Here's a <a target="_blank" rel="ugc" href="https://www.facebook.com/groups/114384998647935">wonderful resource</a> with some lesser-known Franklin Booth works. A list of training images and Kohya_ss settings can be found in the attached training data file. I can share the training images on request. <h2 id="flux-1.0-dev-chsgx0xy6">Flux 1.0 Dev</h2>Same dataset as before, but trained only with a trigger phrase, `detailed pen-and-ink illustration by franklin booth`. To my surprise, it seems flexible enough. His style is detailed, so it helps to find a way to render as large as possible.By default it seems to prefer large white spaces (in the dataset, the ones with white spaces, they were not as big), but I think this can be mitigated by experimenting with prompting and by being explicit about how you want something drawn (e.g. "hatching", "crosshatching", "stippling"). For example, something like `Their shirt has dense thin hatch contour lines.` If you think it is undertrained or underpowered and think it should be retrained, feel free reach out to me.<h2 id="sdxl-xgw7oipez">SDXL</h2>Made for <a target="_blank" rel="ugc" href="https://civarchive.com/bounties/2353">this</a> prematurely closed bounty.Personally, I think the rank32 version is more well-behaved and detailed, but rank16 might have more appeal and flexibility (although maybe it's just because I spent more time testing 16 than 32). The rank16 version can be weak at times, sometimes coloring images or generating photorealistic images. Try adding words like 'monochrome' or 'lithograph' to the positive prompt and 'photo' or 'colored' to the negative prompt. I'm not sure if they are really good at landscapes (especially clouds and trees which frequently occurred in Booth works) as the results at SDXL standard image dimensions can make the line-work difficult to make out.I also tried training a Pony V6 model too, but they were too difficult to control to release. (I think I didn't understand tagging for pony well enough.)

Franklin Booth Style

LoCon DoRA rank 16 with alpha 1. Convolution rank 8 with alpha 4. Max training resolution 1536x1536.

franklin_booth_style_sdxl_v7.safetensors

labels_v04.zip

SDXL 1536 - Rank16

Flux 1.0 Dev - Rank16

SDXL 1024 - Rank32

LoCon DoRA rank 32 with alpha 16. Convolution rank 16 with alpha 8. Max training resolution 1024x1024.

Files with SHA256: 030f6b9f863aec755eca4c24b1ec7c439b0cb6df358334c1badc246d440f0660