license: openrail
tags:
- stable-diffusion
- embedding
- textual inversion
Dreamink
A style embedding for Stable Diffusion v2 (768) of striking stark silhouetted landscapes against colourful backgrounds. Not compatible with SD v1 models. Trained on prompted outputs invoking silhouettes and some historical artists from a model merger of Inkpunk Diffusion and Dreamlike Diffusion.
Prompts
Above images settings:
Prompt 1: v2_dreamink, a sailing ship on a prismatic sea
Prompt 2: v2_dreamink, a cozy library full of bookshelves
Steps: 15, Sampler: DPM adaptive, CFG scale: 7, Seed: 752767199, Size: 768x768, Model: Stable Diffusion 2.1 (768)
Prompt: v2_dreamink, an aurora above a glittering tundra
Negative prompt: dull, muted
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 2748781073, Size: 1024x768, Model: Stable Diffusion 2.1 (768)
Prompt: v2_dreamink, an aurora above a glittering tundra
Negative prompt: dull, muted
Steps: 20, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 2748781073, Size: 1024x768, Model: Stable Diffusion 2.1 (768)
Suggestions
- The sharp lines of the DPM++ samplers work well with Dreamink, and I particularly suggest trying DPM Adaptive out.
- Works best with landscapes, haven't really tried this out with characters and portraits and I think it might struggle with those.
- Definitely slightly overtrained on the sci fi influences of Inkpunk, especially with shorter prompts.
Training
Trained and generated in Automatic1111's Webui
Images generated from a model merge of Inkpunk Diffusion and Dreamlike Diffusion at 0.3, then mostly generated with the following template:
Prompt: Subject matter, (nvinkpunk:0.8), (dreamlikeart:0.8), cel shaded, flat, synthwave, chiaroscuro, by Winslow Homer and Nicholas Pocock and N C Wyeth
Negative prompt: dull, muted, boring, modern
Steps: 20, Sampler: DPM adaptive, CFG scale: 7, Size: 512x512, Model: Inkpunk Dreamlike
Then upscaled using the SD Upscale script to 1024x1024, before being autocaptioned with BLIP with the Preprocess tab under Train and the captions fixed to remove references to rainbows and paintings.
Dataset size: 32
Vector size: 4
Initialisation text: *
Embedding learning rate: 0.003
Batch Size: 4
Gradient accumulation rate: 8
Max Steps: 500
saving an image and embedding every 250 steps
Latent sampling method: deterministic