<h2 id="please-read">please read</h2>V2.1<h3 id="v2.1-update-notes:">V2.1 Update Notes:</h3><ul><li>Fixed: Added the missing image connection to the VQA node that was forgotten in the previous version.</li><li>Optimized: Refined the QwenVQA prompts and optimized the overall VQA settings.</li></ul>You can delete or modify the breast type tags listed in Node 110 according to your preference to control the breast type output by the VQA.<ul><li>⚠️ VERY IMPORTANT, PLEASE NOTE: In your ComfyUI Launcher settings, you MUST enable the SDP optimization scheme to utilize the <code>sdpa</code> acceleration in the VQA node. Using <code>xFormers</code> will cause the system to freeze/hang.</li></ul>V2.1更新内容：之前图片忘了连接到VQA节点。优化QWENVQA提示词，优化VQA设置。可以根据喜好删除或修改节点110中列举的乳房类型tag以控制VQA输出的乳房类型。<h2 id="comfyuisdpvqasdpaxformers">非常重要，注意！comfyui启动器设置，开启SDP方案才能使用VQA节点的sdpa加速，使用Xformers会卡死。</h2><h2 id="v1.0">V1.0</h2>For basic Anima img2img with ControlNet, just as I've heard, the value map (grayscale) yields better results than the depth map. I'm not sure if the standard Illustrious model is also better suited for value maps, but I feel the poses are more accurate.The ControlNet was downloaded from here: <a target="_blank" rel="ugc" href="https://huggingface.co/kohya-ss/Anima-LLLite/tree/main">https://huggingface.co/kohya-ss/Anima-LLLite/tree/main</a>It's paired with a weighted LoRA loader, a batch image loader, JoyTag, WD14, a text concatenate node, a save image node that automatically names and compresses based on the model and time, and a 5-way ControlNet switch node.For the scheduler, using 'normal' makes skin textures more realistic, while 'simple' makes them cleaner基本的anima图生图，搭配controlnet，和我听说的一样，明度图比深度图效果更好。我不知道普通的光辉模型是不是也是更适合用明度图，我感觉动作更准确。controlnet是从这里下载的<a target="_blank" rel="ugc" href="https://huggingface.co/kohya-ss/Anima-LLLite/tree/main">https://huggingface.co/kohya-ss/Anima-LLLite/tree/main</a>搭配有权重lora加载器，批量图像加载器，joytag，WD14，文本合并节点，以模型和时间自动命名并压缩图片的保存节点，5种controlnet切换节点。调度器用normal皮肤质感更写实，simple更干净关于QWEN VQA版本Fixed the default prompts and added QWEN VQA along with a text switcher. Setting 1 is for direct output from joytag, and 2 is for QWEN VQA.The recommended settings for ControlNet and Denoising Strength are: Luminance map, 0.6, 0.6, 0.8. Based on my testing, these are the best settings when not using QWEN VQA. Unlike illustrious, setting the ControlNet weight to 1 will cause the image to collapse, and setting the Denoising Strength to 1 will also cause the image to break.If using QWEN VQA, it takes 150 seconds to generate an image with a long edge of 1680 pixels.Note: Words related to 'occlusion' (or 'covering') in the negative prompts might result in the removal of some clothing.修正了默认提示词，添加了千问VQA,和文本切换器，1是joytag直接输出，2是QWEN VQA，controlnet和重绘幅度 的推荐设置为明度图，0.6,0.6,0.8。经过我的测验，不使用QWEN VQA时，这是最好的设置。它和光辉不同，controlnet为1画面会崩坏，重绘幅度为1画面也会崩坏。如果使用千问VQA，生成一张长边1680的图片需要150秒。负面提示词中的遮挡可能会去除一些衣物。关于修复QWEN VQA的bug看我之前的帖子fix QWEN VQA list bug<a target="_blank" rel="ugc" href="https://civarchive.com/models/2830556/fix-qwen-vqa-output-list-bug">https://civarchive.com/models/2830556/fix-qwen-vqa-output-list-bug</a>I've found that if you use VQA, it's better to use Line or Canny for ControlNet, and the strength can be set to 1, 1 or 1, 0.8.我发现如果使用了VQA，controlnet用line或者canny更好，强度可以调为1,1或者1,0.8 。

Files with SHA256: 206abb2e00881e232e9c69453fab809db8a2a4f62d43f1e419252044448c65f1