<h1 id="version-2-is-here-3m4ftcc6o"><a rel="ugc" href="https://civarchive.com/models/605574">Version 2 is here</a> </h1> On-site trainer does not at allow (at least not easily) making it as a version for some reason, so its its own page. Added a bunch of extra pics to training data so it should be more stable. I was shocked and disappointed to discover Pony could not produce the Oval Office or the Resolute Desk through prompts alone, so I went to Wikimedia Commons, grabbed all the public domain images of the desk without humans in the way and cropped as needed. This got me only 12 images (some black and white) because I wanted to avoid the potential of generating a specific person and was thus unable to use large subsets of potential training data (advice for a revision that includes people at the desk without making every 1boy into carter/reagan/bush/clinton/bush/obama/trump is welcome). There’s a few pics of the desk outside the Oval Office, and the orientations are tagged (the PotUS view is “from behind”, across from him is untagged/front). I was curious how this would turn out with such a limited data set, though it’s clear I’ll need to add ones with people in them since just from the test images it’s clear it’s not really sure how to make people sit at desk and thinks the desk is supposed to leave the legs visible. Please post your images so I get a better idea of how to improve for 2.0Prompt: oval_office, resolute_desk . Recommend “photograph, photo (medium), realistic” as negative. Tagged support include desk, curtains, window, flag, plant, potted plant, indoors, carpet, flag, American flag, chair. The decade is also tagged (e.g., 1990s), though I doubt it will matter with such limited data. Since all training data images were public domain or crops of public domain, I’ve included the training data.

Files with SHA256: b145b4495219ebaf96c4b410f7fb3e5c358cb9f1ff6399d12f2b7eae5a1efdbe