already have a color-block mask, such as red on the left and blue on the right, and where you want to avoid manual pixel calculations.

Combination boundaries and second-pass risks

Combining with ControlNet, IPAdapter, and LoRA

Tool	What it controls	What it does not solve
Regional prompt (Set Area / Regional)	Position and prompt assignment	Identity consistency; attribute bleed, which Set Area does not solve
ControlNet (OpenPose / Canny / Depth)	Pose and structure	Attribute bleed; positional assignment
IPAdapter	Reference-image influence over style or composition	Regional isolation; fixed identity, because IPAdapter influences the overall style
FaceID / LoRA	Identity or style consistency	Regional prompts; attribute bleed

The practical sequence is regional prompts for position → OpenPose for pose → low-strength global conditioning for a shared atmosphere. You must align OpenPose, ControlNet, or IPAdapter positions with the regions manually; they cannot automatically bind “the person on the left.”

Color bleed during a second pass, upscale, or inpaint

If a second pass, upscale, or face-repair step omits regional conditioning, the model may pursue global consistency again and mix hair or clothing colors that were separated in the first stage. The area-composition example warns that hair colors can merge when the second pass has no regional prompt.

Decide whether later passes should continue using regional conditioning or whether local Inpaint is a better fix. If the first pass used Set Area, consider setting the regions again during upscaling so the model does not restore global consistency at the expense of subject separation.

Next steps and related reading

Regional prompting is one part of multi-subject composition. Pose control, identity consistency, reference-image guidance, and high-resolution repair need other tools.

Published:

• Import workflows and install missing nodes: fixes custom-node installation and missing-node issues
• Stable Diffusion prompt template guide: prompt-writing basics for cases where one prompt cannot describe multiple subjects clearly
• ComfyUI ControlNet guide: controls pose and structure alongside regional prompt positioning
• ComfyUI LoRA guide: keeps characters or styles consistent and distinguishes multi-LoRA bleed from regional-prompt bleed

Separate guides will cover IPAdapter, FaceID and face consistency, LoRA training, high-resolution fixes, and Inpainting. Combining those tools with regional prompts can further reduce uncertainty in multi-subject scenes.