AI art is theft, but as picasso said good art borrows, and great art steals. AI artists are no different from other great artists - we steal.
Workflow
https://drive.google.com/file/d/1cWfcvLdEpAPQzqKFtP5Zhk-DJmJ37030/view?usp=drive_link
TRIGGER WORDS
Herbst Photo, herbstphoto
About
As a professional photographer, I have trained a Lora on a set of 500 of my favorite photographs from my personal, (noncommercial) collection that I own the rights to so that you can steal a slice of my aesthetic. As somebody who steals aesthetics from others, I think it is only karmically fair that I am the one who trains and releases a Lora based on my aesthetic.
V3 - Trained with high strength: larger network Dim, 1500 steps, lower learning rate. The file size is larger (500mb) but this is currently the strongest Lora. 100 Images tagged by hand.
All the images in the training data are analog 35mm film: Kodak UltraMax 400, Kodad gold 400, Kodak 250D ENC-2, Kodak 500T ENC-2, Kodak Portra 800, Kodak Portra 400, Kodak Portra 160, Kodak Ektar 100, Kodak T-Mas BW 400, Kodak EXR, Kodak Eastman EXR 5298 500T (1997 Expired) Ilford 3200 BW, Fuji 400, Cinestill 800, Cinestill 50D, Kodak Ektar 100, Kodak ColorPlus 200, Fujifilm Pro 400H, Fujifilm Superia X-TRA 400
***Settings***
Lora strength & Flux guidance.
2.0 is the sweet spot when paired with a Flux guidance of 2.5. This results in the lora having a high strength, giving a balanced amount of imperfections and increasing the tonal difference between shadows and higlights.
.9 is the sweet spot when paired a Flux guidance of 2.0. This results in the lora having a lower strength.
3.0 strength and 4.0 guidance produceces better candid moments, flash photo graphy, and film burns.
3.5 srength and 5.0 guidance produces more abstract images, with blown out highlights and motion blur, while still remaing tasteful
For the strength 0.5 being is the lowest amount to see effects, and 1.5 is the highest without serious changes to the output.
*This version was trained to be quite strong and does not pair well with other loras unless used at it's lowest strength of .9 with a flux guidance of 2.0
I'm not sure why it helps to increase the flux guidance to a ratio of increments of 2:1 to the lora strength. If anyone could elaborate on this concept I would appreciate it.
Scheduler & Sampler
huen & simple - standard baseline
unipc_bh2 & Simple - standard baseline, similar to huen & simple
unipc_bh2 & normal - gives the highest texture by adding contrast and sharpness in the mid tones
unipc_bh2 & ddim_uniform - gives more degredation but tends to alter the output
dpm_fast & sgm_uninform - heavy motion blur and texture but tends to alter the output
Max Shift: 0.0
Base Shift: 8.0 is the sweet spot for 35mm grain texture, 4.0 is a lighter grain, 1.0 is the lightest.
If upscaling, use the SD ultimate upscaler:
Model: 4x_NMKD-Siax_200k model.
Steps:
CFG 2
Sampler: Euler
Scheduler: Normal
Denoise .2
Length - Match to (upscale amount x original resolution)
Height - Match to (upscale amount x original resolution)
*After lots of Testing, I have found that FLux upscaling works best with a tile size that reflects the upscale output resolution at a step of 1. This is good news as it increases the speed of upscaling.
V2 (weakest version)- This version was trained on a smaller data set with 10-20 captions per image to emphasize images with higher imperfections. It is much smaller and will run faster. As well as the trigger words "herbstphoto" and "herbst photo", the model was trained with captions on the following phrases and responds to them well in the prompt: off center, asymmetrical, backlight, light leaks, grainy, grain, film grain, candid, high contrast, film burn, analog texture, partial silhouette, chiaroscuro, highlight bloom, cinema film color, filmic glow, blown out highlights, moody, light sliver, dutch angle, flash photography.
V1 - This Lora is my first and was trained lazily without captions, however it still does a good job of capturing the way a lens the world. It emphasizes asymmetrical composition, strong backlighting, lens flares, candid portraiture, gritty textures, vignetting, and imperfection.
V1 & 2 - This Lora works well at a strength of 2.0 paired with the Xlabs realism Lora at a strength of 1.0
Share your results!
