Artist: hospitallerinaboat (locked)

Staff members involved: Lafcadio, Rainbow Dash, scaliespe, tsukemono, Rainbow Shoebill, Jarlium

HospitallerInaBoat repeatedly disregarded our evidence and findings with no attempt to address them, and on this basis alone we would normally be justified in denying the appeal with no further action. This writeup will explain in detail both our reasoning for deleting HospitallerInaBoat's gallery and the retaliatory raid that ensued.

When HospitallerInaBoat provided PSDs to try and establish legitimacy, a review of their layers revealed details consistent with painting over AI-generated images. Post IDs are included for each piece, and non-staff can view the accused posts at non-Fur Affinity sources; Fur Affinity had already banned this user in advance.

The public version of post #5176721 was the first item we saw from Fur Affinity's investigation. Their staff team had raised two points: a part of the right foot and part of the character's (intentionally broken) horn seeming to disappear into the male's body. However, there is something we have seen from other pieces: the scutes run the entire length of the tail with no green part visible. There is also a separate image that displays a much different style of scutes despite depicting another Skyrim dragon.

The public version of post #4431552 has several issues that compound on one another.

• Distinct seams along the torso show what looks like a one-piece swimsuit. This is a consequence of early non-furry datasets having less data for non-human body colors, and we can explicitly see a vertical seam above the character's vulva, which would basically only exist if a character was wearing that clothing article.
• A white halo covers the character almost entirely, which is consistent with magic wand tools to remove backgrounds from AI-generated materials.
• The character's pink clitoris blends smoothly into the lower body despite the artist's touches otherwise including lots of strong outlines.
• The character's left palm also shows instances of entirely flat coloring that are otherwise not present in the artist's work, with the more detailed parts of the palm retaining consistent AI quirks like tiny highlights.
• Various places exhibit suspicious scale and scute patterns: the scutes on the base of the tail are highly irregular, and the right-hand side of the tail has scutes that overlap the left-hand scutes and create a magical third lane of scutes.
• A series of round brush strokes just behind the character's head suggest that the background was taken from a different source and then modified to obscure certain details. These additional strokes are not in-style with the background and lack noise that would otherwise be present in the background. This would not be an issue on its own - AI-generated backgrounds are not against e621's Uploading Guidelines - but this in addition to the other factors clues us in that something is not quite right.

The public version of post #4630366 has more issues, demonstrating that different artistic processes are being used to make different parts of the image.

• There is an obvious break in the tail on the left side where the detail goes from AI-generated scutes with shaky dark grey outlines to human-painted splotches of light grey on a dark grey background. The red part of the tail also differs between both halves; the AI-generated part plainly has more detailed red coloring, and the human-painted part has essentially flat red coloring.
• Similarly irregular scale patterns line the character's head. Compare the left and right sides of the face, noting places where the scale boundaries are different, where some scales are divided and where others aren't, etc.
• The spikes along the right jaw have simplistic coloring when compared to the parts of the face nearby.
• The bottom parts of the right wing reveal a number of AI-generated details, including thin strokes that break apart and rejoin at arbitrary points, and there are sudden "shears" in the lower wing that suggest parts were copy-pasted to change their position without redrawing.

The PSD for post #5657125 contained localized areas of high complexity with telltale AI smudges or noise, and on top of those, simple dots of yellow and brown coloring. The public-facing copy obscures these details by drawing additional scales and lineart to hide the AI-generated sclera while detailing the yellow eye to make its simplistic nature less obvious.

The public version of post #4661304 has more issues:

• The furry character's back includes two wing bases but only portrays one wing. If one zooms in closely, we can also see different styles of coloring/shading: the left part has smoother coloring, and the right part includes more details consistent with AI.
• The furry character's hair loops back in on itself near the tail, creating a solid loop of hair that connects the right side of the head to the left, and once more near the shoulder. Hair does not work like this.
• The seam in the middle of the furry character's pants has AI noise and a stair-stepping pattern associated with artificial upscales.

The public version of post #4479364 has highly detailed but nonsensical stockings where the pattern changes to either a dotted grid or straight lines. The skirt pleats also don't make sense, and the clothing buttons and folds retain blurry details that are consistent with AI-generated coloring.

