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
Second attempt at reducing the Asian bias from the base model. I feel like I nailed it this time. SFW gens of male subjects are working well, and it isn't feminizing my SelfAI shots.
I re-balanced the HFP dataset to 50/50 male/female subjects, and used this to make a ZiB LoRA to apply to Moody and Dark Beast before merging them in. It's only 5000 steps in at this point but the first test run was too good so I decided to cut v1.1 using it now.
I've been training my character and style LoRAs on vanilla Z-Image Base. They work fine on Turbo derived models, don't listen to any ignoramus who tells you otherwise.
(v1.1) FP16 Compatibility Fix (for non-NVIDIA & ComfyUI users)
I am uploading a new 'fp16' variant of the v1.1.0 model to resolve the "Black Output / NaN in Latent" issue reported, particularly those using SD-WebUI-Forge or non-BF16 LoRAs. Note the weights are still in BF16 but will no longer produce NaNs when dumbly cast to FP16 by dodgy inference pipelines.
What happened: During the merge process for this model, a single attention layer developed weight outliers (values exceeding 65,504). While these weights technically function in BF16 (used by ComfyUI on my NVIDIA hardware), they exceed the numerical range of FP16, which is the standard precision used by some hardware and inference implementations (Forge Neo). When these engines encountered those extreme values, they would overflow into Infinity/NaN, resulting in black images. The fixed version clamps these outliers to a safe range of [-60,000, 60,000], ensuring compatibility across all platforms.
fp8 quant for this version has been added. This was made using https://github.com/silveroxides/convert_to_quant, convert_to_quant -i PsychoMix-ZiT_v1.1.0.safetensors --simple --comfy_quant --device cuda --manual_seed 42 . Note it produces different results to the bf16 model and seems less stable. I tried also without the --simple, it took longer and produced similar a result which failed to render the shadows correctly in the 'girl at blinds' image.
Still working on GGUF quants, should be a matter of days. Would be quicker if I didn't have to upload my free content via a VPN because CivitAI don't deem the Australian market to be worth the investment.
FAQ
Comments (17)
Is this trained on wieners ?
This website can't display whether the file is for the FULL version or the Pruned version! The website is becoming increasingly stupid!
I'm not entirely sure what your intended meaning of 'full' vs 'pruned' is here. These are the DiT weights only (no packaged TE, VAE) which some might mistakenly call that 'pruned'. I have not pruned these weights
Also does not include the training optimizer states, which others might call 'pruned'..
this could be my new favorite ZIT checkpoint; i need more tests, but it looks very promising for vehicles 😎
That capability is coming from the base models, I actually removed all vehicle images from my fine-tuning dataset used here. I really need to throw together a comparison workflow to get SBS images showing the different component models in isolation
@psychologicau i've tested almost all the ZIT checkpoints, and this one appeals to me the most. i even tested my 'Robert Williams' LoRA only to find that the pictures look better when i don't use the LoRA, LOL.
b16 in neo is getting black screen. I tried your v1 no issue in neo.
I had same issue in ComfyUI - but disabling the LORA is the only thing I could find that fixed it.
I'm trying to get neo running in Docker now to test this
@Melodic_Possible_582589 I have forge neo running now (in Docker, built from commit 4c55a97a133b153712e6f9dac5d54afea2cbd713 (HEAD -> neo, origin/neo, origin/HEAD)), and can verify the BF16 model is working fine
...until I tried adding a LoRA, then black output and this in console:
```
webui-1 | Encountered NaN in Latent processing.py :: WARNING
webui-1 | If you are using SageAttention, try --disable-sage
```
@psychologicau Hi, I got it to work by changing the drop down menu at the top right "Diffusion in low bits" to something besides float8 options. Thank You. Testing it out now.
It might be this:
```
Extreme value in model.diffusion_model.layers.8.attention.qkv.weight: 4.352132950612665e+25 (Exceeds FP16 range)
```
I'm digging into the forge code so see if it's derping the UNet into FP16
That was the issue. The range of that weight being outside of fp16 numerical range was causing NaNs when the model had been cast to fp16 (as forge neo was doing). This was working fine in ComfyUI for my NVIDIA cards as it stayed in bf16. I'm uploading a fixed version
@psychologicau thanks for looking into it. Sounds so complicated. I'll be waiting for the updated file :)
@Melodic_Possible_582589 you/re welcome. Was an adventure getting forge running and figuring out the issue. I have a new stage in my safetensors file checks now for outlier weights so it should be easy to prevent next time. Thanks for reporting it.
It took all night over a slow VPN but looks like the 'fp16' file is up now
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.