Grew up watching this show and over the last couple of years have been attempting to model the ship in Sd 1.5, sd2, sdxl and now Flux. The token is zwxem. It really likes wide landscape images, with a high hit rate with those. See more usage tips below.
2.0 has majorly improved details and viewpoint prompt responsiveness.
Key usage points:
zwxem spaceship flying though space, view from front (other words are back, above, below or combinations of those)
Add the word Cinematic to trigger a portion of the dataset which had a lot more movie stills, resulting in far more interesting lighting, at cost of some sharpness.
An example is:
Zwxem spaceship flying though a nebula cloud in space, view from above,cinematic
It works well between .7 and 1.0 Maybe use at 1.0 at first, but you can go a lot lower, and if you leave out the 'flying though space' at .70 for example, it can become a lot more versatile to ship damage etc. Ship damage is much easier to trigger in none space environments, even at 1.0
Some examples here at 1.0 strength, from the same prompt and seed, just changing the viewpoint keyword. You can download them to check for yourself in comfyui: https://www.dropbox.com/scl/fo/emlfemldkdgzanzpdv8vy/AOpylisK_2MGPF_4iYM0fZk?rlkey=503w3v9il0amo812v2m2cgh8l&st=nmz6dnp7&dl=0
Workflow to start you off nice: https://www.dropbox.com/scl/fi/10xj542nuobgvynw25i4g/EnterpriseV2-Workflow.json?rlkey=bsplytaiyfef7om7ojhvl102k&st=urlh5phd&dl=0
My workflow uses the new text encoder by the way, but you dont have to use it, but it does seem to improve output a lot : https://huggingface.co/zer0int/CLIP-GmP-ViT-L-14/tree/main
Misc notes:
This lora is now dimension 64, Alpha 64 - this is necessity from my experiments with V2
No star trek ship words or any other words were used in training, I'm guessing using them in your prompts, especially the ship name may be a negative, due to the poor understanding of spaceships in the base model.
It tends to work better with landscape orientation images! (16:9 or even wider) Just to flag that! Enjoy !
BONUS:
I have a alternate workflow, used to make the image below that combines two ksamplers for varying the Lora strength during generation and also using a great lighting lora in the first 2 steps, and hyper loras for upscaling:
Here is the workflow: https://www.dropbox.com/scl/fo/ge4ndpmyq8nz748dme8rv/AHhTNL-8GNkr__iq5WiRJZM?rlkey=nrasxogpnc7k3hvd4kpcodik4&st=ox21u7wx&dl=0
Here is the lighting lora: https://civitai.green/models/340248/sinfully-stylish-dramatic-lighting
Hyper 8 step lora: https://huggingface.co/spaces/ByteDance/Hyper-FLUX-8Steps-LoRA
