It is debatable whether this warrants being a custom node, but this is a standalone merge node designed for Krea2; no other nodes need to be connected to it.
It blends two Krea2 models based on a specified ratio. The result is saved in BF16 format within the output\diffusion_models\krea2 folder. You generally do not need to adjust the fp8_scale_mode setting.
Merging older FP8 and BF16 models was straightforward because their layer structures were identical. However, with the recent trend toward quantization, Krea2—being a massive model—is often released only in quantized versions. The issue is that layer retention varies during quantization; there are many models with different layer configurations, even among those using the same FP8 scaled (e4m3fn) format. While a standard model has 430 layers, some can have as many as 1,102. This discrepancy arises because different people make different choices regarding which layers to quantize (which adds scaling information and increases the layer count) and which to leave in BF16.
This node handles the merge by temporarily converting models—whether FP8, INT8, or NVfp4—back to FP32, performing the merge, and then saving the output in BF16 format. This allows for the merging of quantized models with different formats. If "Model 2" is left blank, the node simply converts "Model 1" to BF16 and saves it. Since processing and saving occur sequentially layer by layer, VRAM capacity is not a limiting factor, though a reasonable amount of system memory is required. The process runs on the CPU; in my environment, converting a single model's format took about 70 seconds, while merging NVfp4 and INT8 models took just under three minutes. While using the GPU to load both Krea2 models into VRAM would allow for blazing-fast speeds, doing so would drastically limit the number of users who could actually run it.
・Location
model\merging\krea2
・Note
GGUF format is not supported.
Description
Comments (3)
Do you want to do Krea2 checkpoints? I tried it, and it seems that Krea 2 follows prompts worse than Z-image. It seems to draw well, but as soon as I try to explain what the people in the picture are wearing, something strange happens. But it’s beautiful, that’s true. Although not as beautiful as SD3.5. I’ll try your checkpoints anyway — maybe something interesting will come of it.
But I’m still going to use Z‑image for now.
Thanks for the comment.
For me, Krea2 is a model used for generating source assets. Even if I were to create a custom checkpoint for personal use, I wouldn't release it publicly -nor do I intend to do much with it. The reason is simply that I don't like the license, even though most people (including the team behind Civitai) tend to ignore such licensing terms.
Krea2’s main selling point is that it doesn't use synthetic training data - a quality that clearly shows in the image output - though it may have some gaps in terms of content coverage. However, given the active community, those gaps will likely be filled over time. Z-Image is a model I’m quite happy with; it produces decent images and has a permissive license, with the only real downside being the limitations associated with using a refiner.
Having used SD1.5 for a long time, I am convinced that the potential of 6B models is far from exhausted. In fact, the emergence of models like Krea2 and excellent video-generation models is actually helping to drive improvements in the Z-Image model.
This custom node I created was also designed specifically to get that asset-generation Krea2 model working properly in Forge Neo.
@aueki4g467 License? I definitely didn’t read it carefully enough, because it seemed to me that the Krea2 license resembles the Z-image license. Perhaps the English text was too complex for me. The license is important to me because I use models to illustrate paid books. Thank you very much, I’ll be more attentive. And yes, the license is a problem, and it’s a shame the community is so inattentive. For example, SD3.5 is a great model, but it never fully realized its potential because of the license. In any case, the derivative models of Z-image-base are, in my opinion, the best models for creative work so far. I just downloaded Imaginary.
I also used SD1.5 until recently, but I’m not sure anymore because the limited space on the SSD forces me to limit the number of models.