Krea 2 | Six-Model Benchmark Suite for ComfyUI
Compare six different Krea 2 models in a single workflow, explore how they interact with LoRAs, and evaluate their generation performance.
Hello Everyone!
I'm excited to share a workflow that I genuinely hope you'll find useful!
Have you ever struggled to decide which model is best suited for the scene you're trying to create? Have you spent hours running endless tests just to see how a particular model responds to your favorite LoRA?
If that sounds familiar, this workflow might be exactly what you're looking for.
Krea 2 Six-Model Benchmark Suite was designed to let you compare six Krea 2 models within a single workspace. You can examine how different models interpret the same prompt, how they respond to LoRAs, the visual quality of their outputs, and their generation performance.
And the best part? You don't need to build a separate workflow for every model.
A Small Warning Before We Begin
As a general rule, green-colored nodes are intended for user interaction and configuration. Feel free to adjust their settings according to your needs.
However, I strongly recommend leaving the red-colored nodes alone.
Why?
Because even I sometimes spend hours fixing things after making what seemed like a tiny change to one of those red nodes!
In short, leaving the red nodes alone might make life easier for both of us.
1. Getting Started: Decision Panel
Start by opening the Decision Panel.
The first thing we need is a prompt for image generation. Fortunately, you have several options.
Option A — Generate a Prompt from an Image
If you're feeling a little lazy sometimes, just like me, you can upload an image you like and use it as your prompt source.
The workflow uses a Local LLM to analyze the uploaded image and help generate a prompt suitable for Krea 2.
Option B — Just Write Whatever Comes to Mind
You can also describe your scene in your own words, without worrying about prompt-writing rules.
Your description can be messy, incomplete, unstructured, or written in everyday language. That's perfectly fine.
The (Deno) Local LLM Refiner (DenoLocalLLMRefiner) node can transform your text into a more structured prompt suitable for image generation.
Better yet, if your Local LLM supports multiple languages and can translate them into English, you may even write your description in your native language.
However, there's one thing I should mention: having some familiarity with Local LLMs will make your experience much smoother.
Choosing an unsuitable model or using inefficient memory and server settings may significantly increase the time required to generate your prompt.
Option C — Use Your Own Prompt Directly
Maybe you're already great at writing prompts. Or perhaps you have a beautifully structured prompt specifically designed for Krea 2.
No problem!
You can bypass the Local LLM Refiner nodes entirely and start image generation using your own prompt.
Using a Local LLM is completely optional.
Where Can You See Your Final Prompt?
You can view the final version of the prompt that will be used for image generation in the red-colored node labeled "YOUR FINAL (ACTIVE) PROMPT IS HERE", located near the top of the workflow.
This allows you to review the exact prompt being sent to the image-generation pipeline, whether it was generated by a Local LLM or provided directly by you.
2. Krea 2 | Model Selector
The Model Selector allows you to choose six Krea 2 models for comparison.
You can select six different models, or use the same model in multiple slots.
For example, in some of my experiments, I assigned one model to the first three slots and another model to the last three slots. This allowed me to compare two models using different sampling configurations in the same workflow.
I found this especially useful for understanding how models respond to changes in generation parameters.
In other words, you're not limited to comparing different models.
You can also compare the same model under different settings.
3. LoRA Selection — Pixaroma LoRA Loader
I've included the Pixaroma LoRA Loader, which allows you to select multiple LoRAs.
One feature I particularly like is its ability to include LoRA trigger words in the prompt pipeline.
Another useful feature is its integration with Civitai, which helps you access a LoRA's model page, review recommended settings, and check trigger words when needed.
This can be incredibly convenient if you regularly work with multiple LoRAs.
The workflow is designed to apply the selected LoRA or LoRAs to all six models.
However, please remember:
Not every LoRA will behave the same way with every Krea 2 model.
A LoRA may have a strong effect on one model while producing weaker, different, or unexpected results with another.
If you notice significant differences, I recommend checking model–LoRA compatibility through independent tests before assuming there's something wrong with the workflow.
4. Image Dimensions — AusBoss Latent Size
For a meaningful comparison, it's important to generate images using consistent dimensions.
That's why I've included the AusBoss Latent Size node.
The aspect ratio and image dimensions you select here are applied to all six models.
Higher resolutions generally require more processing time and memory.
Personally, I frequently experiment with different aspect ratios, but I usually keep the image dimensions at moderate levels.
This helps me evaluate model behavior more quickly without spending too much time on every test.
5. Seed Control
Using a fixed seed can be very helpful when comparing different models.
For this reason, the Decision Panel includes a centralized seed control that sends the same seed value to all KSamplers.
You can use a fixed seed to examine how different models behave with the same initial seed value.
Of course, using the same seed does not guarantee identical compositions across different models.
However, keeping as many variables as possible consistent can make your comparisons more meaningful.
6. Batch Size — How Many Images?
The Decision Panel also includes a Batch Size control.
This allows you to specify how many images each model should generate per batch.
On my own system, equipped with an RTX 4090 Laptop GPU (16 GiB VRAM) and 64 GiB RAM, I can generally use a maximum batch size of 2, depending on the model and settings.
