Save 4 hours! We auto-setup your workflow! Free!

Drop your workflow.json — we handle every dependency, custom node, and model. Just open the link and run.

Auto-Setup Workflow Json (Free) Now!
ComfyUI > Nodes > ComfyUI-Krea2-Regional

ComfyUI Extension: ComfyUI-Krea2-Regional

Repo Name

ComfyUI-Krea2-Regional

Author
januspluto (Account age: 3193 days)
Nodes
View all nodes(1)
Latest Updated
2026-11-05
Github Stars
0.02K

How to Install ComfyUI-Krea2-Regional

Install this extension via the ComfyUI Manager by searching for ComfyUI-Krea2-Regional
  • 1. Click the Manager button in the main menu
  • 2. Select Custom Nodes Manager button
  • 3. Enter ComfyUI-Krea2-Regional in the search bar
After installation, click the Restart button to restart ComfyUI. Then, manually refresh your browser to clear the cache and access the updated list of nodes.

Visit ComfyUI Online for ready-to-use ComfyUI environment

  • Free trial available
  • 16GB VRAM to 80GB VRAM GPU machines
  • 400+ preloaded models/nodes
  • Freedom to upload custom models/nodes
  • 200+ ready-to-run workflows
  • 100% private workspace with up to 200GB storage
  • Dedicated Support

Run ComfyUI Online

ComfyUI-Krea2-Regional Description

ComfyUI-Krea2-Regional lets you draw image regions, assign each a prompt and LoRA, and generate a cohesive scene in one model pass per sampling step, with controls to reduce style bleed.

ComfyUI-Krea2-Regional Introduction

ComfyUI-Krea2-Regional helps you create images with different prompts and LoRAs in different parts of the same composition. You can draw regions on a canvas, describe what belongs in each one, and assign a character or style LoRA to a specific region. The extension then generates the image in a single model pass per sampling step, rather than generating each area separately and blending the results.

This is useful for scenes with multiple characters, distinct objects, or areas that need different visual treatments. For example, you could place an armored knight on the left and a wizard on the right, give each a separate character LoRA, and keep a shared description for the background and lighting.

Regional control is not an absolute boundary. Nearby styles or character details can still influence one another, particularly where regions overlap. The extension provides mask, attention, and optional detailer controls to help reduce that bleed while keeping the overall image cohesive.

Example ComfyUI workflow

How ComfyUI-Krea2-Regional Works

Think of the canvas as a rough stage plan. You mark where subjects should go and write a description for each marked area. A base prompt describes what the whole scene has in common, such as its setting, mood, and lighting.

During generation, the extension keeps the regional descriptions together but guides each part of the developing image toward its assigned description. It can also limit a regional LoRA's influence to the tokens associated with that region. This lets a character LoRA help shape one subject without intentionally applying it across the entire picture.

The boxes and lassos are guides, not guarantees that a subject will land at an exact pixel position. Two controls help when the generated composition differs from your drawing:

  • Layout hints tell the model where the subjects belong in words, such as “an armored knight on the left.”
  • Adaptive masks can use the model's early attention patterns to adjust a region toward where its subject actually appears.

The base description remains shared, which helps the image feel like one scene. It also means some influence can cross a region boundary. For especially stubborn identity bleed, the optional Regional Detailer can make a second pass over detected subjects after the initial image is generated.

ComfyUI-Krea2-Regional Features

Canvas-based regional editing

Krea2 Regional Builder (canvas) is the main visual node. Draw rectangular or freehand-lasso regions, write a prompt for each, and choose per-region LoRAs from a searchable list. You can also enter a base description, background, and style.

The canvas includes a grid, snapping, guides, a pop-out editor, and a reference background to help with placement. Lasso shapes remain editable: move their points, split an edge by dragging its midpoint, Alt-click a point to remove it, or resize the whole shape with its handles. You can also convert a rectangle to a lasso or back again.

The order of regions matters when they overlap. The region list represents front-to-back depth; use its up and down controls to change which region is in front. An empty region prompt means that region is not used when the builder creates its output, even if its shape is visible on the canvas.

A canvas background is only a drawing aid. It is not automatically supplied to the image-generation model:

  • Connect an image to the builder's image input to show it after the graph runs once.
  • Use Grab BG to display the most recently generated image from ComfyUI's queue history.

Applying the regions

Krea2 Apply Regional takes the builder's regions and base information and returns a regional MODEL and combined CONDITIONING. In a typical workflow, its MODEL goes to the KSampler model input, while its CONDITIONING goes to KSampler positive.

