are transparent))

Model-based Pixel Upscale Workflow

Multiple ControlNets

对脸部等进行化妆，细节增强处理，虽然都是变漂亮，但和美颜磨皮消除细节的过程刚刚相反。相信大家也应该更喜欢这个。

针对Animagine XL V3.1 Lora训练活动而推出的一个简易的Lora批量测试工作流，请在load Lora节点处替换待测试Lora，并调整提示词（绿色节点为正向提示词，红色节点为负向提示词）、出图大小及数量（可在Empty Latent Image节点调整）等相关参数，有相关问题可在评论区提出

Hires fix 只是以较低的分辨率创建图像，将其放大，然后通过 img2img 发送。请注意，在 ComfyUI 中，txt2img 和 img2img 是同一个节点。Txt2Img 是通过将空图像传递到具有最大降噪的采样器节点来实现的。

对脸部等进行化妆，细节增强处理，虽然都是变漂亮，但和美颜磨皮消除细节的过程刚刚相反。相信大家也应该更喜欢这个。

本工作流是SVD的基础模板，主要面向新手朋友。SVD是Stable Diffusion官方发布的生成式视频模型Stable Video Diffusion。只需上传一张图片，调整参数，即可生成一段几秒钟的视频。在本工作流中，上传任意尺寸图片，可生成3秒钟的长边为768像素的MP4视频，消耗约40点算力。参数讲解：基础参数（一般不用改）：Width: 生成视频的宽度；Height: 生成视频的高度；尺寸最好与上传的图片一致，建议在1500以内，为了节省算力，本工作流通过缩放图片，控制在768以内。Video frames: 生成的运动总帧数，建议最大选择25帧FPS: 视频每秒刷新多少张图像，6或者8运动参数（重点修改调试）：CFG: 越接近1，初始图像对画面的控制能力也就越小，建议3左右Motion bucketid: 数值越大运动幅度越大，建议100以内，不超过200Augmentation level: 添加到图像的噪声量，数值越大变化越大，建议0.1以内，不超过0.5CFG、Motion bucketid、Augmentation level要根据实际情况动态调整。

WAN2.1+ControlNet+VACE+丝滑跳舞视频生成，请注意这种跳舞视频模式，不同于传统的姿态控制，是一种自由度更高，且能产生不同效果的更能发挥AI自主创造的图生视频流程，所有的参数或需要说明的注意事项均在工作流程的备注信息里给予了建议，大家可以自己测试，用自己觉得更合理的方式来进行AI视频创作。必须提醒一下，如果用了FusionX加速lora，那么在后面的采样设置中，CFG值应该设置为1，大家可以试试看。

Hi someone, if you are into filmmaking then this is a workflow that might be worth considering with auto prompt from image, export to video by Wan2.1 combined with audio by LLMAudio

海艺的XL出图点模糊，出图后用sd图生图可以清晰一点。各种风格转换。不同模型作用不同。自己更换发挥。可以在sd1.5换上自己喜欢的人物lora。两个条件控制足够了，可以更换。封面是二次元旗袍转真人。

人によっては卒倒しかねない、めちゃ細かいワークフローです。使いやすいように変更箇所はなるべく纏めてあります。

The Anytest ControlNet seems to be some kind of combination of ControlNet models, but no one knows which ones or how it works exactly, but is definitely the best model.It is very good at keeping the pose, proportions and of the input image, while changing style and details. No preprocessing needed.Use a strength of 0.50-0.60 if you are using sketches or non anime pics, of 0.70-0.80 if using anime pics or keeping the original style.Img2img is possible, but may require some tweaking.

