comfyui-hires-fix

ComfyUI 高清修复 (Hires Fix)

先生成低分辨率图，放大后再重采样，获得高质量大图。

---

预注入变量

run_python 已注入：submit_workflow, folder_paths, save_preview

---

核心概念

高清修复流程：

低分辨率生成 → 放大模型 → 高分辨率重采样 → 最终输出
     512²            2x            1024²              1024²

为什么需要 Hires Fix？

• 直接生成 1024² 容易出畸形
• 先生成 512² 再放大，质量更好
• 可以控制重采样强度（denoise）

---

使用方法

方法 1: 使用便捷函数

from comfyui_hires_fix import build_hires_workflow

wf = build_hires_workflow(
    prompt="masterpiece, best quality, a beautiful landscape",
    negative_prompt="low quality, blurry",
    checkpoint="sdxl_base.safetensors",
    
    # 基础生成参数
    base_width=512,
    base_height=512,
    base_steps=20,
    
    # 放大参数
    upscale_by=2.0,  # 512x512 → 1024x1024
    upscaler="4x-UltraSharp.pth",  # 放大模型
    
    # 重采样参数
    hires_steps=15,
    hires_denoise=0.4,  # 0.3-0.5 较好
    
    seed=42
)

# 执行
result = submit_workflow(wf)
if result.get("images"):
    save_preview(result["images"][0])

方法 2: 分步构建（更灵活）

import random

# 1. 基础生成（低分辨率）
wf = {}
wf["1"] = {"class_type": "CheckpointLoaderSimple", "inputs": {"ckpt_name": "sdxl.safetensors"}}
wf["2"] = {"class_type": "CLIPTextEncode", "inputs": {"text": prompt, "clip": ["1", 1]}}
wf["3"] = {"class_type": "CLIPTextEncode", "inputs": {"text": negative_prompt, "clip": ["1", 1]}}
wf["4"] = {"class_type": "EmptyLatentImage", "inputs": {"width": 512, "height": 512, "batch_size": 1}}
wf["5"] = {
    "class_type": "KSampler",
    "inputs": {
        "model": ["1", 0], "positive": ["2", 0], "negative": ["3", 0], "latent_image": ["4", 0],
        "seed": random.randint(0, 2**32), "steps": 20, "cfg": 7.0,
        "sampler_name": "euler", "scheduler": "normal", "denoise": 1.0
    }
}

# 2. 解码为像素
wf["6"] = {"class_type": "VAEDecode", "inputs": {"samples": ["5", 0], "vae": ["1", 2]}}

# 3. 放大模型
wf["7"] = {"class_type": "UpscaleModelLoader", "inputs": {"model_name": "4x-UltraSharp.pth"}}
wf["8"] = {"class_type": "ImageUpscaleWithModel", "inputs": {"upscale_model": ["7", 0], "image": ["6", 0]}}

# 4. 重新编码为 latent
wf["9"] = {"class_type": "VAEEncode", "inputs": {"pixels": ["8", 0], "vae": ["1", 2]}}

# 5. 高分辨率重采样（关键：denoise < 1.0）
wf["10"] = {
    "class_type": "KSampler",
    "inputs": {
        "model": ["1", 0], "positive": ["2", 0], "negative": ["3", 0], "latent_image": ["9", 0],
        "seed": random.randint(0, 2**32), "steps": 15, "cfg": 7.0,
        "sampler_name": "euler", "scheduler": "normal", "denoise": 0.4  # 关键参数
    }
}

# 6. 最终解码和保存
wf["11"] = {"class_type": "VAEDecode", "inputs": {"samples": ["10", 0], "vae": ["1", 2]}}
wf["12"] = {"class_type": "SaveImage", "inputs": {"images": ["11", 0], "filename_prefix": "hires"}}

result = submit_workflow(wf)