【工作流workflow】Z-Image分階段採樣 Z-Image Phased sampling
独自懵逼
Z-Image Base模型具有更好的提示詞理解能力，但是採樣慢；Z-Image Turbo模型美學更好且採樣快，但是受蒸餾影響提示詞理解力較差。所以可以先用Z-Image Base進行構圖採樣，再用Z-Image Turbo完成後續採樣豐富細節。因為第一階段採樣只需10步，提高了抽卡效率，後續採樣也僅需6步即可獲得Z-Image Turbo的美學偏好和細節。
複製以下代碼另存為JSON文檔，然後導入ComfyUI即可獲得工作流。
The Z-Image Base model offers better prompt understanding but samples slowly. The Z-Image Turbo model delivers superior aesthetics and much faster sampling, but its prompt adherence is weaker due to distillation.
A good approach is to first use Z-Image Base for the initial composition sampling (layout & structure), then switch to Z-Image Turbo to finish the sampling and enrich the details. Since the first stage only needs 10 steps, this greatly improves "roll efficiency," and the second stage requires just 6 more steps to inherit Turbo's aesthetic preferences and fine details.
Copy the code below, save it as a JSON file, then import it into ComfyUI to get the workflow.
{
"1": {
"inputs": {
"action": "append",
"tidy_tags": "yes",
"text_a": "",
"text_b": "",
"text_c": ""
},
"class_type": "StringFunction|pysssss",
"_meta": {
"title": "String Function 🐍"
}
},
"20": {
"inputs": {
"seed": -1
},
"class_type": "Seed (rgthree)",
"_meta": {
"title": "Seed (rgthree)"
}
},
"21": {
"inputs": {
"aspect_ratio": "Qwen 3:4 (1140x1472)",
"batch_size": 1,
"custom_width": 2048,
"custom_height": 2048
},
"class_type": "QwenImageLatentAspectRatio",
"_meta": {
"title": "Qwen-Image 图像比例"
}
},
"28": {
"inputs": {
"shift": 3,
"model": [
"48",
0
]
},
"class_type": "ModelSamplingAuraFlow",
"_meta": {
"title": "ModelSamplingAuraFlow"
}
},
"31": {
"inputs": {
"text": [
"1",
0
],
"clip": [
"38",
0
]
},
"class_type": "CLIPTextEncode",
"_meta": {
"title": "CLIP Text Encode (Prompt)"
}
},
"35": {
"inputs": {
"samples": [
"49",
0
],
"vae": [
"43",
0
]
},
"class_type": "VAEDecode",
"_meta": {
"title": "VAE Decode"
}
},
"36": {
"inputs": {
"text": "单膝跪地，deformed, malformed, mutated, disfigured, ugly, bad anatomy, bad proportions, gross proportions, extra limbs, missing limbs, extra arms, extra legs, fused fingers, too many fingers, missing fingers, long neck, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, poorly drawn feet, cloned face, duplicate, morbid, mutilated, out of frame, body out of frame, cut off, lowres, low quality, blurry, pixelated, grainy, jpeg artifacts, oversaturated, underexposed, overexposed, bad art, beginner, amateur, distorted face, watermark, signature, text, error, draft, cartoon, anime, 3d render, painting",
"clip": [
"38",
0
]
},
"class_type": "CLIPTextEncode",
"_meta": {
"title": "CLIP Text Encode (Prompt)"
}
},
"38": {
"inputs": {
"clip_name": "text_encoders/qwen_3_4b.safetensors",
"type": "lumina2",
"device": "default"
},
"class_type": "CLIPLoader",
"_meta": {
"title": "Load CLIP"
}
},
"43": {
"inputs": {
"vae_name": "Flux/ae.safetensors"
},
"class_type": "VAELoader",
"_meta": {
"title": "Load VAE"
}
},
"45": {
"inputs": {
"text": "deformed, malformed, mutated, disfigured, ugly, bad anatomy, bad proportions, gross proportions, extra limbs, missing limbs, extra arms, extra legs, fused fingers, too many fingers, missing fingers, long neck, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, poorly drawn feet, cloned face, duplicate, morbid, mutilated, out of frame, body out of frame, cut off, lowres, low quality, blurry, pixelated, grainy, jpeg artifacts, oversaturated, underexposed, overexposed, bad art, beginner, amateur, distorted face, watermark, signature, text, error, draft, cartoon, anime, 3d render, painting",
"clip": [
"38",
0
]
},
"class_type": "CLIPTextEncode",
"_meta": {
"title": "CLIP Text Encode (Prompt)"
}
},
"46": {
"inputs": {
"text": [
"1",
0
],
"clip": [
"38",
0
]
},
"class_type": "CLIPTextEncode",
"_meta": {
"title": "CLIP Text Encode (Prompt)"
}
},
"47": {
"inputs": {
"filename_prefix": "Step_1_Img",
"images": [
"35",
0
]
},
"class_type": "SaveImage",
"_meta": {
"title": "Save Image"
}
},
"48": {
"inputs": {
"unet_name": "Z-Image/z_image_bf16.safetensors",
"weight_dtype": "default"
},
"class_type": "UNETLoader",
"_meta": {
"title": "Load Diffusion Model"
}
},
"49": {
"inputs": {
"add_noise": "enable",
"noise_seed": [
"20",
0
],
"steps": 30,
"cfg": 4,
"sampler_name": "res_multistep",
"scheduler": "simple",
"start_at_step": 0,
"end_at_step": 10,
"return_with_leftover_noise": "disable",
"model": [
"28",
0
],
"positive": [
"31",
0
],
"negative": [
"36",
0
],
"latent_image": [
"21",
0
]
},
"class_type": "KSamplerAdvanced",
"_meta": {
"title": "KSampler (Advanced)"
}
},
"51": {
"inputs": {
"image": "test.png"
},
"class_type": "LoadImage",
"_meta": {
"title": "Load Image"
}
}
}
Replies (0)
請登錄