PonziReality
PonziReality is a photorealistic, aesthetic-photography focused fine-tune of Z-Image (Base). It prioritizes high-fidelity human subjects, cinematic lighting, and stylized photographic looks: neon noir, saturated teal-magenta LUTs, high-contrast night street scenes, studio portraiture with film grain, and carefully textured costume/fabric detail. The model was trained for realism first, with a secondary emphasis on bold, editorial color grading and compositional polish suitable for fashion, concept portraiture, and moody environmental portrait work.
What this model excels at
Photorealistic human portraits: Accurate facial anatomy, convincing skin textures, directionally consistent catchlights and specular highlights.
Aesthetic lighting styles: Strong neon/gel lighting, split lighting, rim light, tungsten + cyan cross-lighting; looks that read like editorial night photography.
Clothing & materials: Realistic fabrics (leather, wool, silk), pleats and pleated skirts, reflective surfaces, pearls/metal accents.
Cinematic composition: Full-body and 3/4 portraits with strong posing, purposeful negative space, and dramatic perspective.
Low-light, high-contrast scenes: Retains detail in mid-tones and highlights while producing striking shadows.
Artful surreal touches: Controlled stylization (props, head swaps, costume fantasy) while holding onto photorealism when requested.
I would love to see what you guys can come up with, so please share your creations! I would also love to hear your comments and critiques.
Bonus: Drag one of my images into ComfyUI to view my workflow. I would love some feedback!
Special thanks to meep morp for giving me ideas for the showcase images!!
Description
Key Improvements for V1.11
Enhanced Dataset: Performed professional color grading across the entire dataset.
Higher Fidelity: Boosted the Training Resolution and Rank.
Expanded Library: Added curated new images to the training set.
Optimized Merging: Merged with v1 to retain original concepts.
Final merge with the Base Model to ensure flexibility and prevent overfit.
FAQ
Comments (17)
Will there be a FP16 version as well?
euler/simple with cfg 1? steps 30-50?
This is the base model, so you do need CFG.
I personally use exponential/res2s + flux 2 scheduler, 50 steps + 5 steps refiner, CFG 5.5.
I suggest dragging one of the images into comfy to see the workflow, because there's a lot more going on especially with the clownshark setup.
@ponzi thanks
You can use this lora if you don't want to make 50 steps: https://civitai.com/models/2362961/z-image-fun-distill-8-step-lora
@mphobbit Thanks! I kinda do like the negative prompt and extra steps though, especially since I made a really nice setup for enhancing prompts and creating negative prompts all at once. That's the main reason I use the base model over turbo.
Need "Erotic" finetune. :) or NSFW? :D
I only see "Need "******" finetune. :) or ****? :D" but feel free to DM me! hahaha
OHH NVM I SEE NOW! The model can definitely produce NSFW content already, but I'll see what I can do! This might take a while though, cause I don't currently have a NSFW dataset but I think I could easily make one.
Can you link the Clip and VAE you use on your workflow ?
I've also had good results with Qwen3 Reranker 4B F32 as a text encoder, interestingly enough!
https://huggingface.co/prithivMLmods/Qwen3-Reranker-4B-F32-GGUF
Just make sure you add these flags to ComfyUI: --fp32-text-enc --fp32-vae
(Or just --force-fp32 if you want to run the U-Net, VAE, and text encoder in FP32)
Have been getting excellent results with fp32 version + LoRA stack (enable NSFW, Asian mix, etc.). Great work!
Thanks massivetits!