2024-08-18发现同样的png info复制进去后，kaggle出图跟示例图有一些细小差别，感觉很奇怪，可能是显存不同导致导致出图有细微差别。发现原因在于，封面图用的是<code>flux1-dev.safetensors</code>，kaggle测试用的是<code>flux1-dev-bnb-nf4.safetensors</code>模型，因此出图效果有细微差别。I noticed slight differences between the Kaggle generated image and the local generated image after copying the same PNG info, it is really weird. Maybe the VRAM differences result in the image-generating differences.The reason for the discovery is that the cover post image uses <code>flux1-dev.safetensors</code>, while the Kaggle test uses the <code>flux1-dev-bnb-nf4.safetensors</code> model, resulting in slight differences in the rendering effect.<hr />2024-08-16我写了个Kaggle上部署Forge的入门指南，可能会有帮助：<a target="_blank" rel="ugc" href="https://sap7x39yn02.feishu.cn/wiki/WPW2wcrOIiudnEkm9Zwc5YBNnPc">‌‍​‍‬​‌‍​​​​​‬​‌​⁠​​⁠​‬⁠​‌​‌‍​​‬​​‌‌​‬‌​​​​​‍​⁠​​‬免费的云端16G显卡跑Flux！Kaggle上部署SD-Forge2.0指南 - Feishu Docs</a>I wrote a tutorial for deploying Forge-webUI on Kaggle, the scripts might be helpful: <a target="_blank" rel="ugc" href="https://sap7x39yn02.feishu.cn/wiki/WPW2wcrOIiudnEkm9Zwc5YBNnPc">https://sap7x39yn02.feishu.cn/wiki/WPW2wcrOIiudnEkm9Zwc5YBNnPc</a><hr />2024-08-16得益于Zhanglvmin大神的努力，sd-webUI-Forge已经支持FLux1-dev和Flux1-schnell模型，我训练了一个基于<code>flux1-dev.safetensors</code>微调的人像LoRA，已经用Forge体验过，发布出来，欢迎大家体验。Thanks to the efforts of the amazing lllyasviel, sd-webUI-Forge now supports the Flux1-dev and Flux1-schnell models. I’ve trained a portrait LoRA based on <code>flux1-dev.safetensors</code> and have already tested it with Forge. I'm releasing it for everyone to try out.概述：<ul><li>lora是基于<code>flux1-dev.safetensors</code>进行训练的，载入base model和lora后，提示词里加入涉及woman的描述，得到的图片基本都会呈现黄亦玫的面部特征。（黄亦玫是一部电视剧里的角色）</li><li>经过Zhanglvmin的更新，<code>flux1-dev</code>上训练的lora也支持了<code>flux1-dev-bnb-nf4.safetensors</code>模型。</li><li>在comfyUI里一样可以在<code>flux1-dev.safetensors</code>基础上使用该lora。</li></ul>Overview:<ul><li>This LoRA was trained based on <code>flux1-dev.safetensors</code>. After loading the base model and the LoRA, adding prompts related to "woman" will usually result in images featuring the facial characteristics of Huang Yimei (a character from a TV series).</li><li>Thanks to lllyasviel’s updates, LoRAs trained on <code>flux1-dev</code> now also support the <code>flux1-dev-bnb-nf4.safetensors</code> model.</li><li>This LoRA can also be used with <code>flux1-dev.safetensors</code> in webUI-Forge and ComfyUI.</li></ul>使用步骤：<ul><li>Lora<ul><li>下载本lora模型后，放入<code>./stable-diffusion-webui-forge/models/Lora</code>里，<code>flux1-dev.safetensors</code>下载后放入<code>./stable-diffusion-webui-forge/models/Stable-diffusion</code>里。</li></ul></li><li>LoRA使用<ul><li>在prompt文本框里加入lora名称<code>&lt;lora:lora_huangyimei_flux1_dev_v1:0.9&gt;</code>和其他描述性的提示词，点击<code>Generate</code>即可使用本lora进行模型推理。<code>0.9</code>是LoRA的权重，也可以换成<code>0~1</code>的任何小数。</li></ul></li></ul>Usage Steps:<ul><li>LoRA:<ul><li>Download the LoRA model and place it in <code>./stable-diffusion-webui-forge/models/Lora</code>. Download <code>flux1-dev.safetensors</code> and place it in <code>./stable-diffusion-webui-forge/models/Stable-diffusion</code>.</li></ul></li><li>Using the LoRA:<ul><li>In the prompt text box, add the LoRA name <code>&lt;lora:lora_huangyimei_flux1_dev_v1:0.9&gt;</code> along with other descriptive prompts. Click "Generate" to use this LoRA for inference. The <code>0.9</code> represents the LoRA weight, which can be adjusted to any decimal between 0 and 1.</li></ul></li></ul>使用细节：<ul><li>推荐尺寸有<code>768*1152</code>,<code>768*1360</code>,<code>832*1216</code>等等。(训练集的尺寸是<code>768*1152</code>和<code>768*768</code>)</li><li>像素如果用<code>512*512</code>，画出的人物手部容易畸形。原因在于画面太小，留给手部的空间很有限，无法在那么小的像素里画出完好的人手。</li><li>出图人物的清晰度不够高的话，可以把迭代步数从20步提升到40步。</li><li>手部如果有问题，可以锁定种子数，微调迭代步数<code>steps</code>。</li><li>训练集里的图片几乎基本都是半身像、牛仔镜头像，没有全身像。因此使用LoRA生成全身图，效果不好。等后续增加了训练集里图片的数量和质量，可以解决这个问题。</li><li>跑图的时候，不是每一次出图都会符合预期。需要先20步迭代或25步迭代跑多张图，发现满意的图了，锁定种子之后，xyz plot测试恰当的迭代步数。</li><li>flux模型对提示词服从度更高，手指出问题的情况也更少，应对复杂提示词、绘制复杂场面时表现更好。因此可以在flux出图之后，去SD精修。</li></ul>Details:<ul><li>Recommended resolutions include <code>768*1152</code>, <code>768*1360</code>, <code>832*1216</code>, etc. (The training set images are in <code>768*1152</code> and <code>768*768</code> resolutions).</li><li>Using a resolution of <code>512*512</code> can result in distorted hand features due to limited canvas space, which makes it difficult to render hands accurately at such low pixel counts.</li><li>If the generated image lacks sharpness, try increasing the iteration steps from 20 to 40.</li><li>If hand features are problematic, lock the seed and fine-tune the iteration steps (<code>steps</code>).</li><li>The training set primarily consists of half-body and cowboy shots, with very few full-body images. As a result, generating full-body images with this LoRA may produce unsatisfactory results. This issue can be improved by expanding the training set with higher-quality images.</li><li>Flux models have higher prompt adherence and fewer issues with hand rendering. They also perform better when handling complex prompts and scenes. Therefore, you can use flux to generate an image first, then refine it with SD (Stable Diffusion).</li><li>When generating images, the results may not always meet your expectations on the first try. Generating many pictures with one prompt by running iterations with 20 or 25 steps. Once you find a satisfactory image, lock the seed and use an XYZ plot to test for the optimal number of iteration steps.”</li></ul>

Files with SHA256: 8a79e8d3dbbda48945fccaa7f01fb03ea3f166bcb24133df1905ca396a7e00c1