Type | |
Stats | 494 |
Reviews | (64) |
Published | May 16, 2023 |
Base Model | |
Training | Epochs: 12 |
Trigger Words | sdn |
Hash | AutoV2 A534C89053 |
If you enjoy my models and want to help support my releases, I always need a coffee :)
Shh...don't tell Tiron about this!
This is a new LoRA version trained on the same data as my 2.1 checkpoint. These behave very differently! If you want a high fidelity model that sticks closer to the original look, that model works better. This sacrifices the fidelity of that model, but retains the concepts and characters and improves many of them significantly. I recommend using this version moving forward.
Please do not repost or share elsewhere without my permission. If you want to do merges or whatever, feel free, just please don't publicly share without asking or sell them. Also, please don't do anything creepy or gross with this model, and keep the NSFW to yourself (I haven't trained or tried this model that way).
Please DO share your results!
This was trained on the remastered shots from What We Left Behind, on the 1.5 base model at 768 resolution on the 1.5 base model.
This LoRA can stack on just about any base model with manipulation, but for some of my recommendations, please check out my images and these especially (still testing) -
Dreamlike Photoreal - for the most precise results to the original show/training data in my experience
Realistic Vision - Most flexible model, photorealistic characters and dramatic look
Cine Diffusion - dramatic, analog film look, as if it was from a DS9 movie
Models I don't recommend (still testing) -
ChilloutMix - Bad character cohesion
You really need to be a prompt engineer to push this model to its limits. All of my image results are not inpainted, did not use Controlnet/etc., and were not edited in any way. Layering in those will help you get even better results. Please look at my shared results, which include all of the prompts and negative prompts for a wide range of applications.
Pros + cons of using this vs. my checkpoint version -
PROS
Flexibility, ability to stack within any other model
Characters look signficantly better with less prompting
Works at most resolutions flexibly
Works with other 1.5 LoRAS, TIs, etc.
CONS
Loss of original style cohesion
Poor performance on base 1.5 model
Models may produce vastly different results
Less coherence to concepts like uniforms, backgrounds, etc.
Here is my basic prompt. This is based on how I captioned the training data (starred are optional):
sdn, a [closeup/medium/wide] view of [character or object], [species], [gender*], [expression*], wearing [Bajoran/Starfleet/Cardassian] [Operations/Science/Command/Security] uniform*, [lighting descriptors (diffuse glow/contrast lighting/etc.]*, [inside/outside], [location (operations/planet surface]*, [background (blurry/viewport/panel/etc.]* <lora:diffusiondesign_SDN_LoRA_1.12:1.0>
Use “sdn” as the initial keyword. Not technically necessary, and sometimes overweighs training appearance, but keeps coherence of concepts and characters better.
LoRA scale of anywhere between 0.25 and 1.0 work well. Will be very dependent on your prompt.
Use terms like “star trek” “ds9” or “screenshot” etc. to further push certain concepts.
CFG of 7 is good for general purpose. Higher for concepts that vary away from original references. Sometimes, things look a bit too intense and contrasty. Take CFG to 5-6 to get a bit of the original film glow back.
Negative prompts help push results towards more aesthetically pleasing generations on average. My personal go-to is:
bad anatomy, bad proportions, blurry, cloned face, deformed, disfigured, duplicate, extra arms, extra fingers, extra limbs, extra legs, fused fingers, gross proportions, long neck, malformed limbs, missing arms, missing legs, mutated hands, mutation, mutilated, morbid, out of frame, poorly drawn hands, poorly drawn face, too many fingers, ugly
Of course, sometimes you do want deformed or ugly results, so adjust as you need. “Blurry” reduces natural fuzziness of original training so also optional (I negative prompt “blurry” and positive prompt “diffuse glow” for example, to sharpen and keep effect)
Uniforms, specifically Starfleet uniforms, should be keyworded, such as “Bajoran Security Uniform” and “Starfleet Command/Science/Operations Uniform”. Starfleet uniforms don’t always come out in the correct colorways, but can be inpainted if necessary. I also recommend using slight prompting to push what you want, adding in “red black and grey” etc. for example. For starfleet uniforms, Worf’s sash will occasionally appear without prompting. Negative prompt “sash” may help.
Comm badges and pips can sometimes do best with inpainting. Comm badges may appear in duplicates. Comm badges in the LoRA may appear less predictably.
My initial training overweighs Bajoran nose ridges in all species, but especially humans. Use "(bajoran:1.2)" as a negative prompt if you're having them appear a lot. Inpainting usually works well to get rid of them if necessary.
Including actor names can really push characters, especially Miles, Ezri, Nog, Leeta, etc. and I recommend it for basically everyone.
Adding “beautiful” and “handsome” etc. can make images look better.
Bajorans
Using “vedic” does somewhat produce the look, but it also skews generations toward real-life concepts of vedics. Using “Indian” as a negative prompt can help.
Even though I trained and re-trained on Kai Winn, she doesn’t really come out in generations.
Cardassians
The model is trained on Garak, Gul Dukat, and other Cardassians but doesn’t seem to want to “separate” them and most of the results resemble Dukat.
Changelings
You'll need to re-roll a lot to get them to look right.
Ferengis
It really doesn’t understand the difference between them, even though it was trained on Quark, Rom, Nog, Ishka, etc. Actors' names are necessary. Nog comes out best.
Use “uniform” in negative prompts for stronger “Ferengi” clothing.
Humans
Sisko is by far the best trained subject.
Keiko doesn’t really come out even though she was in training data.
Jake Sisko and Kasidy Yates was in the training data but only Kasidy works with the addition of “Penny Johnson Jerald”.
Jem’Hadar
Again, they were trained but definitely not quite enough. They only sort-of resemble them unfortunately.
Klingons
Worf comes out great. Others, not so much.
Use “sash” to get Worf’s sash to generate. It can also be generated/inpainted on other characters.
Worf’s forehead goes a bit wacky but inpaints easily. His nose also randomly turns red. I have no idea why.
Trill
It’s hard to get the spots to pop up. Inpainting is actually difficult with this too, but trying ((leopard spots on skin)) for example can help.
Jadzia and Ezri are both in the training but need actor names to really pop. Unfortunately actor names also lower spots. Spots appear more in LoRA.
Vorta
I trained on multiple Vorta, but it just spits out humans with light skin and black hair and almost gets the ears. More training will be necessary here, unless you can prompt engineer them.
Other Species
No training on Andorians, Benzites, Betazoids, Bolians, Orians, Vulcans, Romulans, etc. Some would pop out with other embeddings or keywords that the general training data knows.