The public version of post #4497778 shows a number of background characters whose details are both highly indistinct and highly detailed. There is smooth lighting along the shoulders of the man with the green object crossing his chest (possibly the diagonal strap of a backpack), and a highly detailed face. The stockings are likely just a checkerboard grid rather than an AI-generated pattern, but the teacher's buttons retain the same uncanny "highly detailed and highly indistinct" style.

The public versions of post #4921866 and post #5272331 include random noise on what would otherwise be a pure-white or pure-grey background. This is a consequence of how AI image generation works, starting with random noise and then trying to "unblur" it to create something that looks like the AI's training data. In the former image, several high-detail bits of coloring can be seen around the character's fingers and hand on top of low-detail AI-generated coloring.

The public version of post #5430061 includes two very different styles of shading: most of the body smoothly transitions from one color to the next (Jarlium testified that they "cannot do this"), and then much simpler feathers, arm scutes, etc.. The feathery tuft between the character's breasts also has a lot of misshapen details that remain consistent with AI trying to draw tufts and is not consistent with the level of care taken in shading the character's midsection.

Most other pieces show some level of background cutouts, indistinct fingers/nipples, differences in quality of coloring/shading, etc.

HospitallerInaBoat saves certain files with a naming scheme reminiscient of Automatic1111, which includes one variable that increments by +1 for every new generation and a second variable that includes the random seed that determines the image's starting point. 00007-383141380-2.jpg would have been the seventh generation (this number resets if generations are deleted later) with the random seed 383141380.

Additionally, inspecting HospitallerInaBoat's deviantArt account reveals $36 spent on AI-generated "adoptables". While this is not sufficient to claim the use of AI imagery in art, it shows that it is not a dealbreaker for them.

• https://www.deviantart.com/furryfarm/art/SOLD-Adopt-Dragon-8-1031343705
• https://www.deviantart.com/aifurrygirl/art/Sorrel-Blazeshroud-Scalie-Red-Dragon-Adoptable-970269774
• https://www.deviantart.com/aifurrygirl/art/Marsala-Scorchtongue-Scalie-Red-Dragon-Adoptable-970770337
• https://www.deviantart.com/sarade45/art/Khibae-Adoptable-CLOSED-975466052

Critically, HospitallerInaBoat's Twitter account had several videos showing off highly suspicious layering, where most of the characters' coloring was reduced to a single layer: post #5326905 demonstrated that the entire character sans the whisker and a single tuft of hair were on a single layer, and a video of post #5511918 shows the artist drawing minor fur details over a finished AI-colored layer. It is highly unusual for artists to work on a single layer except in certain mediums or as a self-imposed challenge, as this necessarily makes every change in coloring destructive: errors cannot be hidden or undone and must be painted over. If not for the other issues raised here, an artist could make a recording demonstrating that yes, they really do work in a single layer and that would simply have to be accepted as fact.

Under normal circumstances, we would just consider the matter closed unless the facts behind the investigation changed, such as the artist providing legitimate proof of these details being non-AI. Things would not stay normal for long: HospitallerInaBoat organized their followers to target Lafcadio for his part in leading the second investigation and consequent deletions, resulting in user report tickets and Discord harrassment.

HospitallerInaBoat proceeded to attack the credibility of janitorial staff, livetweeted the appeal for their followers, disregarded the evidence we provided, and repeatedly sent staff messages in an attempt to derail and distract. None of the artist's recordings or PSDs provided sufficient proof, and in fact a few of these points could only be raised because the artist willingly provided PSDs to us.

It was obvious that the appeal could not proceed like this. Timeouts and slowmode timers were employed to try and force HospitallerInaBoat to be more selective and thoughtful about their replies, and the result was to instead continue a pattern of abusive tweets, bargain for the deletion of tweets, and mock site admins for not being willing to pull rank with Lafcadio. This included trying to coordinate with legitimate artists to coast by with just reputation instead of evidence, calling for mentors who would supposedly back his anecdotes about drawing on single layers, etc.

Evidence beyond a reasonable doubt has been established: HospitallerInaBoat has an unacceptable level of AI assistance, and the abuse against e621 staff suggests he cannot be trusted to make good-faith appeals. This behavior can't be rewarded.