Type | |
Stats | 652 423 |
Reviews | (51) |
Published | Dec 25, 2023 |
Base Model | |
Training | Steps: 3,400 |
Usage Tips | Clip Skip: 1 |
Trigger Words | dw01-3400 |
Hash | AutoV2 B4D8FE4A8F |
Trigger is dw01-3400, but if you rename the .safetensors file you can change the trigger to whatever you want.
e.g.
rename dw01-3400.safetensors to darkworld.safetensors and the trigger becomes darkworld.
This embedding produces trippy/psychedelic/dreamlike results with a night-time vibe.
I've included a gallery showing the effect of with/without the embedding.
The TI will default to a green/purple color palette. If that’s not to your taste, add a color term to your prompt such as “warm color palette”, “pastel color palette” or “triadic colours”.
Tendencies for this TI:
twilight/night/sunset/sunrise
biological rather than technological
glowing dots/bumps/nodules
jellyfish/toadstool type domed shapes
fantastical/magical creatures and plants
auras, halos, auroras, and other glowy things
Adding words such as “glowing”, “vivid”, “magical” and so on will reinforce the embedding.
It's overpowered for short prompts - you really need to be over 20 tokens for controllability. The TI itself is 8 tokens. You can weight it up or down in a1111 webui - I found that weights from 0.5 up to 1.5 were usable. However... I found weighting in a1111 v1.6 and v1.7 to be inconsistent, so I tend to move the TI trigger left or right instead:
nearer the start of prompt = stronger effect
nearer the end of the prompt = weaker effect
A simple prompt that gives some nice images:
a magnificent creature running in a magical landscape, very detailed, intricate, dw01-3400, warm colour palette
For the negative prompt with this TI, I use either nothing or:
toy, replica, scale model, (signature), logo, nude, naked, nudity
CFG: I’ve had good results from 5 to 12.5.
Model: Mostly I use Crystal Clear XL
https://civitai.com/models/122822/crystal-clear-xl
Sampler: mostly tested with DPM++ 2M Karras at 30<--->50 steps, but all the DPM++ samplers gave decent results.
Hires fix: generally I have this on to get more detail in my gens. Usually Latent (nearest-exact), Hires steps = 30, Denoising strength = 0.7. Typically, I upscale by 1.25 or 1.5.
As is usual with Stable Diffusion, changing scale and/or aspect ratio can produce widely varying results!
My aim was to make a TI that produces twilit/night-time dreamlike scenes and entities.
The training set was fully synthetic, generated with SDXL using several LoRAs in combination. Mostly, I used:
Bas Waijers (Art Style) [SDXL]
https://civitai.com/models/180508/bas-waijers-art-style-sdxl
Hieronymus Bosch Creatures + Cracked Paint Style
https://civitai.com/models/188256?modelVersionId=211408
Only black. SDXL