This model is a precision-tuned ZiT merge designed to push the boundaries of fine-grained texture, anatomical accuracy, and atmospheric lighting while maintaining compatibility with LoRAs trained on the base models. I've train my character and style LoRAs on vanilla Z-Image Base. They work fine on Turbo derived models and tend to be more stable during training.
I will upload future models to Hugging Face also, it's a lot faster than dealing with a VPN
https://huggingface.co/PsychoLogic/PsychoMix_ZiT
The Composition
The architecture of this model is a blend of checkpoints and custom refinements:
v1.x
Core Foundation based on Moody Porn Mix v10 for subject pose, with various other models merged for style and consistency (see 'About this version' for specifics for a particular release)
For a more raw/realistic look I recommend using z-image-turbo-flow-dpo with a strength of -0.5 to -0.75 (yes, negative). It was baked into the base model and many (myself included) think it's merely a sloppifier.
v2.x
Core Foundation: A balanced merge of Z-Image Turbo NSFW v7.2 for subject pose, CyberRealistic Z-Image Turbo Catalyst v0.7 for style and technical precision, and selective merge from Famegrid ZIB v1.0, focused on later layers
Feedback & "Show Your Work"
I am actively looking for constructive feedback to improve this and other models. If the model isn't performing as expected for you, please help me fix it!
The "Pics or it didn't happen" Rule: If you encounter issues (artifacts, anatomy failures, or poor prompt adherence), please post your generation with metadata (Seed, Prompt, Sampler). Negative comments without an image and settings typically don't help. If you want a fix, show the failure so I can tune it out in the next merge.
Description
Core Foundation: A balanced merge of Moody Porn Mix v10 for subject pose, CyberRealistic Z-Image Turbo Catalyst v0.7 for style and technical precision.
Atmospheric Weight: Integrated a small amount of DarkBeastZ v8 in later layers to enhance tonal range.
FAQ
Comments (19)
Fair attempt to improve z-image. It's a fun run trying all kind of images. It's my personal experience based on a limited type of compositions and lightings I tried, I found it lackluster both in quality (unnatural sheen in low light, low contrast, over exaggerated skin texture, with patches of different skin types in one body part), and prompt following (especially when it comes to body type, and in addition to that, huge bias for asian facial structure). Also, even though it unlocks some of the nsfw stuff, the pussy has no clitoris most of the time which makes it incredibly creepy and somehow worse than the original zit which blurs the area, synthetic training data does it to models.
Yep, moody has a strong Asian bias, I'm still working to break it
I haven't tried a lot of explicit NSFW gens yet. If you could share some images (with metadata) where this under-performed it would certainly help
@psychologicau I posted several images in PNG format with ComfyUI metadata. It's decent if you don't go NSFW. It's a bit tricky, the model tries to undress the actor, even if you describe the clothing in full detail, and when it does, the skin part that was undressed has even uniform lighting on it, which is not always fitting the scene. So, my understanding: a) Not enough dressed characters in the training data, b) not enough variety in lighting in the scenes.
I tested V1 and for me, the checkpoint is wonderful. I create very good pictures with very good promt adherence. I used er_sde/sgm_uniform 8Steps, almost all the images are error-free. I take out a bit of color with dpo lora. Thanks for the checkpoint, i like it very much.
This is the best nsfw model at the moment. Very clear images, no blurring. It is practically unbreakable by lore and even at the moment of breakdown it retains the image structure for a long time, i.e. to cause mutations you have to try very hard. Huge variety of faces. And finally, endless variability of images per prompt.
Please add FP8 or Q8 GGUF versions of all your models most of us have old hardware that cant run big models :(
I'll look into this. In the meantime, you should be able to use Diffusion Model Loader KJ to load into fp8 dtype
I found this following through the thread of a similar request on UnstableBastard: https://github.com/qskousen/ggufy
Seems to have worked there, will give it a go today
fp8 conversion being tested now, used this:
https://github.com/silveroxides/convert_to_quant
convert_to_quant -i PsychoMix-ZiT_v1.1.0.safetensors --comfy_quant --device cuda --manual_seed 42
@psychologicau Hi ! i can give you the file for converting directly to Q8_0.gguf , Just tried the file i use for my model and it works . ggufy is faster and now it has a GUI it seems .
@BastardOG Thanks, that'd be great :) I'm still working on building/packaging ggufy, was having issue with zig versions and Ubuntu... whole can of worms
q6_k gguf added using ggufy. The q8_0 from that tool produced all black output and q4_k failed. I tried more recent ggml release to no avail.
q8_0 using the script from @BastardOG also produced all black output for me. I'm not sure if I should blame ComfyUI at this point
@psychologicau Damn...blame comfy :)) i have the portable version , if something doesn't work i just install in another folder and move over the models folder and some custom nodes , then just shift+del the old folder
The issue with q8_0 was most likely the same outlier weight as broke the model for forge neo. I'll try and quantize the fp16-fix version
@psychologicau hi, i wrote ggufy, i'm sorry to hear you were having trouble with it. i read what you wrote about the bf16 values, that is interesting, i had not heard that before. if things are simply converting directly across from bf16 to fp16 without any kind of scaling... that would certainly be an issue haha. if you have any problems with ggufy, let me know - i'm always looking for any kind of feedback on it.
@psychologicau i'm curious if you had issues with the prebuilt binary, or if you just prefer to build from source?
@ferretduck I prefer to vet the source and build things myself, and to run things in containers. I have a Dockerfile / docker-compose.yaml set up locally so it's a matter of docker compose build && docker compose run --rm ggufy convert --datatype q8_0 --output-dir /app/output /app/input/PsychoMix-ZiT_v2.0.0.safetensors . I'll make a fork and push these to Github at some point.
The issue I had with conversion was a single BF16 weight in my model being outside of the FP16 range, I guess either it was converted via FP16, or it also exploded the conversion to Q8. Clamping this weight before feeding through ggufy fixed it for me.