Alphgreed - Greed Int8 Showcase
# Fullgreed — Z-Image Base Fine-tune (Photoreal)
Fullgreed is a fine-tune of Alibaba Tongyi's Z-Image (Omni-Base) — the 6B-parameter Single-Stream Diffusion Transformer (S3-DiT) — tuned for photorealistic, phone-camera-authentic portraits and selfies: natural lighting, believable skin and hair texture, and outputs that read as real photos rather than "AI renders."
It is a surgical fine-tune: only the attention and feed-forward projection weights were trained, leaving the base model's norms, embedders, and timestep conditioning untouched. In practice this means Fullgreed keeps everything Z-Image is good at (bilingual prompt following, text rendering, composition) while adding its own photographic character — and it retains the base model's native reference-image editing ability (see Editing below).
---
## Files
| File | What it is | Size | Use it when |
|---|---|---|---|
| fullgreed_i8_plain_comfy.safetensors | INT8 quantization, ComfyUI-native format | 6.3 GB | Recommended for most users — measured 0.03% average weight error vs the full model (visually identical output), half the size, half the memory, loads faster, and needs no custom nodes — runs on stock ComfyUI including ComfyUI Cloud |
| fullgreed_f16.safetensors | Full-precision model | 12.3 GB | Reference quality, LoRA training base, or any tool that doesn't support ComfyUI's quantized format |
Both files produce the same images. If you're unsure, take the INT8.
---
## Requirements (ComfyUI)
Standard Z-Image companion files (same as the official Comfy-Org release):
- Text encoder: qwen_3_4b.safetensors → CLIPLoader, type *lumina2**
- VAE: ae.safetensors (Flux 16-channel VAE)
- Diffusion model: load via Load Diffusion Model (UNETLoader)
- ModelSamplingAuraFlow node with shift = 3
- Latent: EmptySD3LatentImage
Also runs in Draw Things (import as a Z-Image model) and anything else that supports Z-Image.
## Recommended settings
- Steps: 8–20 (quality range 80–30)
- CFG: 1–4 (sweet spot ≈ 1–3; real CFG and a negative prompt work)
- Sampler / scheduler: res_multistep / simple (euler also works)
- Resolution: native around 1024×1024- 4024x4024
---
## Tips
- Slight CFG restraint (≤4) preserves the photographic look; high CFG pushes toward an over-processed render feel.
- The model responds well to camera-language prompts: phone selfie, mirror shot, golden hour, indoor tungsten, shallow depth of field, etc.
- INT8 note: fullgreed_i8_plain_comfy uses ComfyUI's native quantization format — no custom nodes or CUDA extensions, works on ComfyUI Cloud. It is not the same as older int8 builds that required a custom node to decode.
## Credits & license
- Base model: Z-Image (Omni-Base) by Tongyi-MAI, Alibaba Group — see the [Z-Image technical report (arXiv 2511.22699)](https://arxiv.org/abs/2511.22699). This fine-tune inherits the base model's license terms.
- Text encoder: Qwen3-4B (Alibaba). VAE: Flux VAE (Black Forest Labs).
Please generate responsibly. Do not use this model to create images of real people without their consent.