algorithms of the filter might be too strict

[sdwcn]

i had tried with a few checkpoints, as shown above. Not sure if it's possible to have a weight option in order to adjust filtering rules?

[aria1th]

Sure, I'll add some thresholds and options for filtering (maybe, within 9 hours)

[sdwcn]

Sure, I'll add some thresholds and options for filtering (maybe, within 9 hours)

how's it going?

[aria1th]

@sdwcn I tried some more setups, but seems like it has weird bias on white dress... The first plan was implementing labels to blur, but maybe it should have some very custom setups (such as thresholding by probability) Today, I'll release with some basic and custom setups, but I think model was roughly underfitted for border-images, resulting 'some' exposure to be censored

[sdwcn]

noted, I'd like give a try,

[sdwcn]

how shall I understand the diff between safe questionable & explicit?

levels_to_idx = {
        'safe': [1,3,4],
        'questionable': [1,3],
        'explicit': []
    }

according to tags in comments, ['drawings', 'hentai', 'neutral', 'porn', 'sexy'] does qustionable equals [hentai + porn]?

[aria1th]

Yes, safe = all (hentai porn sexy) questionable = hentai+porn explicit = none

[sdwcn]

switched to "questionable", it still bias on white dresses, it sounds good if a probablity setting is enabled. cheers

[aria1th]

@sdwcn I can offer the probability option but its hyperparameter may require tremendous amout of test 😭