It's something I didn't quite work out yet. For exampe, I've been training up a Giger model and released a first version at 1600 steps while it trained more.
It hit 3600, and I've been experimenting, and while it feels like it's much more Giger, is there a use in a model like that if it is harder to control? And how do I work out how 'hard' it is to control.
A piece of artwork by HR Giger. The artwork depicts a large space, the camera looking upwards to an opening in the ceiling. There is fog in the distance. In the center is a pale organic mass inside, intertwined with mechanical structures, set against a backdrop of swirling, abstract forms. Dominated by shades of brown and red, the scene includes tubes and pipes that merge with what appear to be fleshy, cancerous growths, suggesting a fusion of nature and machinery. Grotesque, warped female and male anatomy are spread around the surface of the machinery. The landscape is textured with intricate details, where the organic components have bulbous and porous surfaces, while the mechanical parts feature organic ribbed and linear designs.
It feels like each of those captures what I was asking for. It obviously becomes more surreal and abstract as the steps increase, but Giger is very abstract and would not likely have the feet and hands of the first image in the manner they are. Or would he? I don't know.
A piece of art by HR Giger. The artwork depicts a complex and intertwined composition of skeletal, humanoid figures and serpentine forms. The central figure is arched, with visible ribs and elongated limbs. Surrounding this figure are other creatures, similarly skeletal, with elongated skulls and limbs that intertwine with each other. The scene is rendered in monochrome tones, blending organic and mechanical elements in a fluid, continuous motion. The artwork is noisy and messy with scratches and canvas marks in places and desaturated colours
Again, it feels like each one captures what I've asked for, and becomes more and more Giger as you progress. Maybe I need to come up with a way of testing how flexible something is - ie. take a picture such as above and add something to it.
It definitely becomes more 'ah that's Giger' rather than 'that's Giger inspired' with each one, so I guess it would depend on what people are looking for (and ignoring that you could weight the LoRa prompt).
A piece of art by HR Giger. The artwork shows a pale, serene humanoid face with striking white eyes is displayed in the center, embedded within a complex array of mechanical and organic components. This face is crowned by a headpiece adorned with intertwining skulls and elongated tendrils resembling snakes. boils and lesions cover her body, and her skin is decaying and rotting beyond her face. skulls are placed above her head with pipes snaking in and out of the orifices. Flanking the woman are two skeletal figures with ribbed torsos and elongated limbs, seamlessly integrated into the metallic and tubular backdrop. It portrays a fusion of human and mechanical elements, creating a haunting yet intriguing composition.
This one definitely loses flexibility as it progresses. At 3600 steps, the 'flanking skeletal figures' are very minimal and abstracted. There are skulls above the head, though the pipes become less 'snaking'. All three feel Giger to me in their own ways. That prompt is also weighted to 1.6 though give it the overpowering Giger levels. Here it is at 1.0:
Where it does retain the skeletons. If anything, 1.0 still feels a little underdeveloped, even at 3600.
This is where it's tricky. Some prompts capture the feel the LoRa is going for perfectly - some feel like they need a push. Some feel like they need to be reigned back in.
Perhaps the answer is to provide multiple steps per release? I don't know.