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 > Krea 2 Reference

ComfyUI Extension: Krea 2 Reference

Repo Name

ComfyUI-Krea2-Reference

Author
konokoaz (Account age: 1342 days)
Nodes
View all nodes(1)
Latest Updated
2026-07-15
Github Stars
0.04K

How to Install Krea 2 Reference

Install this extension via the ComfyUI Manager by searching for Krea 2 Reference
  • 1. Click the Manager button in the main menu
  • 2. Select Custom Nodes Manager button
  • 3. Enter Krea 2 Reference 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

Krea 2 Reference Description

Krea 2 Reference adds ComfyUI conditioning that combines reference images with text prompts to preserve character identity, transfer visual style, and guide new scenes without copying layouts.

ComfyUI-Krea2-Reference Introduction

ComfyUI-Krea2-Reference adds reference-image conditioning for Krea 2 workflows in ComfyUI. It lets you guide an image generation with one or more reference images while still allowing your text prompt to control the scene, pose, composition, and story.

A simple example is:

“The character from Picture 1, in the exact art style of Picture 2, riding a bicycle through a rainy city.”

The first image can provide the character’s identity, while the second image provides the visual style. The result is designed to preserve the character and style information without simply copying the reference images into a collage.

The extension is useful when you want to:

  • Keep the same character recognizable across multiple images.
  • Transfer an art style from one image to another.
  • Combine a character reference with a separate style reference.
  • Create comic pages or multi-panel scenes with a consistent subject.
  • Use text to describe a scene while reference images provide visual identity.
  • Avoid the common problem where a reference image overwhelms the prompt and produces a character sheet, collage, or unrelated composition.
  • Control whether a reference is treated as a character, a style, or a general image reference.
  • Use up to four reference images in a single conditioning setup.

The extension uses Krea 2’s vision-language text encoder and a dedicated reference LoRA. This allows the reference information to be included as visual tokens that the model can interpret alongside the written prompt.

How ComfyUI-Krea2-Reference Works

Reference Images Become Part of the Prompt

In a normal text-to-image workflow, the model receives written instructions such as:

“A girl walking through a forest in a watercolor style.”

With this extension, images can be included as named visual references. The model can interpret instructions such as:

“The character from Picture 1, in the exact art style of Picture 2, walking through a forest.”

The images are not used only as pixels to copy. They are converted into visual information that the Krea 2 vision-language encoder can understand and combine with the written prompt.

A useful way to think about this is that the prompt becomes a small visual briefing:

  • Picture 1 tells the model which image contains the character.
  • Picture 2 tells the model which image contains the style.
  • The rest of the prompt describes what the character should do and where the scene should take place.

Character and Style References Have Different Jobs

A character reference is used to preserve identity details such as:

  • Face and hairstyle.
  • Clothing.
  • Accessories.
  • Distinctive colors.
  • Body shape.
  • Recognizable features.

A style reference is used to influence details such as:

  • Brushwork.
  • Line quality.
  • Color treatment.
  • Lighting.
  • Rendering method.
  • Historical or artistic appearance.

For example:

“The character from Picture 1, in the exact art style of Picture 2, standing beside a train station.”

Picture 1 should show the subject clearly. Picture 2 should show the desired visual style, preferably without an unrelated subject that you want the model to reproduce.

Prompt Words Assign Reference Roles

The extension uses the words in your prompt to determine how each image should be interpreted.

Use terms such as:

  • character1, character2, character3, or character4 for identity references.
  • style1, style2, style3, or style4 for style references.
  • image1, image2, image3, or image4 for general image references.
  • Picture 1, Picture 2, and so on as a more natural alternative.

For example:

The character1 holding an umbrella in the style2, standing on a city street at night.

Here, the first image is treated as the character reference and the second image is treated as the style reference.

If an image is connected to the node but is not mentioned in the prompt, the node automatically adds a character-style reference clause for it. Explicitly naming the role is recommended because it gives you more control.

If a reference is described as both a character and a style, the character role takes priority.

The Reference LoRA Helps the Model Follow the Scene

The reference LoRA is designed to improve how Krea 2 follows reference instructions. Without it, the model may focus too strongly on the reference and produce:

  • Several copies of the same character.
  • A character sheet instead of the requested scene.
  • A collage that resembles the layout of the reference.
  • A scene where the written action is ignored.

The LoRA helps separate the roles of the reference and the prompt. The reference supplies identity or style information, while the prompt remains responsible for the scene.

Reference Information Is Not the Same as img2img

This extension does not work like traditional img2img. It does not simply denoise an existing image toward a new result. Instead, it provides reference information through Krea 2’s vision-language conditioning.

This means you can use a reference image while asking for a completely different:

  • Pose.
  • Background.
  • Camera angle.
  • Composition.
  • Activity.
  • Art style.

