Loop Upscale - Automated upscale workflow for different scenarios版本Flux.1_v2.0 (ID: 1034012)

Buy me a coffee ☕

Workflow Explained (in Chinese):

Flux.1_v2.0: https://www.bilibili.com/video/BV19GDjYiErK/

Flux.1_v1.0: https://www.bilibili.com/video/BV16nSaYpE5A/

Features

This workflow adopts an iterative approach, automatically determining the necessary parameters and process based on the image size. It upscales by a factor of 2 each time, repeating the process until the target size is reached.

　

Solved Issues

Compared to common upscaling workflows, this workflow primarily addresses the following issues:

1. When the original image size is too small, forcing an upscale results in significant deviations from the original.

2. Different sizes of original images require specific parameter adjustments or even different upscaling processes to achieve optimal results.

3. When using prompts to guide AI in Tiled upscaling, the AI may sometimes experience "hallucinations" generating unwanted elements.

　

Requirement

• Models

  1. Any Flux.1 unet model as the base model

     .\ComfyUI\models\unet\

  2. VAE: ae.safetensors

     .\ComfyUI\models\vae\

  3. CLIP: t5xxl_fp8_e4m3fn.safetensors, clip_l.safetensors

     .\ComfyUI\models\clip\

  4. Upscale: 2xNomosUni_span_multijpg_ldl

     .\ComfyUI\models\upscale_models\

  5. Torbo LoRA: FLUX.1-Turbo-Alpha

     Download and rename the file, put it into the path:

     .\ComfyUI\models\loras\Flux.1\

  6. (Optional) LoRA that makes the visual style more realistic, it can be replaced by other LoRA: zyd232ChineseGirl_Flux1D_v1_2

     .\ComfyUI\models\loras\Flux.1\1_Character-Female\

  7. ControlNet: Flux.1-dev-Controlnet-Upscaler

     Download and rename the file, put it into the path: .\ComfyUI\models\controlnet\Flux.1\

• Custom Nodes

  • After importing the workflow for the first time, please use the "Install Missing Custom Nodes" button in ComfyUI Manager menu to get all the nodes needed.

　

Tips

1. Generally, you only need to adjust the "Target Size" in the "Input" group and select the image to upscale.

2. The higher the "Similarity" value, the closer the result is to the original image; the lower the "Similarity" value, the more detailed the result, but with reduced resemblance to the original. Generally, when upscaling portraits, the "Similarity" value should be between 0.97 and 0.99 to preserve facial features. In other cases, you can try lower values, as 0.85 to 0.98 can still yield good results.

　

__________________________________________________________________

支持我的工作 ☕

工作流详解：

Flux.1_v2.0: https://www.bilibili.com/video/BV19GDjYiErK/

Flux.1_v1.0: https://www.bilibili.com/video/BV16nSaYpE5A/

特点

这个工作流采用了循环的办法，根据图片的尺寸，自动判断放大时所需的参数和流程，每次放大2倍，不断循环，直至目标尺寸,

　

解决的问题

对比目前常见的放大流程，这个工作流主要解决了以下问题：

1. 原图尺寸太小时，若强行一次性放大，结果和原图的偏差较大,

2. 对于不同尺寸的原图，都要做针对性的调参，甚至要用不同的放大流程，才能得到较好的结果,

3. 使用提示词引导，并进行Tile分块放大时，AI有时候会产生“幻觉”，绘制出我们不想要的东西,

　

必需

• 模型

  1. 任意 Flux.1 unet 模型

     .\ComfyUI\models\unet\Flux.1\

  2. VAE: ae.safetensors

     .\ComfyUI\models\vae\

  3. CLIP: t5xxl_fp8_e4m3fn.safetensors, clip_l.safetensors

     .\ComfyUI\models\clip\

  4. 放大模型: 2xNomosUni_span_multijpg_ldl

     .\ComfyUI\models\upscale_models\

  5. 加速 LoRA: FLUX.1-Turbo-Alpha

     下载并重命名，放到以下路径中:

     .\ComfyUI\models\loras\Flux.1\

  6. (可选) 使风格更加真实的 LoRA，可以用其他 LoRA 代替: zyd232ChineseGirl_Flux1D_v1_2

     .\ComfyUI\models\loras\Flux.1\

  7. ControlNet: Flux.1-dev-Controlnet-Upscaler

     下载并重命名，放到以下路径中:

     .\ComfyUI\models\controlnet\Flux.1\

• 自定义节点

  • 首次导入工作流程后，请使用 ComfyUI Manager 菜单中的 “Install Missing Custom Nodes” 按钮获取所需的所有节点,

　

提示

1. 通常来说，你只需要在“Input”组中修改“Target Size”，并选择要放大的图片,

2. “Similarity”值越大，结果和原图越接近；“Similarity”值越小，结果的细节越多、但是会降低和原图的相似性,通常来说，对人物进行放大时，“Similarity”值应为0.97-0.99，以保持人物的相貌,其他情况下，你可以尝试更低的值，0.85-0.98 都可以得到不错的结果,

描述:

1. Added a new toggle "Image to Prompt".

2. Added a new toggle "Tiny Details"; a corresponding parameter was also added.

3. Refactored the workflow to reduce coupling and make the structure more intuitive, making it easier to modify functionalities.

4. Added a counter to record the number of times the image has undergone the looping process. When the counter is at 0, the image will be forced to undergo a ControlNet redrawing after being doubled in size, regardless of its original size. This eliminates the need to downscale larger-than-1024px images to improve upscale quality, as was done in versions v1 and v1.1. Thus, the original image can be processed automatically without loss of detail, regardless of its size.

—————————

1. 新增了一个开关，用于控制是否使用“图片倒推提示词引导重绘”。

2. 新增了一个开关，用于控制图片循环至大尺寸时是否添加更多小细节；同时添加一个对应的参数，用于控制这些小细节的重绘程度。

3. 重构了工作流，降低了其耦合性，同时使结构更直观，以便于修改功能。

4. 添加了一个计数器，记录图片经历循环的次数。当计数器为0时，无论其尺寸大小，强制其在放大2倍后进行一次ControlNet重绘。这样一来，就不用像v1和v1.1版本那样，在处理大于1024px的图片时、为了放大质量而缩小图片了；如此一来，无论原图大小如何，都能够自动处理，不会损失原图的细节。

训练词语: