Qwen-Image-Edit Uncensor+Nudify
550
5
25
550 5 25
This LoRA is experimental. (All previews include unedited single-shot generations with just the 8-step LoRA and this model)
It was first trained exclusively on a censored->uncensored dataset with ~3k image pairs for 4 epochs at 512x512 LR 2e-5.
At this point the LoRA would often oftset the elements of an image by a little bit on higher resolutions.
Then it was further trained for a single epoch at 1024x1024 LR 5e-5.
This fixed the offset issue. But the model still clearly didn't understand genitals at all. It detected censor bars and mosaic censoring very well and tried to fix them but it just didn't know how to fill them properly.
I figured I wouldn't be able to make it a decent uncensoring LoRA no matter what so I just added ~650 clothed->nude (girls-only) pairs to the dataset and trained on all of them for 2 extra epochs at 768x768 LR 2e-5.
The result:
detects bar/mosaic censorship fairly easily
due to large and diverse dataset, it can handle a wide range of styles and poses but struggles with doujinshi
can remove clothes while preserving character position and other elements in the image that should not be changed
sometimes it needs encouragement to remove leg-wear and head-wear (remove hat, remove pantyhose, remove shoes, etc...)
still does not understand genitals and struggles with nipples.
Usage:
load the Lightning-8steps LoRA and set steps to 8 and CFG to 1
(optional) load a nipples and/or p*ssy Qwen-Image LoRA(s) with strength 0.3~0.75
load this LoRA with strength 1
trigger words: uncensor, remove all clothes
tip: you can reinforce the model by adding 'completely nude' and/or 'remove (hat/bikini/bra/panties/pantyhose/skirt/shirt/etc...)'
My conclusion from all my experiments:
We won't be able to ever get a decent nudify/uncensor LoRA for Qwen-Image-Edit until someone does a whole NSFW finetune on Qwen-Image (which can then be extracted as a LoRA and loaded into the Edit model).
Alternatively we would need decent genitals/nipples exclusive LoRAs trained on Qwen-Image with a large and diverse dataset (the ones available at the time of writing are very limited and biased towards a specific body type and style - and even then they are still not great)
模型信息
'%3e%3cg%3e%3cpath d='M6.9375%2c1.875L14.8125%2c1.875C15.5374%2c1.875000858307%2c16.125%2c2.462627%2c16.125%2c3.1875L16.125%2c11.0625C16.125%2c11.78737%2c15.5374%2c12.375%2c14.8125%2c12.375L6.9375%2c12.375C6.21263%2c12.375%2c5.625%2c11.78737%2c5.625%2c11.0625L5.625%2c3.1875C5.625%2c2.462626%2c6.21263%2c1.875000343323%2c6.9375%2c1.875ZM6.9375%2c3C6.83395%2c3%2c6.75%2c3.0839499999999997%2c6.75%2c3.1875L6.75%2c11.0625C6.75%2c11.16605%2c6.83395%2c11.25%2c6.9375%2c11.25L14.8125%2c11.25C14.9161%2c11.25%2c15%2c11.16605%2c15%2c11.0625L15%2c3.1875C15%2c3.0839499999999997%2c14.9161%2c3%2c14.8125%2c3L6.9375%2c3ZM11.25%2c13.5C11.25%2c13.1893%2c11.50184%2c12.9375%2c11.8125%2c12.9375C12.1232%2c12.9375%2c12.375%2c13.1893%2c12.375%2c13.5L12.375%2c14.8125C12.375%2c15.5374%2c11.78737%2c16.125%2c11.0625%2c16.125L3.1875%2c16.125C2.462626%2c16.125%2c1.875000171661%2c15.5374%2c1.875%2c14.8125L1.875%2c6.9375C1.875%2c6.21263%2c2.462626%2c5.625%2c3.1875%2c5.625L4.5%2c5.625C4.81066%2c5.625%2c5.0625%2c5.87684%2c5.0625%2c6.1875C5.0625%2c6.49816%2c4.81066%2c6.75%2c4.5%2c6.75L3.1875%2c6.75C3.0839499999999997%2c6.75%2c3%2c6.83395%2c3%2c6.9375L3%2c14.8125C3%2c14.9161%2c3.0839499999999997%2c15%2c3.1875%2c15L11.0625%2c15C11.16605%2c15%2c11.25%2c14.9161%2c11.25%2c14.8125L11.25%2c13.5Z' fill='%239DA2A8' fill-opacity='1' style='mix-blend-mode:passthrough'/%3e%3c/g%3e%3c/g%3e%3c/svg%3e)
This LoRA is experimental. (All previews include unedited single-shot generations with just the 8-step LoRA and this model)
It was first trained exclusively on a censored->uncensored dataset with ~3k image pairs for 4 epochs at 512x512 LR 2e-5.
At this point the LoRA would often oftset the elements of an image by a little bit on higher resolutions.
Then it was further trained for a single epoch at 1024x1024 LR 5e-5.
This fixed the offset issue. But the model still clearly didn't understand genitals at all. It detected censor bars and mosaic censoring very well and tried to fix them but it just didn't know how to fill them properly.
I figured I wouldn't be able to make it a decent uncensoring LoRA no matter what so I just added ~650 clothed->nude (girls-only) pairs to the dataset and trained on all of them for 2 extra epochs at 768x768 LR 2e-5.
The result:
detects bar/mosaic censorship fairly easily
due to large and diverse dataset, it can handle a wide range of styles and poses but struggles with doujinshi
can remove clothes while preserving character position and other elements in the image that should not be changed
sometimes it needs encouragement to remove leg-wear and head-wear (remove hat, remove pantyhose, remove shoes, etc...)
still does not understand genitals and struggles with nipples.
Usage:
load the Lightning-8steps LoRA and set steps to 8 and CFG to 1
(optional) load a nipples and/or p*ssy Qwen-Image LoRA(s) with strength 0.3~0.75
load this LoRA with strength 1
trigger words: uncensor, remove all clothes
tip: you can reinforce the model by adding 'completely nude' and/or 'remove (hat/bikini/bra/panties/pantyhose/skirt/shirt/etc...)'
My conclusion from all my experiments:
We won't be able to ever get a decent nudify/uncensor LoRA for Qwen-Image-Edit until someone does a whole NSFW finetune on Qwen-Image (which can then be extracted as a LoRA and loaded into the Edit model).
Alternatively we would need decent genitals/nipples exclusive LoRAs trained on Qwen-Image with a large and diverse dataset (the ones available at the time of writing are very limited and biased towards a specific body type and style - and even then they are still not great)