For exact pose control, the original readme recommends using a pose image as the initial latent and setting the denoise strength around 0.70–0.78.

ComfyUI-Krea2-Reference Features

Krea 2 Reference Encode

The TextEncodeKrea2Reference node combines your written prompt with up to four reference images.

It provides:

  • Text prompt input.
  • Up to four reference-image inputs.
  • Per-image detail settings.
  • Optional masks.
  • Optional background removal.
  • Optional layout splitting.

The output is intended to connect to the Krea 2 sampling workflow in place of a regular text-conditioning node.

Character References

Use a character reference when you want the generated image to preserve a subject’s identity.

Example:

The character1 wearing a raincoat and walking through a forest.

A clear, single-subject image usually works better than a character sheet. Character-sheet references can cause the model to generate multiple views or multiple copies of the subject.

Character reference details are especially important when the subject has:

  • A distinctive hairstyle.
  • A specific clothing design.
  • Important accessories.
  • Unusual colors or markings.
  • A recognizable face.

Style References

Use a style reference when you want to borrow the visual language of an image without reproducing its subject or composition.

Example:

The character1 walking through a city in the exact art style of style2.

Style references can influence:

  • Color palette.
  • Brush or line style.
  • Shading.
  • Texture.
  • Rendering style.
  • Overall artistic mood.

If the style is strongly associated with a particular franchise, the reference may also pull recognizable characters or objects from that franchise into the result. A more general style image is usually easier to control.

You can also reinforce a style with text:

The character1 in the style2, rendered with flat watercolor washes, soft paper texture, and delicate ink outlines.

This is useful when you want the style reference to affect the character as well as the background and overall scene.

General Image References

The image1 through image4 and Picture 1 through Picture 4 names can be used when you do not want to describe an image strictly as a character or style reference.

For example:

Create a poster based on Picture 1, with the subject from Picture 2 in the foreground.

For predictable results, use characterN and styleN when the role is clear.

Image Detail

Each reference has an imageN_detail setting. The available levels are:

Detail settingApproximate reference sizeRecommended use
low256 pxBroad style, color, or pose hints
normal512 pxGeneral-purpose reference use
high1024 pxCharacter identity and clothing details
max1280 pxVery detailed references when those details matter

Higher detail gives the model more visual information, but it can also require more memory and processing time.

Recommended starting points:

  • Use high for a character with important facial, clothing, or accessory details.
  • Use low or normal for a style reference.
  • Use low for a simple pose or composition hint.
  • Increase detail only when the result is missing information that is clearly visible in the reference.

Image Masks

The imageN_mask input allows you to focus on a selected region of the reference.

The node crops the reference to the mask bounding box and fills the surrounding area with white before encoding it. This is useful when:

  • The subject occupies only a small portion of the image.
  • You want to remove distracting background content.
  • A reference contains several objects but only one should be used.
  • You want to focus on a face, outfit, prop, or other specific area.

A tight mask can make the subject clearer. However, masking away important context may remove useful information such as the subject’s full clothing or overall silhouette.

Background Removal

The bg_removal input can be connected to a LoadBackgroundRemovalModel node. This allows the extension to automatically clean the background from a reference before encoding it.

Background removal can help when:

  • The background contains strong shapes that compete with the subject.
  • You are using a product, character, or object reference.
  • The subject should be separated from its original environment.
  • You want to reduce unwanted scene-layout copying.

Background removal is not always necessary. For style references, the original background may contain the artistic qualities you want to preserve.

Layout Split

The layout_split option controls how the reference is used during the early part of sampling.

The first fraction of the sampling steps is conditioned on text only. This gives the model time to establish the requested scene before reference information is introduced.

A value around 0.25 can help reduce unwanted collage-like layouts in zero-shot use. It is generally unnecessary when the reference LoRA is enabled and working correctly.

Use layout splitting when:

  • The output copies the reference composition too closely.
  • The image becomes a collage or character sheet.
  • The written scene is being ignored in favor of the reference layout.

Layered Reference LoRA

The Krea2ReferenceLoraLoader loads krea2_reference_v1.safetensors and provides separate strength controls for different parts of the LoRA.

The recommended starting values are:

strength_text_fusion: 1.0
strength_dit_body: 0.0

This configuration enables the reference addressing, anti-bleed, and style-transfer behavior while leaving the base model’s image rendering mostly unchanged.

You can increase strength_dit_body to approximately 0.3 when a difficult prompt needs stronger scene adherence. Higher values may introduce more identity drift because the DiT portion of the LoRA can pull the result toward the LoRA’s training distribution.

A regular LoraLoaderModelOnly can also be used with a strength around 0.7–1.4. This provides a single strength control, but higher strengths may cause more identity drift than the layered setup.

