/z-image-txt2img

name: z-image-txt2img
description: Build Z-Image txt2img workflows. RedCraft checkpoint, Z-Image Turbo/Base LoRAs, ControlNet, and sampler presets
globs:
  - "**/*.json"

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`](../comfyui-launch-flags/SKILL.md).

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)

| Component | Node | Model | Notes | |-----------|------|-------|-------| | **Checkpoint** | `CheckpointLoaderSimple` | `redcraftRedzimageUpdatedJAN30_redzibDX1.safetensors` | 17GB, bundles UNET+CLIP+VAE |

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, so no separate loaders are needed.

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

| Component | Node | Model | Notes | |-----------|------|-------|-------| | **UNET** | `UNETLoader` | `z_image_turbo_bf16.safetensors` | Not currently installed | | **CLIP** | `CLIPLoader` (type=`qwen_image`) | `qwen_3_4b.safetensors` | Not currently installed | | **VAE** | `VAELoader` | `z-image-ae.safetensors` | 320MB. Flux VAE architecture but different weights — NOT the same file as Flux's `ae.safetensors`. From `Comfy-Org/z_image_turbo` (`split_files/vae/ae.safetensors`) |

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

| Component | Node | Model | Notes | |-----------|------|-------|-------| | **UNET** | `UNETLoader` | `z_image_base_bf16.safetensors` | Not currently installed | | **CLIP** | `CLIPLoader` (type=`qwen_image`) | `qwen_3_4b.safetensors` | Not currently installed | | **VAE** | `VAELoader` | `z-image-ae.safetensors` | 320MB. Flux VAE architecture but different weights — NOT the same file as Flux's `ae.safetensors` |

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

| Preset | Steps | CFG | Sampler | Scheduler | Notes | |--------|-------|-----|---------|-----------|-------| | **Distilled Fast** | 10 | 1.0 | euler | simple | Quick iteration | | **Standard** | 30 | 4.0 | euler | simple | Full quality |

Z-Image Turbo

| Preset | Steps | CFG | Sampler | Scheduler | Notes | |--------|-------|-----|---------|-----------|-------| | Author recommended | 14 | 1.0 | res_2s | simple | CopaxTimeless author pick | | Beauty/fashion | 10 | 1.0 | euler_ancestral | beta | Smooth skin, fashion photography | | **Sharpest** | 10 | 1.0 | dpmpp_sde | beta | Sharpest, most natural (560-image test) |

Z-Image Base (Two-Stage)

Stage 1, primary generation:

| Parameter | Value | |-----------|-------| | Steps | 22 | | CFG | 4.0 (range 4–7) | | Sampler | res_2s | | Scheduler | beta | | Denoise | 1.0 |

Stage 2, detail refinement (optional img2img pass):

| Parameter | Value | |-----------|-------| | Steps | 3 | | CFG | 4.0 | | Sampler | res_2s | | Scheduler | normal | | Denoise | 0.15 |

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

| Aspect | Resolution | Notes | |--------|-----------|-------| | Square | 1024x1024 | Standard | | Square (native) | 1328x1328 | Higher quality at native resolution | | Portrait 3:4 | 896x1152 | | | Portrait 5:8 | 832x1216 | | | Portrait 9:16 | 768x1344 | | | Landscape 16:9 | 1280x720 | |

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 to