High resolution ++ for Krea2 (Patreon)

Ever since releasing 《High resolution for Krea2》, I’ve been wanting to make further improvements based on it. Today I found a method that can greatly enhance image quality without altering the overall structure of the original image. Let’s first check the comparison between the cover image and the earlier version.

The image on the left is the input image, the one in the middle is generated by the previous workflow, and the one on the right is generated by this workflow. Isn't the difference obvious? Do you like the final generated result?

Let's look at more sets of sample comparisons.

How is it? Isn't the effect quite noticeable? If you are interested in this result, keep reading.

First, let’s bring out our old friend — the Krea2edit node. This node comes with a fixed edit workflow, and I only made minor adjustments to its workflow to achieve the current effect.

Nearly all edit workflows follow a similar structure. In the upper section, the model, CLIP and conditions are input into the Ksampler. In the lower section, the dimensions of the image are extracted and injected into the latent space before feeding into the Ksampler. The Ksampler then outputs the latent representation, which is decoded by the VAE to generate the final image.

In the sampling process of KSampler, the overall structure of the image is computed first, followed by the refinement of image details. I leverage this working principle by connecting two KSamplers in series. The first one computes the overall image structure while loading the

《 High resolution for Krea2》 LoRA to lock the structure of the original image to the maximum extent. The output latent is then scaled up 2x (to generate more details) and fed into the second KSampler for subsequent detail refinement. Two points need attention here:

The parameters of the two KSamplers must be identical. The only settings to adjust are start at step and end at step. The first KSampler handles sampling for the first eight steps, so its parameters are set to 0-8. The second KSampler performs the subsequent sampling, so its parameters are set to 8-12.

The latent fed into the second KSampler is scaled up 2x, which equivalently doubles the image resolution. However, the original image dimension information stored in the positive prompt has not undergone this 2x scaling. Directly feeding this into the second KSampler will produce ghost images. To resolve this issue, I have specially developed a dedicated node. Now you only need to connect the NO8D-Remove Krea2 Reference Latents node before the positive prompt input, and images can be generated normally. This node can be found within NO8D-Controls.

https://github.com/no8d/ComfyUI-NO8D-controls

Don't forget, the most important EDIT-LORA : High resolution for Krea2

This is the workflow for this session. Isn't it simple? Here's what the complete workflow looks like.

OK! That's all. If you enjoy my sharing, you are welcome to join my private club. If you are already a member, you can get the project files for this sharing directly below.

Fv2pWtzrdOKXF1QCCBmhX-9QsNdfAP0xEcegS7JSk6w

🫡 May the AI-power be with you, see you soon !🫡

Disclaimer: My model or workflow is solely for learning and communication purposes and shall not be used for any commercial activities. All responsibilities for the generated content, including its legality, ethical implications, or potential misuse, are borne entirely by the user. By using this model or workflow, you agree to this principle. Please use it responsibly within the scope of local laws.