multimodalart/ctrl-x
8
1<!doctype html>2<html lang="en">3 4 5<!-- === Header Starts === -->6<head>7 <meta http-equiv="Content-Type" content="text/html; charset=UTF-8">8 9 <title>Ctrl-X</title>10 11 <link href="./assets/bootstrap.min.css" rel="stylesheet">12 <link href="./assets/font.css" rel="stylesheet" type="text/css">13 <link href="./assets/style.css" rel="stylesheet" type="text/css">14</head>15<!-- === Header Ends === -->16 17 18<body>19 20 21<!-- === Home Section Starts === -->22<div class="section">23 <!-- === Title Starts === -->24 <div class="header">25 <div class="logo">26 <a href="https://genforce.github.io/" target="_blank"><img src="./assets/genforce.png"></a>27 </div>28 <div class="title", style="padding-top: 25pt;"> <!-- Set padding as 10 if title is with two lines. -->29 Ctrl-X: Controlling Structure and Appearance for Text-To-Image Generation Without Guidance30 </div>31 </div>32 <!-- === Title Ends === -->33 <div class="author">34 <a href="https://kuanhenglin.github.io" target="_blank">Kuan Heng Lin</a><sup>1</sup>* 35 <a href="https://sichengmo.github.io/" target="_blank">Sicheng Mo</a><sup>1</sup>* 36 <a href="https://bklingher.github.io" target="_blank">Ben Klingher</a><sup>1</sup> 37 <a href="https://pages.cs.wisc.edu/~fmu/" target="_blank">Fangzhou Mu</a><sup>2</sup> 38 <a href="https://boleizhou.github.io/" target="_blank">Bolei Zhou</a><sup>1</sup>39 </div>40 <div class="institution">41 <sup>1</sup>UCLA 42 <sup>2</sup>NVIDIA43 </div>44 <div class="note">45 *Equal contribution46 </div>47 <div class="title" style="font-size: 18pt;margin: 15pt 0 15pt 0">48 NeurIPS 202449 </div>50 <div class="link">51 [<a href="https://arxiv.org/abs/2406.07540" target="_blank">Paper</a>] 52 [<a href="https://github.com/genforce/ctrl-x" target="_blank">Code</a>]53 </div>54 <div class="teaser">55 <img src="assets/ctrl-x.jpg" width="85%">56 </div>57</div>58<!-- === Home Section Ends === -->59 60 61<!-- === Overview Section Starts === -->62<div class="section">63 <div class="title">Overview</div>64 <div class="body"> 65 We present <b>Ctrl-X</b>, a simple <i>training-free</i> and <i>guidance-free</i> framework for text-to-image (T2I) generation with structure and appearance control. Given user-provided structure and appearance images, Ctrl-X designs feedforward structure control to enable structure alignment with the structure image and semantic-aware appearance transfer to facilitate the appearance transfer from the appearance image. Ctrl-X supports novel structure control with arbitrary condition images of any modality, is significantly faster than prior training-free appearance transfer methods, and provides instant plug-and-play to any T2I and text-to-video (T2V) diffusion model.66 <table width="100%" style="margin: 20pt 0; text-align: center;">67 <tr>68 <td><img src="assets/pipeline.jpg" width="85%"></td>69 </tr>70 </table>71 72 <b>How does it work?</b> Given clean structure and appearance latents, we first obtain noised structure and appearance latents via the diffusion forward process, then extracting their U-Net features from a pretrained T2I diffusion model. When denoising the output latent, we inject convolution and self-attention features from the structure latent and leverage self-attention correspondence to transfer spatially-aware appearance statistics from the appearance latent to achieve structure and appearance control. We name our method "Ctrl-X" because we reformulate the controllable generation problem by 'cutting' (and 'pasting') structure preservation and semantic-aware stylization together.73 </div>74</div>75<!-- === Overview Section Ends === -->76 77 78<!-- === Result Section Starts === -->79<div class="section">80 <div class="title">Results: Structure and appearance control</div>81 <div class="body">82 Results of training-free and guidance-free T2I diffusion with structure and appearance control, where Ctrl-X supports a diverse variety of structure images, including natural images, ControlNet-supported conditions (e.g., canny maps, normal maps), and in-the-wild conditions (e.g., wireframes, 3D meshes). The base model here is <a href="https://arxiv.org/abs/2307.01952" target="_blank">Stable Diffusion XL v1.0</a>.83 84 <!