The controls most likely to affect your image are:

SettingWhat it changesWhen to try it
adaptive_masksAdjusts masks based on where subjects appear during early generation. The default, refine boxes, keeps the search near your drawn regions; free allows more movement.A subject is near its box but the regional effect does not follow its shape.
layout_in_baseAdds placement information to the shared prompt. The default uses position hints; full JSON supplies a more structured layout.Subjects repeatedly appear in the wrong places.
exclusive_masksGives overlapping areas to the strongest region; a tie goes to the front region. Off by default.Two regional effects compete in the same area.
restrict_img_attnLimits image-to-image influence across regions. Off by default.Character features bleed strongly between neighboring subjects.
restrict_end_percentSets how long that attention restriction remains active.You want early separation but softer, more unified lighting and edges later.
base_loras_exclude_regionsKeeps base LoRAs out of subject regions. Off by default.A scene-wide style LoRA is changing regional character identities.
region_lock_strength, region_lock_start, region_lock_endControl identity anchoring over part of the sampling process.A regional subject loses its identity as generation progresses.
unmaskable_layersChooses what happens to LoRA layers that cannot be limited by image position: apply globally or skip. The default is apply globally.You need to decide whether those particular LoRA effects may influence the whole image.
dynamic_vram_modeControls model management for regional sampling. The default is regional safe.Leave it at the default unless you are deliberately testing the older DynamicVRAM behavior or a build with the upstream issue fixed.

For tightly packed subjects, keep mask growth and feathering modest and leave some space between drawn regions. Strong attention restriction may improve separation but can make the result look assembled from separate pieces. A restrict_end_percent around 0.4–0.6 is one way to isolate subjects early, then let later steps bring the scene together.

The default regional safe DynamicVRAM mode addresses a reported host-buffer accumulation issue after repeated Krea 2 restaging. It changes management of the regional model clone for sampling; it does not turn off DynamicVRAM throughout ComfyUI.

Regional LoRAs

Regional and base LoRAs are selected from models/loras. The extension supports the PEFT/diffusers, kohya, and supported LoKr formats described in the original documentation. Some tucker or convolutional LoKr variants are skipped with a warning. Global LoRAs applied with regular ComfyUI LoRA loaders can still be used alongside regional LoRAs.

A regional LoRA influences only part of the image, so a strength that works globally may look weak in a small region. The author suggests trying approximately 1.3–1.6 for regional identity LoRAs rather than assuming a global strength near 1.0 will have the same effect. Treat that as a starting point and adjust to the image.

Optional second-pass detailing

Krea 2 - Regional Detailer can improve likeness or identity after the first image has been generated. Instead of relying only on the original box, it can use detected subject masks to find where a subject actually appeared, work on that subject, and composite the result back through its silhouette. Regions without a matched detection remain untouched.

Connect a decoded image and subject masks, or detector SEGS, along with the base model, VAE, builder regions, and builder base_conditioning. Use the builder's conditioning here, not the combined conditioning from Apply Regional. The original documentation mentions SAM3 person detection and Impact Pack's Ultralytics person or face detectors as possible sources of detections.

For a gentle identity correction, the documented starting range for denoise is 0.12–0.18; the stated default is 0.18. Higher values around 0.2–0.25 push likeness more strongly, while 0.35 or above can start changing the composition inside the selected subject. The recommended detail_target setting is auto, which varies the detail area with subject size. A partial-denoise pass may look smoother at 3–5 steps than at eight.

Keep shift at model default for Krea 2 Turbo so the detail pass inherits the first pass's schedule. The per-crop auto shift mode is intended for RAW. Without a detector, the Detailer falls back to region boxes; detection is preferable when subjects overlap.

Other nodes and workflow options

  • Krea 2 Empty Latent Image creates a Krea 2-sized starting latent. Its aspect-ratio presets and megapixel control also provide width and height values that can be connected to the builder.
  • Krea2 Regional LoRA and Krea2 Regional Prompt let you assemble regions through node chains instead of using the canvas.
  • Krea2 Regions from Ideogram JSON converts compatible caption JSON into regions without a canvas. The canvas builder can import the same kind of JSON and then let you edit it.
  • The builder's Caption button can run a connected captioner and import its layout JSON without running a full Krea 2 generation. The supplied captioner prompt and wiring notes describe a setup using the Qwen3.5 VL text encoder already loaded for Krea 2.

You can load the supplied example workflows by dragging a workflow .json file onto the ComfyUI canvas.

ComfyUI-Krea2-Regional Models

This extension works with Krea 2; the provided documentation does not describe a separate set of downloadable models for the extension. It does distinguish between Krea 2 Turbo and RAW when setting up the Detailer: Turbo should keep shift at model default, while the per-crop auto option is for RAW.

