From what I saw in the images, the model "burned", if I can say so, but to make it easier to understand roughly, "the merge burned", I know this because I tried to merge models with different architectures at the beginning of my work as a developer and this happened due to two factors:

1: models with different architectures being merged.

2: embedded used in fusion and merge do not go to the merged model, which can cause this type of failure.

I found this out the hard way, or rather I don't know.

But here are some settings that have saved me sometimes

Sampler_names: Most that have the " cfg_pp " function

It could at least make the images sharper, not that it will solve anything, but those blurs will be less visible.

@Sr_RockwellyDesu Yes there seems to be something fundamentally wrong with the latest models of @CHROMEKIDD . It is the same problem with HANA-4-CHROME HUGE. It looks like the metadata of those models claim it to be v-pred causing ComfyUI to automatically use v-pred settings, which cause this issue. If you include the ModelSamplingDiscrete node and set it to eps with zsnr negative, you will get "normal" images. But with most settings those images are complete burned, so it looks like something like the DMD2 LoRA is also included. LCM Karras with CFG 1.5 and 12 steps then gives fine results, but it still doesn't match the ones from the showcase images. A proper set of generation parameters would be really helpful here... :/