Tuned for a more realistic look than the bare base model: better skin, light, and texture, with less of the plastic, AI-skin tell.
Strong on photo, portrait, and mood imagery.
Fast Turbo-style sampling: about 8 DiT steps.
Runs in plain Python via Diffusers, or as a single merged UNET in ComfyUI.
Apache-2.0: free to use, free for commercial use, and redistributable.

Limits

Weak at legible text, including Hangul, Devanagari, and most non-Latin script. Route text-bearing surfaces elsewhere. Also weaker on graphic, print, and typography-heavy work than on photo and portrait.

Turbo-style sampling runs without classifier-free guidance (guidance_scale=0 / ComfyUI cfg 1), so negative prompts are ignored. Put avoidances in the positive prompt instead.

Quick start (Diffusers / Python)

Enterprise and production default. After the Hugging Face upload includes transformer/, load that Diffusers transformer and keep the text encoder / VAE from the base pipeline.

bash

pip install -U torch transformers accelerate safetensors
pip install -U diffusers

python

import torch
from diffusers import ZImagePipeline, ZImageTransformer2DModel

transformer = ZImageTransformer2DModel.from_pretrained(
    "redrob-labs/redrob-image",
    subfolder="transformer",
    torch_dtype=torch.bfloat16,
)

pipe = ZImagePipeline.from_pretrained(
    "Tongyi-MAI/Z-Image-Turbo",
    transformer=transformer,
    torch_dtype=torch.bfloat16,
)
pipe.to("cuda")

prompt = "A documentary portrait in natural window light, shallow depth of field"
image = pipe(
    prompt=prompt,
    height=1024,
    width=1024,
    num_inference_steps=9,  # 8 DiT forwards
    guidance_scale=0.0,     # required for Turbo
    generator=torch.Generator("cuda").manual_seed(42),
).images[0]
image.save("redrob-image.png")

Optional: pipe.enable_model_cpu_offload() on smaller GPUs.

Quick start (ComfyUI)

File	Put under	Source
`redrob-image.safetensors`	`models/diffusion_models/`	this repository
`qwen_3_4b_fp8_mixed.safetensors`	`models/text_encoders/`	Comfy-Org/z_image_turbo
`ae.safetensors`	`models/vae/`	same Comfy-Org pack

UNETLoader -> redrob-image.safetensors
CLIPLoader -> qwen_3_4b_fp8_mixed.safetensors (type: lumina2, ComfyUI loader type for this text encoder)
VAELoader -> ae.safetensors
Sampler: 8 steps, cfg 1, res_multistep / sgm_uniform

Load workflows/redrob-image-api.json for a minimal working graph. The graph zeros out negative conditioning (ConditioningZeroOut); do not expect a negative text prompt to change the image.

Files

Path	Role
`redrob-image.safetensors`	ComfyUI merged UNET (LFS, ~12 GiB)
`transformer/`	Diffusers layout (built at HF upload)
`workflows/redrob-image-api.json`	Minimal ComfyUI API graph
`README.md` / `README.ko.md`	Model card (English / Korean)
`LICENSE`	Apache License 2.0
`NOTICE`	Attribution

transformer/ is not in git. On Hugging Face upload, scripts/push_hf.sh converts the Comfy UNET into Diffusers format (or copies a prebuilt TRANSFORMER_DIR).

Convert a local Comfy UNET yourself:

bash

python scripts/comfy_to_diffusers_zimage.py \
  --input redrob-image.safetensors \
  --output-dir transformer/

License

Apache License 2.0. Copyright Redrob. Built by Janghoon Lee (이장훈). Upstream attribution is in NOTICE. Redistributors keep NOTICE with the weights.

Attribution

Copyright: Redrob (redrob.io)
Built by: Janghoon Lee (이장훈) (@savagemanage)
Repository: redrob-labs/redrob-image
Base model: Tongyi-MAI/Z-Image-Turbo - Apache License 2.0