Fullgreed — Z-Image Base Fine-tune (Photoreal, INT8)

Fullgreed is a fine-tune of Alibaba Tongyi's Z-Image — 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. Fullgreed keeps everything Z-Image is good at (bilingual prompt following, text rendering, composition) while adding its own photographic character.

Released model

Both files produce the same images — the INT8 is effectively lossless (measured 0.03% average weight error vs the full-precision weights — visually identical output) at half the size and memory, and it needs no custom nodes: it runs on stock ComfyUI, including ComfyUI Cloud.

Companion files are included in this repo for convenience:

• qwen_3_4b.safetensors — text encoder (Qwen3-4B)
• money_vae_f16.safetensors — the standard Flux 16-channel VAE repackaged in fp16 for a smaller download; interchangeable with the official ae.safetensors
• workflows/alphgreed_workflow.json — ready-made ComfyUI workflow (text-to-image + SeedVR2 upscale stage)

Loading (ComfyUI)

• Diffusion model: Load Diffusion Model (UNETLoader) → greed_int8.safetensors
• Text encoder: CLIPLoader → qwen_3_4b.safetensors, type lumina2
• VAE: VAELoader → money_vae_f16.safetensors
• 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
• CFG: 1–4 (sweet spot ≈ 1–3; a negative prompt works)
• Sampler / scheduler: res_multistep / simple (euler also works)
• Resolution: native around 1024×1024, up to ~4K

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.
• The INT8 file uses ComfyUI's native quantization format — no custom nodes or CUDA extensions. It is not the same as older int8 builds that required a custom node to decode.

Credits & license

• Base model: Z-Image by Tongyi-MAI, Alibaba Group — see the Z-Image technical report (arXiv 2511.22699).
• Text encoder: Qwen3-4B (Alibaba). VAE: Flux VAE family (Black Forest Labs).
• License: Apache-2.0, matching the Z-Image base model.

Please generate responsibly. Do not use this model to create images of real people without their consent.