z-image-txt2img

Z-Image Text-to-Image Workflows

⚠️ Launch flag: Z-Image does not sample correctly under --use-sage-attention (black / garbled output). Launch ComfyUI with --use-pytorch-cross-attention for Z-Image. See comfyui-launch-flags.

Overview

Z-Image is a 6B-parameter image generation model from Alibaba's Tongyi Lab using a Scalable Single-Stream DiT (S3-DiT) architecture. It uses a Qwen text encoder (not CLIP-L/T5). Its VAE shares the Flux VAE architecture (same tensor shapes, so the file is the same 320MB size) but ships different weights — it is NOT byte-identical to Flux's ae.safetensors and must be kept as a separate file (z-image-ae.safetensors) to avoid clobbering the Flux VAE. Two variants:

1. Z-Image Base (and RedCraft finetune) — Full model, supports negative prompts, LoRA training, ControlNet. 10-30 steps.
2. Z-Image Turbo — DMD-distilled, 8-10 steps, no effective negative prompts (CFG baked in).

Models

RedCraft Redzimage DX1 (Installed — Combined Checkpoint)

RedCraft is a Z-Image Base finetune by the RedCraft team. Designed for faster inference than stock Z-Image Base. Uses CheckpointLoaderSimple since it's a combined checkpoint — no need for separate loaders.

Z-Image Turbo (Separate Components — May Need Download)

Z-Image Base (Separate Components — May Need Download)

Conditioning

TextEncodeZImageOmni (Built-in)

For Z-Image separate component loading. Supports reference images via CLIP Vision:

Required Inputs:
  - clip: CLIP
  - prompt: STRING (multiline)
  - auto_resize_images: BOOLEAN (default true)

Optional Inputs:
  - image_encoder: CLIP_VISION (for reference images)
  - vae: VAE
  - image1-3: IMAGE (up to 3 reference images)

Outputs:
  [0] CONDITIONING

CLIPTextEncode (For RedCraft Checkpoint)

When using CheckpointLoaderSimple, standard CLIPTextEncode works since the checkpoint bundles the correct tokenizer:

{
  "class_type": "CLIPTextEncode",
  "inputs": { "clip": ["<checkpoint>", 1], "text": "<prompt>" }
}

Sampler Settings

RedCraft DX1

Z-Image Turbo

Z-Image Base (Two-Stage)

Stage 1 — Primary generation:

Stage 2 — Detail refinement (optional img2img pass):

Negative Prompts

RedCraft / Z-Image Base

Supports negative prompts at CFG > 1.0:

3D, ai generated, semi realistic, illustrated, drawing, comic, digital painting, 3D model, blender, video game screenshot, screenshot, render, high-fidelity, smooth textures, CGI, masterpiece, text, writing, subtitle, watermark, logo, blurry, low quality, jpeg, artifacts, grainy

Z-Image Turbo

Negative prompts are not effective — CFG is baked in via distillation. Use the positive prompt to guide away from unwanted elements instead.

Recommended positive-side avoidance template:

over-smooth skin, plastic skin, doll face, anime, CGI, waxy texture, blurry face, fake pores, exaggerated makeup, over-sharpening, unrealistic symmetry, flat lighting, low detail skin, extra fingers, distorted anatomy

Resolutions

Dimensions must be divisible by 16.

LoRA System

ZImageTurbo LoRAs

Located in loras/ZImageTurbo/ with subfolders:

• style/ — Style LoRAs (e.g., TurboPussyZ_v2.safetensors)
• concept/ — Concept LoRAs (e.g., body from below.safetensors, ZITnsfwLoRA.safetensors)
• character/ — Character LoRAs (e.g., NSFW_master_ZIT_000008766.safetensors)
• action/ — Action LoRAs

Use with Z-Image Turbo base model. Typical LoRA strength: 0.6–1.0.

ZImageBase LoRAs

Located in loras/ZImageBase/ with subfolders:

• style/ — Style LoRAs (e.g., NSGIRL-Z-Image-LoRA-By-MM744.safetensors)
• concept/ — Concept LoRAs

Use with Z-Image Base or RedCraft. Typical LoRA strength: 0.6–1.0.

Z-Image-Aesthetic-Base v1

General aesthetic improvement LoRA:

• File: Z-Image-Aesthetic-Base v1.safetensors (352MB)
• Settings: euler_ancestral + beta, 30 steps, CFG 4, strength 0.6–1.0

Applying LoRAs

{
  "class_type": "LoraLoader",
  "inputs": {
    "model": ["<checkpoint_or_unet>", 0],
    "clip": ["<checkpoint_or_clip>", 1],
    "lora_name": "ZImageTurbo\\style\\TurboPussyZ_v2.safetensors",
    "strength_model": 0.8,
    "strength_clip": 0.8
  }
}

Note: When using CheckpointLoaderSimple for RedCraft, model output is index 0 and CLIP output is index 1. When stacking multiple LoRAs, chain them sequentially.

