So I've been playing with those Chroma Radiance models (very underbaked I guess, the then latest one was okay, but the one released a day later was absolute shit). I noticed it decides to denoise some stuff it can't recover from. This happened with Euler, etc. I thought, huh, I wish there was some mechanism for it to correct its mistakes.
Anyway, I noticed that Euler ancestral worked much better. I tried it with current Chrome (v48) and it works even better there. Pic related, Euler ancestral and DPM++ 2s ancestral "converge" to the first image, while ~all of the non-ancestral samplers converge to the second image. Keep in mind that I'm using a lora based on a Belgian comic book, so the one on the left is more correct. Give Chroma with ancestral samplers a chance.