Greetings, denizens of this trash pile! I bring you another in a line of ill-conceived experiments from my GPU, this time a first for me, on the Anima platform!I've been recently infatuated with the model and its capacities. Not only that, but I'd say for me the model is 90% there to what I want, which lets me truly screw around and find fun things to train.One of them is this model. I've taken a dataset of a truly eclectic mix.There's the soft-visual, painterly end, Zdzisław Beksiński, Steve R. Dodd, and other softer, flowier, curvier lines and shapes, then it moves smoothly through Dark Fantasy style crunch, all the waaay out to 10_4_(10f0ur), who's got a thick, gritty, sharp way of drawing. With, admittedly, some Omegaprocessor in there, because, well, it's my jam. And Marathon is somewhere square in the middle, so, my guess is it's going to be weird.And it was!The overall result is a model with a distinct lineart density, that responds well to concepts, and has a distinct look to it. Mechanical details seem to be strongly expressed and generally the model does really nice mecha-cyborg work. It also seems to be strongly weighted towards backgrounds and scenery, which I'd wager is down to the dataset.I think what largely saved this from becoming incomprehensible soup was the fact Illustrious and its Qwen LLM text encoder actually do handle language rather than token soup, so training with mixed tags (i.e. both Danbooru tagged images AND natural language captions) results in generally higher chances for the model to catch on. Schmaybe.Mileage may vary.

Files with SHA256: 9a4e1714890f4cd71a3e4c30504da2e79fd59b6a10a17d4cc31b19f9d18eac13