ControlNet

ZImageFunControlnet (Built-in)

Experimental built-in node for Z-Image ControlNet. Patches the model with a control signal:

Required Inputs:
  - model: MODEL
  - model_patch: MODEL_PATCH (from ControlNet loader)
  - vae: VAE
  - strength: FLOAT (default 1.0, range -10 to 10)

Optional Inputs:
  - image: IMAGE (reference/control image)
  - inpaint_image: IMAGE
  - mask: MASK

Outputs:
  [0] MODEL (patched)

Z-Image-Turbo-Fun-Controlnet-Union

A unified ControlNet supporting multiple condition types:

• Canny, HED, Depth, Pose, MLSD
• Strength: 0.65–0.80 (v2.1 recommended range)
• Best paired with res_2s, res_5s, or res_2m samplers + beta57 scheduler

Complete Workflow: RedCraft DX1 (Fast, 10-Step)

{
  "1": { "class_type": "CheckpointLoaderSimple", "inputs": { "ckpt_name": "redcraftRedzimageUpdatedJAN30_redzibDX1.safetensors" }},
  "2": { "class_type": "CLIPTextEncode", "inputs": { "clip": ["1", 1], "text": "<positive prompt>" }, "_meta": { "title": "Positive" }},
  "3": { "class_type": "CLIPTextEncode", "inputs": { "clip": ["1", 1], "text": "" }, "_meta": { "title": "Negative" }},
  "4": { "class_type": "EmptyLatentImage", "inputs": { "width": 1024, "height": 1024, "batch_size": 1 }},
  "5": { "class_type": "KSampler", "inputs": {
    "model": ["1", 0],
    "positive": ["2", 0],
    "negative": ["3", 0],
    "latent_image": ["4", 0],
    "seed": 42, "steps": 10, "cfg": 1, "sampler_name": "euler", "scheduler": "simple", "denoise": 1
  }},
  "6": { "class_type": "VAEDecode", "inputs": { "samples": ["5", 0], "vae": ["1", 2] }},
  "7": { "class_type": "SaveImage", "inputs": { "images": ["6", 0], "filename_prefix": "redcraft" }}
}

Complete Workflow: RedCraft DX1 with LoRA Stack

{
  "1": { "class_type": "CheckpointLoaderSimple", "inputs": { "ckpt_name": "redcraftRedzimageUpdatedJAN30_redzibDX1.safetensors" }},
  "2": { "class_type": "LoraLoader", "inputs": {
    "model": ["1", 0], "clip": ["1", 1],
    "lora_name": "Z-Image-Aesthetic-Base v1.safetensors",
    "strength_model": 0.8, "strength_clip": 0.8
  }},
  "3": { "class_type": "LoraLoader", "inputs": {
    "model": ["2", 0], "clip": ["2", 1],
    "lora_name": "ZImageBase\\style\\NSGIRL-Z-Image-LoRA-By-MM744.safetensors",
    "strength_model": 0.7, "strength_clip": 0.7
  }},
  "4": { "class_type": "CLIPTextEncode", "inputs": { "clip": ["3", 1], "text": "<positive prompt>" }},
  "5": { "class_type": "CLIPTextEncode", "inputs": { "clip": ["3", 1], "text": "<negative prompt>" }},
  "6": { "class_type": "EmptyLatentImage", "inputs": { "width": 896, "height": 1152, "batch_size": 1 }},
  "7": { "class_type": "KSampler", "inputs": {
    "model": ["3", 0],
    "positive": ["4", 0],
    "negative": ["5", 0],
    "latent_image": ["6", 0],
    "seed": 42, "steps": 30, "cfg": 4, "sampler_name": "euler", "scheduler": "simple", "denoise": 1
  }},
  "8": { "class_type": "VAEDecode", "inputs": { "samples": ["7", 0], "vae": ["1", 2] }},
  "9": { "class_type": "SaveImage", "inputs": { "images": ["8", 0], "filename_prefix": "redcraft_lora" }}
}

Prompt Style

Natural language descriptions work best (uses Qwen LLM tokenizer, not CLIP):

Good: "Professional headshot of a confident businesswoman in her 30s, natural makeup, soft studio lighting, neutral gray background, sharp focus on eyes, Canon EOS R5"
Bad: "masterpiece, best quality, 1girl, businesswoman, studio"

VRAM Considerations

• Always clear_vram before switching to Z-Image from another model family
• RedCraft is one of the most VRAM-efficient quality models available

Tips

1. RedCraft DX1 with 10 steps / CFG 1.0 is surprisingly fast and high quality for quick iteration
2. For maximum sharpness with Turbo LoRAs, use dpmpp_sde + beta scheduler
3. The Z-Image-Aesthetic-Base v1 LoRA at 0.6–0.8 strength noticeably improves output quality across all Z-Image Base variants
4. Z-Image excels at photorealistic human generation — it's the go-to for portrait and fashion photography
5. When switching between Turbo and Base LoRAs, use the matching base model variant