Overview:

Trained on a physical set of 97 beautiful scanned vintage photos, many being candid solo and family portraits dating back to the 1910's-60's, Memoria faithfully reproduces the dust, scratches, and scrapes many images of that period are known for. Beyond basic scuff, due to many of the photos appearing to have been taken by family members, you'll sometimes get endearing qualities like hilariously missed focus, some awkward framing, and a variety of neat era specific milky bokeh and subtle lens distortions that really add a lot of charm to the final output. Thanks to some new approaches on my end, the scanned dataset itself is tack sharp, so image quality should stay crisp regardless. Don't be afraid to push the outside the lora's comfort zone either, I've created some great unholy cinematic nonsense with it as well so feel free to poke and prod.

I really enjoyed working on this one and learned quite a bit in the process. As far as image quality is concerned, it's by far my best model (better scanner + higher dpi). Not to mention I've recently received about 200 more images to scan and add to the dataset so keep your eyes peeled for future updates!

Recommended Settings:

LoRA Weight: Full strength of 1 works great.

Models: v1-5-pruned-emaonly (trained on), realisticVisionV5.1 (VAE), realisticVisionV20_20, ProtoGen_X3.4-pruned-fp16

VAE: vae-ft-mse-840000-ema-pruned

Clip skip: 1

Resolution: Training data was 768x768 and all standard resolutions up to 512 x 1024 seem fine, 1024 x 1024 is where quality seems to fall off.

Sampler: DPM ++ 2M SDE Karras, DPM SDE Karras, DPM++ 2M Karras, DPM ++ 2S a Karras

Steps: 20-30

CFG: 7-8

Example Prompt:

Model: v1-5-pruned-emaonly

Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 8, Seed: 183580767, Size: 456x680

Highres Fix: Upscale by 2x, Upscaler: 4x_NMKD-Superscale-SP_178000_G, Hires Steps: 20, Denoise: 0.36

Posetive: xyzVntg <lora:Memoria_xyzVntg:1>, candid portrait photo, loving kind eyes, joyful, cinematic, (sharp focus:1.1), 4k

Negative: cartoon, painting, illustration

A note on dataset diversity:

One unfortunate quality of the dataset used is a serious lack of racial diversity. This is unsurprising given the times these images were taken and who would and wouldn't have had access to camera equipment back then in the first place, but still, worth mentioning. By default, you're very likely to see white faces, but I found that specifying other skin colors and ethnicities (or even google translating to other languages) worked quite well, so give that a shot if you're looking for more diversity in your output.