Rule of Thirds Composition
This model was trained to avoid those repetitive centered images AI usually creates, creating compositions that follow the "Rule of thirds", meaninig subject (or focal point) is a bit towards a side (1/3rd or the image).
UPDATE:
Second version has improved a lot in terms of success rate and less artiacts, however is not perect yet. When composition fails, rising the weight works at the expend of style contamination.
As you can see, the mode complex the prompt in terms of quality tokens, the higher weight it may require.
Description
FAQ
Comments (5)
I'd love to get this working, but can't seem to in the Civitai generator. I keep getting normal central composition. I'd gladly keep some of the other artifacts you mention if I could get some of these results.
I will keep working on it to improve it, I think I confused it a little inclouding subjects on both left or right, I will just mirror them to have them all on a side, for instance... adding more variated samples should improve it at well since most came rom backgrounds featuring sunsets... .kinda contaminating the results.
Very interesting. I'm not even at novice-level when it comes to shot framing and scene composition, but I've heard of the Rule of Thirds. I was under the impression that SD models already followed it, either by some internal mechanism or just the fact that when you use tags like "best quality" and "masterpiece," you're more likely to pull from trained data that followed the rule. Note that my assumption is likely not 100% correct. I'm eager to try this out and see how well it brings new variety to generations.
There should be more LoRAs like this, and I'd say that working to improve it would be a very worthwhile endeavor.
I hope you can improve this in the future. The odd centeredness and odd negative space has always been how I spot T2I made images in the wild, so fixing those goes a long way to making generations just that more "real".
Made a second version that improves it to some degree, not perfect yet tho!