-- Adjust the number of rows and columns (EVERY project differs). -->85 <table width="100%" style="margin: 20pt 0; text-align: center;">86 <tr>87 <td><img src="assets/results_struct+app.jpg" width="100%"></td>88 </tr>89 </table>90 <table width="100%" style="margin: 20pt 0; text-align: center;">91 <tr>92 <td><img src="assets/results_struct+app_2.jpg" width="85%"></td>93 </tr>94 </table>95 </div>96</div>97 98<div class="section">99 <div class="title">Results: Multi-subject structure and appearance control</div>100 <div class="body">101 Ctrl-X is capable of multi-subject generation with semantic correspondence between appearance and structure images across both subjects and backgrounds. In comparison, <a href="https://arxiv.org/abs/2302.05543" target="_blank">ControlNet</a> + <a href="https://arxiv.org/abs/2308.06721" target="_blank">IP-Adapter</a> often fails at transferring all subject and background appearances.102 103 <!-- Adjust the number of rows and columns (EVERY project differs). -->104 <table width="100%" style="margin: 20pt 0; text-align: center;">105 <tr>106 <td><img src="assets/results_multi_subject.jpg" width="90%"></td>107 </tr>108 </table>109 </div>110</div>111 112<div class="section">113 <div class="title">Results: Prompt-driven conditional generation</div>114 <div class="body">115 Ctrl-X also supports prompt-driven conditional generation, where it generates an output image complying with the given text prompt while aligning with the structure of the structure image. Ctrl-X continues to support any structure image/condition type here as well. The base model here is <a href="https://arxiv.org/abs/2307.01952" target="_blank">Stable Diffusion XL v1.0</a>.116 117 <!-- Adjust the number of rows and columns (EVERY project differs). -->118 <table width="100%" style="margin: 20pt 0; text-align: center;">119 <tr>120 <td><img src="assets/results_struct+prompt.jpg" width="100%"></td>121 </tr>122 </table>123 </div>124</div>125 126<div class="section">127 <div class="title">Results: Extension to video generation</div>128 <div class="body">129 We can directly apply Ctrl-X to text-to-video (T2V) models. We show results of <a href="https://animatediff.github.io/" target="_blank">AnimateDiff v1.5.3</a> (with base model <a href="https://huggingface.co/SG161222/Realistic_Vision_V5.1_noVAE" target="_blank">Realistic Vision v5.1</a>) here.130 131 <!-- Demo video here. Adjust the frame size based on the demo (EVERY project differs). -->132 <div style="position: relative; padding-top: 50%; margin: 20pt 0; text-align: center;">133 <iframe src="assets/results_animatediff.mp4" frameborder=0134 style="position: absolute; top: 2.5%; left: 0%; width: 100%; height: 100%;"135 allow="accelerometer; autoplay; encrypted-media; gyroscope; picture-in-picture"136 allowfullscreen></iframe>137 </div>138 </div>139</div>140 141<!-- === Result Section Ends === -->142 143 144<!-- === Reference Section Starts === -->145<div class="section">146 <div class="bibtex">BibTeX</div>147<pre>148@inproceedings{lin2024ctrlx,149 author = {Lin, {Kuan Heng} and Mo, Sicheng and Klingher, Ben and Mu, Fangzhou and Zhou, Bolei},150 booktitle = {Advances in Neural Information Processing Systems},151 title = {Ctrl-X: Controlling Structure and Appearance for Text-To-Image Generation Without Guidance},152 year = {2024}153}154</pre>155 156 <!-- BZ: we should give other related work enough credits, -->157 <!-- so please include some most relevant work and leave some comment to summarize work and the difference. -->158 <div class="ref">Related Work</div>159 <div class="citation">160 <div class="image"><img src="assets/freecontrol.jpg"></div>161 <div class="comment">162 <a href="https://genforce.github.io/freecontrol/" target="_blank">163 Sicheng Mo, Fangzhou Mu, Kuan Heng Lin, Yanli Liu, Bochen Guan, Yin Li, Bolei Zhou.164 FreeControl: Training-Free Spatial Control of Any Text-to-Image Diffusion Model with Any Condition.165 CVPR 2024.</a><br>166 <b>Comment:</b>167 Training-free conditional generation by guidance in diffusion U-Net subspaces for structure control and appearance regularization.168 </div>169 </div>170 <div class="citation">171 <div class="image"><img src="assets/cross_image_attention.jpg"></div>172 <div class="comment">173 <a href="https://garibida.github.io/cross-image-attention/" target="_blank">174 Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, Daniel Cohen-Or.175 Cross-Image Attention for Zero-Shot Appearance Transfer.176 SIGGRAPH 2024.</a><br>177 <b>Comment:</b>178 Guidance-free appearance transfer to natural images with self-attention key + value swaps via cross-image correspondence.179 </div>180 </div>181</div>182<!-- === Reference Section Ends === -->183 184 185</body>186</html>187