Color and Identity Control

The colors in a character reference are treated as part of the character’s identity. If the reference character wears a red dress, the model may continue to preserve that red dress even when the text prompt requests a different color.

To change a color or make the character monochrome, change the reference itself. For example:

  • Use a grayscale version of the character reference.
  • Add text such as “the entire image is black and white.”
  • Use the modified reference with the same character description.

The original readme notes that increasing strength_dit_body does not solve this color issue. It may instead pull the image toward the colorful distribution learned by the LoRA.

Multi-Panel and Comic Images

The extension can be used for multi-panel images and comics. A consistent character reference can help preserve details such as:

  • Hair and accessories.
  • Clothing.
  • Freckles.
  • Jewelry.
  • Character colors.
  • Recognizable facial features.

For comic layouts, the recommended layered LoRA setup is particularly useful because it helps the written scene and panel instructions remain in control.

For wide canvases, watch for duplicated figures. A 1024 × 1024 canvas is described in the original readme as a reliable starting point.

ComfyUI-Krea2-Reference Models

Krea 2 Checkpoint

The Krea 2 checkpoint provides the main image-generation model. In ComfyUI, it is loaded with UNETLoader.

This model creates the final image and interprets the text and reference conditioning supplied by the extension.

Qwen3-VL-4B Text Encoder

Krea 2 uses the Qwen3-VL-4B vision-language model as its text encoder. In ComfyUI, it is loaded with CLIPLoader using the krea2 type.

This encoder is responsible for understanding both:

  • The written prompt.
  • The visual information from the reference images.

The original readme specifies that the encoder should use the bf16 version. Using an fp8 version can break the vision tower, which prevents the reference images from being interpreted correctly.

Qwen Image VAE

The qwen_image_vae is loaded with VAELoader. It converts between the model’s internal image representation and the visible image used by ComfyUI.

It is part of the Krea 2 workflow but does not determine the character or style role of the references by itself.

Krea 2 Reference LoRA v1

The reference LoRA is provided as krea2_reference_v1.safetensors.

It was trained on approximately 2,700 reference-following triplets, including character and style relationships. Its main purpose is to improve:

  • Reference addressing such as “the character from Picture 1.”
  • Scene-text adherence.
  • Character and style separation.
  • Anti-collage behavior.
  • Style transfer without copying the original layout.
  • Normal text-to-image behavior when no references are used.

The recommended layered setup is:

strength_text_fusion: 1.0
strength_dit_body: 0.0

Use a higher strength_dit_body only when the prompt needs additional scene adherence and the resulting identity changes are acceptable.

Model Selection and Performance

The original context does not describe multiple versions of the Krea 2 checkpoint or multiple reference LoRA releases. The main model choices are therefore the base Krea 2 model, the Qwen3-VL-4B encoder, the Qwen image VAE, and the reference LoRA.

Higher image-reference detail and multiple references generally require more memory and may take longer to process. Start with one clear character reference and one style reference before adding more images.

ComfyUI-Krea2-Reference Examples

The extension includes example images whose PNG files contain complete ComfyUI workflow information.

You can drag and drop these files into ComfyUI:

The example workflow uses the following general structure:

Load Image
Load Image
Krea 2 Reference Encode
Layered Reference LoRA
KSampler

The examples were generated with Krea 2 Turbo using 8 steps and CFG 1.0.

Troubleshooting ComfyUI-Krea2-Reference

The result looks like a collage or character sheet

This usually means the reference layout is being copied too strongly or the scene prompt is not being followed clearly.

Try the following:

  1. Use a single-pose character reference instead of a character sheet.
  2. Describe the desired action and environment explicitly.
  3. Name the roles in the prompt, for example:
    The character1 walking through a market in the style2.
    
  4. Set the layered LoRA to:
    strength_text_fusion: 1.0
    strength_dit_body: 0.0
    
  5. Try a layout_split value around 0.25 if the issue continues.
  6. Use a 1024 × 1024 canvas instead of an unusually wide layout.

The model ignores the requested scene

The reference may be too dominant, or the prompt may not clearly explain what should happen.

Try the following:

  1. State the action, location, and composition directly.
  2. Use explicit reference names such as character1 and style2.
  3. Avoid attaching several images that are not needed.
  4. Reduce the detail level of style references.
  5. Keep strength_dit_body at 0.0 before testing higher values.
  6. If the prompt is especially difficult, try increasing strength_dit_body gradually toward 0.3.

The wrong image is treated as the character

The extension uses the wording in the prompt to assign roles.

Check that:

  1. The character reference is connected to the expected image input.
  2. The prompt uses the correct number, such as character1 or Picture 1.
  3. Style references are labeled separately, such as style2.
  4. You are not describing the same image as both a character and a style. If both roles are used, the character role takes priority.

The character’s colors do not change

The reference image’s colors are considered part of the character’s identity. Text alone may not be enough to replace them.