Your actual limit may vary depending on model size, resolution, and memory management.
Feel free to increase the batch size as much as your GPU and system memory can handle!
However, I recommend starting with Batch Size = 1, especially during your first tests.
7. Private Panels — Independent Settings for Each Model
Just below the Decision Panel, you'll find six Private Panels, one for each model.
These panels correspond directly to the order of models in the Model Selector.
Model Selector Slot 1 → Private Panel 1
Model Selector Slot 2 → Private Panel 2
Model Selector Slot 3 → Private Panel 3
Model Selector Slot 4 → Private Panel 4
Model Selector Slot 5 → Private Panel 5
Model Selector Slot 6 → Private Panel 6
Each panel allows you to configure the corresponding model's Sampler, Scheduler, CFG, and Steps independently.
This is particularly useful when comparing models that require different sampling configurations.
But there's something important to keep in mind!
Whenever you change a model in the Model Selector, remember to review the corresponding Private Panel settings.
Settings that work well for one model may not be appropriate for another.
If you overlook this, you may end up with misleading results that don't accurately reflect a model's potential.
Each Private Panel also includes a Save Image node for saving that model's output individually.
8. Output Comparison Grid
This is one of my favorite parts of the workflow!
Output Comparison Grid is a custom node designed to combine the images generated by all six models into a single comparison image.
It was developed using the Image Compare (RMBG) node from the AILab node package as a starting point.
To use this feature, you'll need to install the corresponding custom node package listed in the workflow's dependencies.
The Output Comparison Grid combines the six generated images into a single layout, displaying each image together with its model name.
This makes it much easier to evaluate visual quality and see how differently each model interprets the same scene, without constantly switching between separate images.
Layout Options
The Layout setting at the bottom of the node provides three arrangements:
6×1: Six images in a single horizontal row.
3×2: Three columns and two rows.
2×3: Two columns and three rows.
My personal favorite is 3×2.
However, the default setting displays all six images side by side. So don't be surprised when you see that arrangement on your first run!
9. Everything Ready? Hit RUN!
You've prepared your prompt, selected your models, configured your LoRAs and resolution, and reviewed the Private Panel settings.
Now you're ready to press RUN!
The workflow will process the six selected models sequentially through their respective KSamplers, generating images suitable for visual comparison.
Each model runs using its individually configured sampling settings while sharing the common prompt, seed, and image dimensions.
Generation time will depend on several factors, including model size, selected LoRAs, resolution, sampling parameters, and your computer's performance.
10. Where Are the Images Saved?
During generation, each model's individual image is saved separately.
Once the six outputs are available, the Output Comparison Grid is generated as well.
By default, the workflow organizes the results into date-based folders inside your ComfyUI Output directory.
The general file structure looks like this:
YOUR OUTPUT FOLDER / yyyy-MM-dd / hh-mm-ss_model-name_widthXheight.png
Within the same date folder, you'll find both the individual model outputs and the combined comparison grid.
There's a reason I chose this naming system.
When you start testing multiple models with different settings, keeping track of which image belongs to which model and generation session can quickly become confusing.
This structure helps keep everything organized and makes it easier to revisit earlier tests.
11. Performance Scoreboard — Beyond Visual Quality
Once generation is complete, don't forget to check the Performance Scoreboard!
Visual quality is obviously important, but generation speed and GPU memory usage can also play a significant role in choosing the right model.
The Performance Scoreboard was designed to help you evaluate the performance of all six models.
It provides model-specific generation timing and VRAM measurements.
You can also save your benchmark results in the following formats:
CSV: For reviewing and analyzing results in spreadsheets.
TXT: For archiving measurements as plain text.
PNG: For saving or sharing a visual performance report.
This allows you to evaluate models based not only on image quality, but also on their computational performance.
Workflow Timers
Next to the Performance Scoreboard, you'll find two separate timing systems.
Timer 1 — Live Elapsed-Time Display
This timer displays the elapsed workflow time in a large, easy-to-read format.
It's particularly useful when you've zoomed out to view the entire workflow while generation is running.
Timer 2 — Generation History and Timing Comparison
This timer is designed to help you compare generation times across multiple runs.
For example, after changing the sampler, scheduler, step count, or resolution, you can examine how those adjustments affect generation time.
This way, you can explore not only which model produces the images you prefer, but also which settings allow it to operate more efficiently.
A Few Final Words
My goal in building this workflow wasn't simply to run six models at once.
I wanted to reduce the time I spent constantly switching models, adjusting parameters, reloading LoRAs, and opening individual images just to compare the results.
I've found this approach very useful in my own experiments, and I hope it makes your workflow easier too.
Of course, results will vary depending on your hardware, selected models, and LoRA combinations.
Rather than treating this workflow as a tool that definitively determines which model is superior, I encourage you to think of it as a practical comparison environment that helps you discover the models and settings best suited to your own creative goals.
I hope you enjoy experimenting with it as much as I do.
Happy experimenting, and may your next generations be your best ones yet!