The extension also works with ComfyUI's native Krea 2 reference-latent path. In this mode, reference content acts as shared context available to all target regions. Regional character LoRAs are not applied to the reference tokens. This compatibility does not mean that a separate reference image can be assigned to each region.

What's New with ComfyUI-Krea2-Regional

The supplied documentation identifies release 1.2.0 as tested against ComfyUI 0.29–0.30, but does not provide a dated, version-by-version changelog. The following capabilities are described as additions or current behavior; their exact release versions are not specified:

  • Regional-safe DynamicVRAM management is the default to work around the reported repeated-restaging host-buffer issue without disabling DynamicVRAM globally.
  • Adaptive mask routing and base-prompt layout hints help regional effects follow generated subjects and improve their placement.
  • Regional Detailer auto targeting can choose a suitable detail area based on subject size instead of treating every region as another full regeneration.
  • Native Krea 2 reference-latent compatibility allows references to remain shared context while keeping regional LoRAs off reference tokens.
  • Video VAE support, including WanVAE, is described as working through the regional workflow and Detailer.

Troubleshooting ComfyUI-Krea2-Regional

A drawn region has no effect.

  1. Check that its region prompt... field contains a description. A shape with no prompt is skipped.
  2. Look for the “no prompt — not used” label in the region list.
  3. Confirm that the builder's regions output is connected to Krea2 Apply Regional.

A reference image does not appear behind the canvas.

  1. If you connected the builder's image input, run the graph once. The preview arrives with the node's execution output.
  2. If you want to draw over a recent result instead, use Grab BG after generating an image.
  3. Remember that the canvas background is a placement aid, not image input for generation.

Subjects do not land inside their boxes.

  1. Start with the builder's layout_in_base setting at position hints so the shared prompt describes each subject's location.
  2. Try refine boxes under adaptive_masks to let masks follow subjects near your drawn areas.
  3. Leave enough room between subjects and avoid excessive mask growth in a crowded layout.

Characters or LoRA styles bleed into one another.

  1. Try adaptive_masks first, especially if a subject's actual silhouette differs from its box.
  2. Turn on exclusive_masks if regions overlap.
  3. If stronger separation is needed, try restrict_img_attn. Consider ending that restriction partway through sampling to preserve softer transitions.
  4. If a base style LoRA is affecting character regions, try base_loras_exclude_regions.
  5. For remaining likeness problems, use the Regional Detailer with subject detections.

The Detailer changes too much of the image.

  1. Lower denoise toward the documented 0.12–0.18 range.
  2. Use detail_target = auto and consider 3–5 steps for a partial-denoise pass.
  3. Connect a subject detector so the Detailer can work from silhouettes rather than broad region boxes.
  4. For Turbo, leave shift at model default.

The canvas looks wrong in ComfyUI.

If you are using the experimental Nodes 2.0 rendering mode, try disabling it and viewing the builder in ComfyUI's standard graph renderer.

ComfyUI reports a host-buffer error after repeated runs.

Check that Apply Regional's dynamic_vram_mode remains at its default, regional safe. The keep DynamicVRAM choice retains the older regional behavior and is intended for deliberate testing.

Frequently Asked Questions

Does each region require its own generation pass?
No. The main regional generation uses one model pass per sampling step. The optional Detailer is a separate pass after generation.

Can I use my usual global LoRA as well as regional LoRAs?
Yes. The documented workflow supports global ComfyUI LoRA weight patches alongside regional LoRAs.

Will my canvas background appear in the generated image?
No. Displaying an image behind the regions does not, by itself, send that image to the generation model.

Do native reference latents belong to individual regions?
No. In the documented compatibility path, native references are shared context for the target regions.

What happens when regions overlap?
Their order determines which is in front. With exclusive_masks enabled, the strongest region claims an overlapping token, and an exact tie goes to the front region.

Learn More about ComfyUI-Krea2-Regional

  • Example workflows show how the nodes connect in ComfyUI.
  • Captioner prompt and wiring notes explain the documented image-to-region captioning setup.
  • Architecture notes offer a deeper explanation of regional attention, LoRA routing, and caching for readers who want the technical details.
  • The FreeFuse paper provides background on the adaptive-routing approach referenced by the extension.

ComfyUI-Krea2-Regional Related Nodes

RunComfy
Copyright 2025 RunComfy. All Rights Reserved.

RunComfy is the premier ComfyUI platform, offering ComfyUI online environment and services, along with ComfyUI workflows featuring stunning visuals. RunComfy also provides AI Models, enabling artists to harness the latest AI tools to create incredible art.

ComfyUI-Krea2-Regional detailed guide | ComfyUI