Version 2.0 update: after I connected the grid pattern sometimes visible in bright or dark parts of the images with the fp4 training with OneTrainer, I switched to fp8 training, which got rid of the pattern. But I have to redo all my LoRAs to fix them all. For the dataset I used only the images with real small glasses, but they come still out a bit on the bigger side. I decided to publish both versions - the 40 epoch version is the normal one. If you really like the John Lennon glasses try 100 epochs. The images are a bit more like plastic and the colors a bit bled out, but the glasses are smaller. Nothing changed in usage, except the tag is now &lt;lora:WireframeGlasses-SmallRound-step0720:1&gt; - or step01800 respectively.<hr />Flux seems pretty limited if you want glasses - mostly the more round sort with a thicker frame. I intended to make a all-in-one LoRA first, but that didn't work - if you want to specify between normal and small glasses, round, rectangular and oval shapes you had to train each one separately. This creates small round wire-frame glasses - hard to find today but in use for a long time. If the result is not as round or small, a &lt;lora:WireframeGlasses-roundSmall-60:1&gt; tag increases the influence. It has also the potential of creating several colors, in the examples is a gold one.

Wire-frame glasses - small round

Version 2.0 is trained with fp8, getting rid of the grid pattern sometimes visible in the images. The dataset was cleaned up to just the real small ones, otherwise the output would be more normal-sized round glasses. This alternative version is somewhat overtrained but produces even smaller glasses.

Files with SHA256: c09dd0aa063547918e96cb51390324398b662bb5aeb117f49241ff46f41c3b36