I had been waiting for this from the usual suspects, since SDXL and Pony version do exist, but it appears to not be happening any time soon. I guess if you want something done...So, basic harness panel gag for Illustrious base models. Technically SFW but... you know. If you don't, ask your mom. Illustrious can already do the concept(s) ("panel gag" and "<s>face</s> gag harness") but it's not reliable, plus it really likes to put holes everywhere, which means that even if you use an SDXL LoRA, <a target="_blank" rel="ugc" href="https://civarchive.com/posts/10064901">you will more likely than not end up with something like a plug gag</a>. This LoRA is intended to fix this issue. It appears that it still wants to grab on to anything it can in the center, especially in close-ups. Trained on a mix of photos, illustrations and hand-picked AI images with relevant photoshop applied where applicable. Unlike some of the existing purely photo-based ones, it won't pull your image massively towards realism. That last point is also the reason there's a Pony version of it now (apart from it being easy since I already had the dataset, naturally): It's more docile than the existing LoRAs if you want source_anime or source_cartoon, but a lot less so if you want realism. The SDXL version is pretty much just to complete the set, but it works well with Animagine and other illustration-focused checkpoints.Recommended prompt is "panelgag, panel gag" and "face harness" for v0.5/"gag harness" for v1.0. "covered mouth" can help if you still get mouth artifacts. Strength 1.0 seems to work fine, go lower for close-ups and higher for wide shots if necessary as per usual. The "gag harness" tag has a tendency to imply body harnesses, put "breast harness, bondage harness" into the negatives if that's a problem. You can also try leaving the "gag harness" prompt out completely, which is especially helpful if you want it to be partially covered by hair or headwear of some kind.<s>You can choose any color,</s><s> </s><s>as long as it's black.</s> Turns out color can be done, it just might require upweighting of the color prompt.

Harness Panel Gag

This evidently works, but prompting is very scuffed – see trigger words, adjust that phrase as needed but the structure is the one I've found most reliable (thankfully, with Z-Image you can always just go into more detail if you have to, see the nurse example). It's also not very reliable for anything that isn't a straight-on portrait. I will probably retrain this with better captions and more epochs when the trainer is working better.

Files with SHA256: 92d4442dbcf08968958fe46ad3cfa86dd8e673bc96fff868c4334d7db10a8e08