Try the following:

  1. Edit or convert the character reference to the desired color treatment.
  2. Use a grayscale reference for a monochrome result.
  3. Reinforce the change in the prompt, for example:
    The entire image is black and white.
    
  4. Do not expect strength_dit_body to solve this issue. Increasing it can instead pull the result toward the LoRA’s colorful training distribution.

The style reference affects the wrong parts of the image

A style reference may contain a recognizable subject, franchise character, or strong composition. The model can sometimes transfer more than the intended artistic style.

Try the following:

  1. Use a style image with a simple or unrelated subject.
  2. Choose an image where the brushwork, lighting, and rendering are more important than the depicted character.
  3. Describe the desired style in text as well as providing the image.
  4. Lower the style image detail setting to low or normal.
  5. Use a mask if only a particular area of the style image should be examined.

The reference details are missing

The image may be encoded at too low a detail level.

Try the following:

  1. Set imageN_detail to high.
  2. Use max only when very small details are essential.
  3. Crop or mask the reference so the subject occupies more of the image.
  4. Use a clear, well-lit reference with the important features visible.
  5. Avoid adding several low-quality references that conflict with one another.

The model generates duplicate characters

Duplicate figures are more likely with:

  • Character sheets.
  • Wide canvases.
  • References containing several views of the same subject.
  • Prompts requesting multiple poses without clear panel or composition instructions.

Try the following:

  1. Use one clear, single-pose reference.
  2. Start with a 1024 × 1024 canvas.
  3. Describe the intended number and placement of subjects.
  4. Use layout_split if the reference composition is being copied.
  5. For comics, describe the panel layout directly and use a consistent character reference.

CFG or the negative prompt appears to have no effect

With Krea 2 Turbo at CFG 1.0, the negative prompt is skipped. This is expected behavior rather than an error.

You can test a different CFG or sampling configuration if your workflow supports it, but the original examples use Turbo with 8 steps and CFG 1.0.

The vision reference is not being read

Check the text encoder configuration:

  1. Confirm that the Qwen3-VL-4B encoder is loaded through CLIPLoader.
  2. Confirm that the loader type is krea2.
  3. Use the bf16 encoder.
  4. Avoid the fp8 encoder, which can break the vision tower.
  5. Verify that the reference image is connected to the correct input on TextEncodeKrea2Reference.

Frequently Asked Questions

Can I use only one reference image?

Yes. A single image can be used as a character reference, a style reference, or a general visual reference. You can also describe the style entirely with text while using the image only for the character.

Can I use both a character and a style image?

Yes. Connect the character image as one reference and the style image as another. Then identify them in the prompt, for example:

The character1 in the exact art style of style2, sitting in a cafe.

How many reference images can I use?

The reference encode node supports up to four reference images.

Do I have to use the reference LoRA?

The node can provide reference conditioning, but the reference LoRA is designed to improve reference addressing, scene adherence, anti-bleed behavior, and style transfer. The recommended setup uses the layered reference LoRA.

Should I use a character sheet as a reference?

Usually, no. Character sheets can cause the model to generate multiple views or duplicated subjects. A single clear pose is generally easier to control.

Can I change the character’s clothing color with text?

Not reliably. Reference colors are treated as part of identity. Edit the reference itself when a specific color change is important.

Can I use this for exact pose matching?

The extension is primarily for reference-based identity and style control. For a more exact pose, use a pose image as the initial latent and try a denoise value around 0.70–0.78.

Can I use a style reference without copying its composition?

Yes. Describe the desired scene clearly, use the image as a style reference, and consider reducing its detail level. The reference LoRA is intended to separate style from layout.

Why does increasing the LoRA strength change the character?

The DiT portion of the LoRA can introduce identity drift at higher strengths. Start with strength_text_fusion: 1.0 and strength_dit_body: 0.0, then increase the DiT strength only when necessary.

What's New with ComfyUI-Krea2-Reference

The provided context does not include a version history or release notes. No specific updates can be listed reliably without inventing information.

Learn More about ComfyUI-Krea2-Reference

The most useful learning resources included with the extension are the embedded workflow examples in the examples/ directory. Because each PNG contains its ComfyUI graph, you can inspect a working setup by dragging an example into ComfyUI and tracing how the nodes are connected.

Useful starting points include:

When experimenting, change one variable at a time:

  1. Start with one clear character reference.
  2. Confirm that character1 is understood.
  3. Add a style reference as style2.
  4. Adjust image detail only after the roles work correctly.
  5. Test layout_split if the output copies the reference layout.
  6. Adjust strength_dit_body only after trying the recommended layered LoRA settings.

The original extension repository is available at ComfyUI-Krea2-Reference on GitHub.

Krea 2 Reference 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.

Krea 2 Reference detailed guide | ComfyUI