Addressing a training quick from earlier SimpleTuner versions

Update model_index.json

Update model_index.json

Update scheduler/scheduler_config.json

Update scheduler/scheduler_config.json

set alpha to one so that we remove residual noise like a champ

100250 steps of training, auto-pushed, enabled datasets: mj-general-1mp, nijijourney, laion-hd-exif-1024, laion-hd-safe-1mp, exif, experimental, ethnic, sports, gay, architecture, shutterstock, cinemamix-1mp, nsfw-1024, anatomy, bg20k-1024, yoga, photo-aesthetics, text-1mp, text-embeds

96375 steps of training, same settings

95125 steps of training, same settings

94250 steps of training, filtered captions, no bg20k

92250 steps of training, same settings

87375 steps

86500 steps

85625 steps, timesteps 0-400 with 2x multiplier and recached embeds, aspect crops

84750 steps, added unsplash, timesteps 0-1000, maximum image size and target downsample with crops

83750 steps, shutterstock

83125 timesteps 0-400, seems to be struggling, we will see

80375 steps, 0-400, input perturbation

79125 steps, input perturbations 0.1 and all aspect bucketed data, no crops

72750 steps, timesteps 0-1000

67500 steps, now with text embeds in the cloud

66500 steps, same settings

65750 steps, added nijijourney back, timesteps 0-1000

63500 steps

62625 steps, same config

60375 steps, added photo-aesthetics data

58625 steps. since the last commit, we trained on 400-1000 and then back to 0-400 but with reduced bias from 50x to 5x.

53250 steps on v1, since last ckpt we have biased timesteps 0-400 and things improved

45750 steps, weird results

44875 steps

42750 steps continuation of prior

39750 steps, before restarting on recached all-aspected data

34000 steps, maybe deteriorated, not sure

24375 steps on prev commit

fresh v1 to 21125 steps on niji-nsf-cinema-b mix, b disabled early on

3200 steps continued from v1+multidataset_1600 branch

1600 steps fresh on v1 with multidataset trainer

Update model_index.json

v2 beta1, 2200 steps

v2 beta 1 model, 1800 steps at bsz=96 (24 * 4)

3600 steps one-shot trained from revert. photo dataset finetuning, fixed micro conditioning inputs

Revert unet to cb6f860 71400 steps, after fixing min image tracking

Update scheduler/scheduler_config.json

v1 model retrained on new master dataset to 4800 steps

another 18800 steps on 118400 ckpt, starting to deform some subjects. approx 55 epochs of the early dataset.

reset 118400 branch to 0 steps and then did 11800 steps on a photo dataset for realism finetuning. hands got harmed.

71400 steps, after fixing min image tracking

57800 steps

prev checkpoint at 42600 steps carried forward to 52000 steps without offset noise, improved coherence and reduced crust further

36000 + 21800 + 14200 + 42600 steps, 118400 total - continued training from 30k steps on to 42600 steps using cropped squares with conditional inputs