Model BF16 Merger with 2 Loras. It does pretty much everything without adding Loras. I find it much better without Loras than with. I use it with a simple workflow in comfyui and sd forge neo classic with Euler Beta57 or simple in 8 steps.

The only downside is its 20 GB size.

The NVFP4 version is lighter, good quality, comparable to the BF16 version, faster (19 seconds on an RTX 3060 with ComfYUI), still has the problem with the Loras, works much better without them.