This model allows for overhead perspectives. Additional supporting descriptions allow finetuning the perspective. Try things like {overhead | side | rear | front view looking at the woman | high above | over shoulder | front over head } view overheadview <span style="background-color:rgb(26, 27, 30);color:rgb(193, 194, 197);font-family:-apple-system, BlinkMacSystemFont, &quot;Segoe UI&quot;, Roboto, Helvetica, Arial, sans-serif, &quot;Apple Color Emoji&quot;, &quot;Segoe UI Emoji&quot;;font-size:14px">it does not necessarily play well with others, but it can work well if you carefully avoid drawing attention to things other models or checkpoints are trained to display front-on. in short, mention the things you want the model to focus on like 'hair focus' for a top-down view and don't mention things like 'naked' if you want a view of the back. say instead 'unclothed' and 'leaving'. It is hard to avoid triggering viewpoints of other models but it does work.

overheadview

I trimmed the image database to remove confusing images and attempted to teach it some groupings with new terms. Interestingly, it learnt the poses well except for doggiepose, independently of the viewpoint, which means they trigger a forward perspective at times. It still struggles with exactly topdown.. With regards to the other terms, sometimes it listens, sometimes it does not - it appears to depend on the seed, but the results are good enough I feel and there does not appear to be any overfitting - just confusion with the perspective and posture combinations. Here are the terms that can be added to overheadview trigger word; Word Percentagetopdown 21.10 shoulderdown 9.35 backdown 5.76 frontdown 8.63 walkingpose 1.44 standingpose 16.79 kneelingpose 3.12 lyingpose 5.04 doggiepose 5.04 frontfocus 4.08 backfocus 7.67 chestfocus 3.36 sidefocus 2.64 bumfocus 1.20 facedown 2.40 faceup 2.40

Files with SHA256: a0b929db82b4f620636f74d670826c4289fc5603ffeedc612a22ce43d188a460