Color Match - Bad results. Is it Krita AI or is it my models?

[hansnolte]

I have tested with
FLUX.2 [klein] 9B (flux-2-klein-9b-nvfp4.safetensors)
FLUX.2 [klein] 4B (flux-2-klein-4b.safetensors)
and
Qwen Image Edit 2511 (qwen-image-edit-2511-Q6_K.gguf)

Prompt:
remove the teapot

In FLUX.2 [klein] 9B and FLUX.2 [klein] 4B are bad colors.
In Qwen Image Edit the gradient is bad, please zoom in.
You see my results in the History of the Krita file.

Since I don't know whether the color changes and the destruction of the color gradient come from my models or from Krita AI, it would be nice if others tested it.

Latest ComfyUI with Krita AI 1.48.

Krita file to download (start ComfyUI first)
ColorTest.zip

By
Hans

[Acly]

You need to use a selection.

[hansnolte]

Hi Acly,

many thanks for your very fast answer.

You need to use a selection.
But with an selection it doesn't remove the reflection and the shadow of the teapot

[Acly]

Then make the selection bigger :P

[lokpro]

(What I want to say, but written with the help of AI):

The inability of image editing models to keep raw texture is a common challenge in generative AI and digital manipulation. This often happens because the model "re-interprets" the image pixels rather than just modifying specific areas, leading to a loss of original fine detail or "waxy" artifacts.
Why Models Lose Raw Texture

• Generative Over-Smoothing: Many AI models, particularly diffusion-based ones, tend to produce "plastic" or "perfectly smooth" skin and surfaces because they are trained on high-quality, often filtered datasets.
• Lossy Processing: Editing an image often involves re-sampling or resizing, which can be compared to "denaturation"—once the pixels are manipulated and saved in a lossy format like JPEG, the original raw texture cannot be perfectly restored.
• Re-encoding Artifacts: When a model performs an edit, it often decodes the image into a "latent space" and then re-encodes it. This process can discard high-frequency noise (which we perceive as grain or texture) in favor of more stable, low-frequency shapes.