最新的FLUX大模型结合lora和Controlnet用于图片的不同风格转绘。 This workflow is all built on top of FLUX. Xlabs-flux-controlnetV3 is great. FLUX's lora is also great.The details are optimized through subsampling, and the final image is of very high quality. （这套工作流全是基于FLUX搭建的。Xlabs-flux-controlnetV3很棒。FLUX的lora也很棒。通过二次采样进行细节优化最终出图质量非常高。） TEST： Black Myth Wukong style repainted（黑神话悟空风格转绘） You need to upgrade to the latest Comfyui, and the nodes that appear can basically be installed directly in the manager（需要升级到最新的Comfyui，出现的节点基本都可以在manger里面直接安装。） flux： https://huggingface.co/black-forest-labs/FLUX.1-dev/tree/main Xlabs: https://github.com/XLabs-AI/x-flux controlnet: https://huggingface.co/XLabs-AI (v3) lora： https://civitai.com/models/681691 https://www.liblib.art/modelinfo/3945f0e21a4b4bb5965b4319d927dd4f?from=search https://www.liblib.art/modelinfo/f00b5b28b87b4935a99f651499495ffc?from=search Xlabs-lora： https://huggingface.co/XLabs-AI/flux-lora-collection/tree/main

(0607)MinorFix Background will be better in Magmix10 This workflow uses the SD 1.5 IPAdapter and OpenPose for the character pose. It requires the latest version of the IPAdapter Node and uses new weights "weak input" to keep most of the clothes' appearance (looks better than the ootd node in some situation). If you want to reproduce the results like the example picture, you can get the model you like from Civitai, provide the character pose reference, face image, clothes, and background to complete it in this workflow. The webp picture provided in Asset shows the effect of the reference picture under each type of weight. WD14 About editing WD14 tags. In FinalPrompt, if you see an unwanted label, you can add it to the blue grid (Unwanted) and add the required label to the red grid. Some users cannot use WD14, so they can bypass the WD14 node and enter prompt words that match their clothes and background in the red box. WD14 is not a necessary condition for the process, it just helps you with the prompt words, but there are often detection errors. Entering the prompts yourself can sometimes produce better results. == 这个工作流使用了SD1.5的IPAdapter与Openpose达成基本换衣的效果。它使用了IPAdapter的权重 "weak input"用以尽量维持衣服的外观 (某些情况下它比OOTD还要好，特别是在人物姿势与衣服完全不同的时候). 如果你想产生像是范例那样的结果，你可以从Civitai取得模型，提供背景、人物动作、衣服、支持上身与下身的衣服，以及背景。 WebP档案呈现IPAdapter在各种权重下，衣服的表现。 关于编辑WD14的标签。 在FinalPrompt中，若看到不要的标签可以将它加入蓝色的格位排除(Unwanted)，并在红色的格位加入需要的标签。 有些使用者无法使用WD14，可将WD14节点bypass掉，并自己在红色格中，输入符合衣服与背景的提示词。 WD14并非流程的必要条件，只是协助你进行提示词，但常有侦测错误，自己输入提示有时候会有比较好的结果。 目前这个工作流无法呈现衣服的细节，但AI换衣的理想是在哪种人物姿势下都能够穿上相对应的衣服，这个做法只是让它比较像一些而已。

# Flux 📌 Face + Upscale by 🚩 Introduction to Workflow Flux 📌 Face + Upscale by 🚩 It is a consistent magnification workflow developed by Soil Sight Studio, mainly used for image processing and generation, particularly performing well in facial extraction and image super-resolution. ##1、 Model correlation 1. Rich types of models: Workflow involves multiple models, including Unet model, Clip encoder, VAE, etc. These models each undertake different functions, such as the Unet model for core image processing, Clip encoder for text image association, and VAE for image encoding and decoding. 2. Model source and placement: The model is mainly sourced from platforms such as Civitai and Hugging Face, and has clear placement path guidance. For example, the Flux Dev Q5 GGUF model is placed in the models/unet/directory, while the clip_1 model is placed in the models/clip/directory. At the same time, detailed download links are provided to facilitate users in obtaining and configuring the required models. ##2、 Workflow Process 1. * * Image Input and Preprocessing * *: Firstly, the image is loaded through the LoadImage node, and then the AutoDropFaces node may be used for face extraction, automatically detecting and cropping facial regions in the image to prepare for subsequent processing. Image scaling operations can also be performed, such as the ImageScaleToMegapixels node, which can scale the image at a specified magnification. 2. Model loading and setting: Load Unet model, VAE model, Clip model, etc. separately, and set and obtain the model and related parameters through SetNode and GetNode nodes. For example, loading the Unet model through the UNETLoader node, loading the VAE model through the VAELoader node, and loading the Clip model through the DualCLIPLOADer node. 3. Image Generation and Processing: Using the KSampler node for image sampling, generate new images based on the input model, positive conditions, negative conditions, and Latent images. The ControlNetApply Advanced node applies the ControlNet model to further control and optimize the image, adjusting the generation effect of the image by setting different parameters. 4. * * Post processing and output * *: The generated image can be previewed through the PreviewiImage node or saved locally using the SaveImage node. During this process, some post-processing operations can also be performed, such as the PlaySound | pysssss node playing specified sound prompts when image generation is complete. ##3、 Copyright Notice This workflow is independently developed by SoilSighStudio and is not for sale. Copyright✦ SoilSighStudio ✦

