Raw²diance V2 Final (SDXL/NSFW/Universal)

Rawdiance Final V1

---

Snapshot of my current merging project and testing checkpoint for my loras.

Without any new finetunes for SDXL its hard to improve current top models any further. I try my best (and not break the rest).

Rawdiance V1 Final

Outputs crispy detailed and beautiful images while trying to look as less synthetic as possible. This was my goal from start. Yeah i know - synthetic robots in the samples. Pfff..

Non DMD2 Version since you need to load the DMD2 Lora anyway to tune when using more Loras and it can be used for Lora training. I still recommend DMD2 / LCM / CFG1 / 12 Steps. Use common settings for generation with other samplers. When using DMD2 Lora with other Loras together start at 0.75 for both and up CFG to 1.3.

May post a DMD2 version if demanded.

Have fun and do not forget to post your creations and feedback.

---

Proto1:

DMD2 Version; Use LCM, CFG1, 12 Steps. Other settings may work but not tested.

Proto2:

Minor balancing of DMD2 to work better with Euler-A and other uniform samplers. CFG can now go between 1-2. Start with 10 Steps / CFG 1.5 for Euler-A to reduce noise.

12 Steps for LCM, CFG 1.25.

Reduced likelihood of unwanted buttcracks.

---

Tip: You can load the DMD2 Lora with slightly negative weight for fine tuning output. (Necessary for stronger samplers like DPM). Lowering DMD2 weight needs higher CFG.

When using Character-Loras start with weights about 0.7.

Note regarding my prompts: I use Invoke, + is used to up weights like (word) = word+ . Stuff in <> are embeddings.

Feedback, suggestion and/or requests welcome.

---

WIP; Will add more samples over time...

---

Mirror and other Stuff: https://huggingface.co/nphSi/