Ongoing Updates<ul><li>(9/16): I have had some major life events happen over the last month that have forced my attention away from this work (good things, fortunately); apologies to anyone who has been following this model's development. :) I have several candidates for a v3, but have had no time to test. I am actively working on finding that time!</li></ul><ul><li>(8/18): An experimental <code>fp16</code> version of the <code>v1</code> model has been uploaded. Initial testing seems to indicate that the results are at least comparable.<ul><li>If you can run the <code>bfloat16</code> version, <code>bfloat16</code> will probably perform better.</li></ul></li><li>(8/14): A <code>v2</code> checkpoint has been uploaded. Please read the "about this version" notes in the drop-down panel to the right.</li></ul><h3 id="heading-10715">CanvasXL: Unleash Your Creativity</h3>Presenting CanvasXL, an SDXL v1 Dreambooth fine-tune project which aims to empower your creativity like never before! Trained on a dataset of thousands of highly aesthetic, artistic AI-generated images, CanvasXL sets out to bring your vision to life with unparalleled beauty and creativity within the open-source image generation space.I have tried my best to summarize the model, my intentions with it, and my plans around it in the text below. I hope you enjoy the fruits of my labor and create some beautiful images! If you like and want to support my work, you are welcome to <a target="_blank" rel="ugc" href="https://bmc.link/canvasxl">buy me a coffee.</a>Features, Specifications, and Notices:<ul><li>Versatility: CanvasXL produces a wide array of subjects and represents novel concept combinations impressively across many contexts.</li><li>Resolutions: Trained on 44 different resolutions, with particular preferences for 1024x1024 and 704x1216 (mobile portrait-style). Some subjects will be replicated better on one resolution than the other, currently.</li><li>Model Type (Base vs Refiner): Currently, a fine-tune of the SDXL v1 <code>base</code> model is presented. This model can be used for <code>txt2img</code> and <code>img2img</code>, however I plan to fine-tune the <code>refiner </code>model as well, once I have the time to figure it out!</li><li>Convenient Use: A ComfyUI <code>workflow.json</code> is available for easy set up of the model, to help you get started. <a target="_blank" rel="ugc" href="https://pastebin.com/Uqgde3vt">You can copy it from here.</a></li><li>Format: The model is currently presented in <code>bfloat16</code> format. If anyone needs a <code>fp16</code> version, please leave a comment and I will create it when I have the time!</li><li>Demo Images: The generation parameters for the images should be available. The prompts were directly copied from a selection of nice-looking images on a popular AI image generation site. The 20 demo images were cherry picked from a set of approximately 300 images, and are completely unedited.</li><li>Future Plans: I have amassed a dataset of ~30k well-tagged, good-quality images. I am currently pruning the dataset to increase the quality, cleaving about 60% of the total images. However, I have only made it through half of the total dataset, so the training set should expand and improve as I have the time to work through the set. As significant additions are made to the training set, new training runs will be performed on my local machine and uploaded.</li><li>Content Notice: The training dataset is entirely free of NSFW, nudity, violent, or otherwise potentially-highly objectionable imagery. As a result, this model should not be expected to be able to replicate such subjects well or even at all, as it is explicitly intended to produce non-objectionable imagery.</li><li>Usage Notice: Please be aware that you are solely responsible for any content that you create using this model. In addition, your use of this model implies that you accept an agreement to not use it to produce harassing, harmful, illegal, or otherwise highly-objectionable imagery.</li></ul>Resolution Buckets:CanvasXL's training encompasses various resolution buckets, ranging from (512, 1664) to (1792, 512). Due to variance across training runs, the general distribution is described:<ul><li>bucket 0: resolution (512, 1664), count: low</li><li>bucket 1: resolution (512, 1728), count: low</li><li>bucket 2: resolution (512, 1792), count: low</li><li>bucket 3: resolution (576, 1472), count: low</li><li>bucket 4: resolution (576, 1536), count: low</li><li>bucket 5: resolution (576, 1728), count: low</li><li>bucket 6: resolution (640, 1344), count: low</li><li>bucket 7: resolution (640, 1408), count: low</li><li>bucket 8: resolution (640, 1472), count: low</li><li>bucket 9: resolution (640, 1536), count: low</li><li>bucket 10: resolution (640, 1600), count: low</li><li>bucket 11: resolution (704, 1216), count: HIGH</li><li>bucket 12: resolution (704, 1280), count: low</li><li>bucket 13: resolution (704, 1344), count: low</li><li>bucket 14: resolution (704, 1408), count: MID</li><li>bucket 15: resolution (704, 1472), count: low</li><li>bucket 16: resolution (768, 1152), count: low</li><li>bucket 17: resolution (768, 1216), count: low</li><li>bucket 18: resolution (768, 1280), count: low</li><li>bucket 19: resolution (768, 1344), count: low</li><li>bucket 20: resolution (832, 1088), count: low</li><li>bucket 21: resolution (832, 1152), count: low</li><li>bucket 22: resolution (832, 1216), count: MID</li><li>bucket 23: resolution (896, 1024), count: low</li><li>bucket 24: resolution (896, 1088), count: low</li><li>bucket 25: resolution (896, 1152), count: low</li><li>bucket 26: resolution (960, 896), count: low</li><li>bucket 27: resolution (960, 1024), count: low</li><li>bucket 28: resolution (1024, 896), count: low</li><li>bucket 29: resolution (1024, 960), count: low</li><li>bucket 30: resolution (1024, 1024), count: HIGH</li><li>bucket 31: resolution (1088, 832), count: low</li><li>bucket 32: resolution (1088, 896), count: low</li><li>bucket 33: resolution (1152, 832), count: low</li><li>bucket 34: resolution (1216, 704), count: low</li><li>bucket 35: resolution (1216, 768), count: low</li><li>bucket 36: resolution (1216, 832), count: low</li><li>bucket 37: resolution (1280, 768), count: low</li><li>bucket 38: resolution (1344, 704), count: low</li><li>bucket 39: resolution (1344, 768), count: low</li><li>bucket 40: resolution (1408, 704), count: low</li><li>bucket 41: resolution (1600, 640), count: low</li><li>bucket 42: resolution (1728, 576), count: low</li><li>bucket 43: resolution (1792, 512), count: low</li></ul>

CanvasXL

CanvasXL v2 Release NotesOver the past week, I have performed a number of training experiments with the goal of producing an improved version of the model. Sadly, my initial, as-planned training run for v2 did not produce especially appealing results. However, an interesting model was produced during subsequent testing runs. This version is not a 1:1 upgrade from v1, however I feel that it has an interesting versatility, so I wanted to share the checkpoint.This checkpoint has been trained on a much larger subset of my total collected dataset, and as a result it tends to combine subjects in more interesting and versatile ways. However, this presents the double-edged sword that it is less tuned towards defaulting to an appealing aesthetic. While I have not had time to experiment too much, mixing and matching v2 and v1 as txt2img and img2img may produce interesting results!At this point I have three short-term goals planned for this project:<ul><li>Produce FP16 versions of the checkpoints (priority 1)</li><li>Produce a fine-tune of the refiner model (priority 2)</li><li>Work towards a "best of both worlds" v3 (priority 3)</li></ul>The priority has largely been set in consideration of how much time and effort is expected to be required to tackle each item. Expect fp16 versions in the next few days, updates on the refiner in the same time-frame, and with a little luck, a refiner release, and perhaps even a v3 by week's end. No promises though! I am hoping that I can tune v2 using a similar strategy to v1, to produce v3. :)

Files with SHA256: 53761582ee1f4bfd312c95caaa9271d65b1c6d665de3396b1302200da23fb84f