Index 0 Error: A couple of people have experienced an error "There's nothing in Index 0" and I had it happen to me today. Has the Florence Large model selected, worked fine for Inpainting an image, I then swapped to a different image and that's when the error happened. Changed the model to base, went through fine. Tried some other versions of Florence and the error came back. Changed image and went back to Large and no error. Seems like some versions of Florence are failing to pick up the prompt for what needs to be masked and giving empty data to the co-ordinates node, hence the error. So if you do get it, try base which seems to be the least affected by it, or try changing your mask prompt. Worst case scenario if none of that works is to switch to manual mask and try auto again with the next image you want to Inpaint. Update 4: Added an invert Mask node. If you put person into the Florence prompt for the auto masking and enable the invert mask node, you can use it to Inpaint the background. Update 3: Added a couple of LoRAModelLoaderOnly nodes. Update 2: Added the ModelSamplingFlux node, not quite sure how I missed that. In addition, swapped out the Load and Resize Image node for a standard load image node and also added an image switch node to the Manual mask group. A couple of times I switched to that but forgot to add the picture to the load image in the Florence group, so the result was not what I expected. Seemed silly having to load the same image twice, so I've put that with the switch to swap masks. Have both set to true if you're using Florence and SAM2 for auto masking, false if you need to mask something by hand.

[Workflow introduction] online running URL ： https://www.runninghub.ai/post/1851088322331561986 After opening the workflow, you need to upload two images. One is a face image, and a frontal portrait is preferred here; the other is a clothing image. First, use Runninghub Captioner to reverse the prompt words, then use Apply PuLID Flux to perform facial migration, and then combine the FLUX model to generate a new image. Then, use the SUPIR model to optimize the image. Finally, generate the ID photo layout effect through the HivisionNode node. The built-in parameters have been set, and you are welcome to experience it. [Usage scenario] You only need to upload two images to generate ID photos with different outfits. You can change them according to different needs to generate a variety of images. Convenient, fast and practical, come and try it! [Key Node] PULD FLUX [Model Version] Flux Model Name: flux1-dev-fp8-Kijai.safetensors SUPIR Model Name: SUPIR-vOQ.ckpt [LoRA Model] None [ControlNet Application] None [K Sampler] CFG: 1.0 Scheduler: beta Noise Reduction: 1.00 Sampler: euler

联系我：15884404095（微信同号）。1.定制专属工作流（根据具体需求制作，更实用）2.定制Lora（让生成图片的模特、产品、风格等保持一致性）3.获取AI绘画视频教程（可在下面的更多中找到教程工作流）4.定制AI视频（根据具体需求生成特定视频）更多类似实用工作流可在个人主页获取有其他需求可以评论出来，会根据需求考虑更新功能：手部修复使用方法：在绿色框处上传要修复的手部图片，涂抹需要修复的手部区域，右上角点击开始生图就可以了。如果修复的图片细节不完美可以在更多工作流使用高清放大，完善细节。