Took me a while to check out Krea2, but holy mother of god it's good. Definitely the first model I've used that I could contemplate not using detailers (still using detailers though). But quality is one thing, actually listening to you is another- here as well we have something of a quantum leap. So I paired the wonderful basic VL encoder with my Simple Qwen GGUF setup. You technically do not have to prompt at all,not a single word. Each mode has an appropriate system instruction. Without user input it will look at your source and make a prompt from it.

I like this turbo model a lot. Works great. Go give it thumbs.

But any turbo model will do. Or full model, just edit my step slider to go higher.

The usual stuff to go with it Qwen3VL encoder and vae. And SAM model, and the upscaler of your choice. I like remacri. And each detailer needs a bbox detector of course.

This started out as a simple batch processor - take a folder of old crappy SD images from ancient days (aka two years ago) and pass them all through Krea2 to clean up and detail with one click, everything automated to prompt itself and run until every image is done.

It still does this of course, but it grew into a general-purpose Krea2 workflow and supplanted my existing setup. So I spent a few days cleaning it up so I could share it with you. This is great motivation for getting the functions debugged. Please note though that I claim no great Krea knowledge. I learn new stuff by picking a basic template and pulling in my own modules, or porting the core architecture into one of my workflows. I'm quite sure that there is a better way to do the detailing. The bit I'm confident is the the backbone. The key here is to be able to work with large numbers of old images, to process them without losing their filenames, and to be able to use a local LLM to guide a custom unique refine for each image.

I have not yet integrated full edit/swap functionality. There is, in the meantime, a Preserve Face option that will, when enabled, keep your subject's face from being resampled in the initial and refine ksamplers. It will be detailed if the face detailer is on though. It's just a mask though - denoise strength must be low to keep your head in position. This is a stopgap, I have not yet started playing around with the nodes/LoRAs that are for editing specifically.

This might be useful if you want to:

• Process 1000 awful a1111 generations overnight to transform them from a gigabyte of trash into 100MB of slightly nicer slop. Dial it it, put in your folder name, press run.

• Try Krea2 with straight I2I at any denoise strength, or T2I with a super-precise recreation prompt by Qwen. (It really does a great job.)

• Use Krea2 as a detailer only - set denoise to 0, turn off refine and turn on your details with high strength to preserve the source image any only fix what you need to - the four default detailers are Face, Nipples, The Womanly Slot, and the Manly Pole. Each detailer uses it's own conditioning and LoRAs. TBH my detail nodes are a bit long in the tooth. I'm planning on completely rebuilding this with the newer tools.

I am uploading the WF as a folder - it's all the exact same json, I have just activated the appropriate groups for a few scenarios. The ONLY difference between them is the state of the bypass toggle switches. For my example images I am going to upload composites of source and final output, with banners for the mode. The tags are mostly correct, but sometimes it says something stupid like i2v, or has wrong setting listed. Either way, they all have the metadata for that run. Don't run those though! The WF is in various states of completion in the examples, so there could be serious bugs. Use metadata to glean settings only.

Here is an overview of some modes. There is a plethora of notes in the workflow, as is my wont.

Text to Image

* Disable source image and i2i describe. If you set the LLM bypass to true, your user prompt will be used as-is. If LLM is enabled, your prompt will be cleaned up and embellished per the system instructions on the left. Qwen can write an awesome prompt for you. Just give it a sentence or two. It won't care how perverse and twisted your filthy mind is. You can also use the 'Override' prompt box. That will ALWAYS be the main prompt if enabled.

Image to Image

• Enable source image and i2i describe. Load your source image. Just like above, you can use a straight user prompt, if LLM is enabled it will describe your image (really well). The user prompt is for any special instructions/image mods you want. More info in the group.

• You can do i2i in several ways - enabling the i2i encode group perform the classic i2i - you set the denoise strength with the slider to determine how much freedom the ksampler has. If you are trying to preserve features/just going for a once-over to clean up the image, you may want to disable the refine stage, or lower the refine denoise. You can tweak that however you like.

• Because the combo of the LLM and the Qwen encoder, you can actually also do i2i with the i2i encode group OFF. This really can do an amazing recreation of the image without using the source pixels at all! Very cool, you should definitely try this too.

• With either method, you can put any changes you need into the USER prompt and the LLM will obey. Amazingly well. Change your character completely, or just features, or the background. No limits when not encoding original pixels, more rigid when you are- the extent of flexibility being the strength of the denoising.

• Simple detailing - if you need to fix one specific thing only, for example a botched beaver, or nasty nipples, or a putrid penis. Turn off refine, set main denoise to 0.0, turn detailing denoise way up, like 0.8 and up. Enable face protection if you don't want to lose a good face. Crank up those detailer LoRAs and see what happens. (Don't worry if the main sampler still runs, it won't be used.)

Batch

Does a whole folder. You only need to enter the path of your source folder and one for your outputs. Your original filename is preserved with an suffix to denote kreafication. You set comfy to instant and click run once when you are ready. It won't stop until it runs out of slop to process. It keeps track with a temporary json in the source folder (it disappears when the batch is done). The images are automatically scaled and cropped to closest workable resolution, so you don't have to prep them. Some of my instructions have strict rules about removing watermarks, corner logos, crap that you want removed. Some don't, because it can interfere with text that is wanted. So choose the one that best fits.

Core nodes were prioritized wherever possible to keep the graph clean and stable. Below is the list of required and optional custom nodes needed to run the workflow properly.

Tip: If you are new to the LLM setup, check out my Simple Qwen Template Workflow to test and iron out any kinks before running it inside this full pipeline.

Essential Custom Nodes

These nodes are integral to core logic, batching, math, and model architecture:

• Simple Qwen GGUF – The heart of the prompt-generation and analysis pipeline.

• ComfyUI Impact Pack & Impact Subpack – Core logic, detection, and workflow processing.

• ComfyUI-KJNodes – Essential utilities and tensor manipulations.

• ComfyUI-VideoHelperSuite – Video loading, batching, and combination nodes.

• comfyui-dream-video-batches – My usual batch processing management setup.

• rgthree's ComfyUI Nodes – Workflow routing, muting, and quality-of-life primitives.

• CRT-Nodes – Processing and pipeline utilities.

• ComfyUI-mxToolkit – Logic and parameter routing tools.

• ComfyMath – Dynamic resolution and calculation nodes.

Nodes you could replace, but it will void the warranty

The workflow is designed around these nodes, but they can be bypassed or substituted if you prefer:

• save-image-extended-comfyui

  Drives automated file naming and folder sorting. If you replace this, ensure you re-map all folder paths, prefix strings, and dynamic keys.

• Pixaroma

  Visual comparison tools. Highly recommended for evaluating output changes, but can be replaced with standard pairs of preview nodes.

• ComfyUI-Lora-Manager

  LoRA organization. Not strictly mandatory for execution, but essential for quick switching and trigger phrase handling.

• Krea 2 Conditioning Control

  Conditioning adjustment. Optional; can be bypassed or replaced with alternative conditioning nodes. I highly suggest you visit the repo to understand how this and similar conditioning modifiers work. It isn't exactly straightforward. When using a mod that is already NSFW is can be superfluous, but not always. Play around with it, you can get really interesting result.

• ComfyUI-JakeUpgrade

  Aspect ratio / dimension toggles. Used here for the instant width/height SWAP feature. You can substitute any standard resolution selector if preferred. Just make sure not to mess with anything downstream- move on those H+W noodles on the output of the node.