ZIT-ZEUS-DARKMOP YARA "New Engine"
Detailers or Lora's are not really needed.
res_2 - beta 8 Steps / res_2s - bong_tangent / er_sde - sigmoid_offset / ipndm -. beta / euler - beta (10 Steps)ZIT-ZEUS-DARKMOP SIX-ONE (6.1) (reup)
Please be so kind as to let me know if anything doesn't work or could be improved—I'll fix it in the next version :)
A Simple Workflow for DARKMOP: HERE
DARKMOP XL (FP8) Version INFO!!! FP16 OK :)
Users with a 5950X CPU should not run this; your Core 3 may have failed—unfortunately, this is a known issue with the Ryzen 5000 series. (applies only to the FP8 version)
This version is more or less a test run. I liked some of the results it produced with 12 steps using the “ipndm - beta” sampler, or “ipndm - sgm_uniform.” The model is inspired by raw, vintage, and analog photography. There may well still be some bugs in it; I was only able to test the model myself with 3 images (FP8) before my Ryzen threw a Core 3 error. The FP16 version runs without any issues. However, it may also contain bugs. USERS with a 5950X CPU dINFO & SETTINGS:
Online Generation: I have noticed that online generators using the “Euler” sampler do not always achieve the results that this model is actually capable of. My recommendation for the sampler, fast and good details: “er_sde - sigmoid_offset.” Unfortunately, not all online generators offer this combination. ANIFAN Version: Anime & Fantasy
Trigger: (Anime) (Illustration) (test it ;) )
(euler - simple / er_sde - sigmoid_offset 8 Steps CFG: 1.0)V6 = DARKMOP 6 (AIO) Version
V6 AIO Sampler: dpmpp2m - beta 7 - 10 Steps / CFG 1.0 - 1.5
(Place here: \ComfyUI\models\checkpoints\)V5 = SIMBA (Other Series) ((You can find it on my profile under “Models”))
V4 = MopPhoria Realistic and authentic (Better Skin, Similar to Qwen Hyphoria)
V3 = VMH Realistic and authentic (Similar to Qwen Hyphoria)
V2 =DarkMop Real and fantasy / (more Fantasy)
V1 = All-rounder (Successor to my ZIT-ZEUS-ONE model)
MOP = "Moody or Pervy" XD (not "Mixture of Perverts") from GBRX
We classify this model as an all-rounder ;)
We all know that some details can only be improved with LoRa.
All sample images for this checkpoint were created without LoRa. (Sampler settings er_sde - simple 8 steps, cfg 1.0)
Feel free to put it through its paces ;)
V6 Sampler Settings:
dpmpp2m - beta / 7 Steps / CFG 1.0 - CFG 1.5)
V3 & V4 Sampler Settings:
Samplers Normal: 8 Steps
Euler simple (Speed)
ddim beta (Speed good)
er_sde simple (realistic more Details) (perfect)
er_sde sigmoid_offset (realistic more Details)
er_sde kl_optimal (Nice Colors ;) )
dpmpp_2m_sde beta (Less skin texture)
dpmpp_2m_sde simple (more Less skin texture)
Samplers: Medium / High Q. 8 Steps
res_5s beta57
Samplers 4 Play XD (8 Steps)
res_2s sigmoid_offset
Samplers: High Quality 8 Steps
res_2s bong_tangent
res_3s bong_tangent
res_5s bong_tangentLet me know which settings gave you the best results.
Description
DarkMop 6.1
FAQ
Comments (6)
fp8 please
It's actually surprisingly easy to convert a fp16 to fp8. You can do it yourself even on a potato computer. It uses ram/cpu not gpu for conversion and takes less than 5 minutes. If you don't know how, just ask gemini/claude/chatgpt, i actually had claude-code do it for me.
Yara is probably the best Z image Turbo model I've ever used. And maybe not just Z image. It even handles simple queries well - I haven't had time for complex ones yet.
There's almost no bad anatomy, all the actions I ask for are correct. Cool model.
Hey @herkus_baronas631, thank you so much for the feedback! It’s great to hear that the anatomy and the action-handling are working out for you. I spent a lot of time fine-tuning the balance to keep it stable even with simple prompts. Stay tuned – I’m currently experimenting with some new architectural hybrids to push the detail-limit even further. Happy generating
You asked to tell you what's wrong so that you can fix it in the next versions.
I don't know if it's a bad thing or if it's just the 'genetics' of Z image - when making smoking people, sometimes cigarettes get stuck somewhere near the fingers. Sometimes the fingers get deformed.
That would be an observation from previous versions. But such things happen with all models, except maybe Qwen.
Please add Q4-8 GGUF or FP8.
Not all of us can run huge models :(