Test to make a sort of "Highres Fix" LoRA. Trained over top of v1-5-pruned-emaonly.

The idea is to remove the need for using Highres Fix, but I'm having decent results by using it with Highres Fix. So the first pass would be the model as usual, then for the second pass only you would load the LoRA. That way your original image isn't affected, but for the second pass it would be more stable.

I want to stress that this is a test, and may be completely useless to you. Depending on how I feel about results I get from further testing I may continue trying to improve this.

bucket 0: resolution (512, 2048), count: 1
bucket 1: resolution (576, 1792), count: 1
bucket 2: resolution (640, 1536), count: 10
bucket 3: resolution (640, 1600), count: 5
bucket 4: resolution (704, 1408), count: 61
bucket 5: resolution (704, 1472), count: 164
bucket 6: resolution (768, 1280), count: 179
bucket 7: resolution (768, 1344), count: 525
bucket 8: resolution (832, 1216), count: 3480
bucket 9: resolution (896, 1152), count: 7294
bucket 10: resolution (960, 1088), count: 576
bucket 11: resolution (1024, 1024), count: 2382
bucket 12: resolution (1088, 960), count: 370
bucket 13: resolution (1152, 896), count: 4348
bucket 14: resolution (1216, 832), count: 5292
bucket 15: resolution (1280, 768), count: 409
bucket 16: resolution (1344, 768), count: 1055
bucket 17: resolution (1408, 704), count: 164
bucket 18: resolution (1472, 704), count: 212
bucket 19: resolution (1536, 640), count: 56
bucket 20: resolution (1600, 640), count: 40
bucket 21: resolution (1664, 576), count: 23
bucket 22: resolution (1728, 576), count: 10
bucket 23: resolution (1792, 576), count: 18
bucket 24: resolution (1856, 512), count: 5
bucket 25: resolution (1920, 512), count: 2
bucket 26: resolution (1984, 512), count: 3
bucket 27: resolution (2048, 320), count: 2
bucket 28: resolution (2048, 384), count: 1
bucket 29: resolution (2048, 448), count: 1
bucket 30: resolution (2048, 512), count: 5
mean ar error (without repeats): 0.030218674326178727