Type | |
Stats | 5,205 |
Reviews | |
Published | Mar 2, 2023 |
Base Model | |
Training | Steps: 3,000 Epochs: 1 |
Hash | AutoV2 3716CF5228 |
Analog Film portraits LoRA
Contents: How to use - Trigger words - Training - Future
Getting a real film look to portrait photo generations was quite difficult, many models produce overblow colors, or overexpose images at base, so I trained a LoRA to work with SD 1.5 models. This LoRA is able to produce a generic film look to generations, giving the subtle film grain/noise to an image often found in film, and is able to replicate the higher dynamic range of film as well. I explain further down why portraits made using SD can often look bright and overexposed.
How to use:
LoRA strength: 0.3 to 0.7 - Higher strengths may produce artifacts
Should work well with any photo realistic model for portraits (see training), the sample images were made with Realistic Vision V1.4 - Have tested on a couple of other models and works well. May be a little over trained, so higher strengths are not advised, will try to fix this in the future.
This LoRA WILL change the shape of faces, backgrounds, and the overall feel of an image, and may produce vastly different results on the same seed. This is due to the training style.
I highly recommend using Control Net and other tools if you want to regenerate a specific image.
For generating photos of certain types of people, try to be specific in your descriptions: “Latin female wearing an oversized hoodie” - “Asian male with blue hair” as otherwise you will get similar looking faces.
This LoRA is not tested or intended for NSFW generations.
Trigger words:
Film - Analog film
Works without trigger words as well, but extra prompts enhance the effect. Also works well with prompts such as: Sunset - Hard light - Soft Light etc.
Training:
This LoRA was trained using around 30 portraits taken on various types of color film stock. Included in the training data was various ethnicities, genders and ages to try to balance its ability to work on any model or person. The LoRA is particularly good at faces, but will often try to put buildings, and different artifacts/items in places they don’t belong at higher strengths. This is due to my laziness during tagging. Will try to fix this for the future, and try to take as much bias to certain types of face shapes and objects out of the model as possible while still producing excellent portraits. Was a little overtrained, for about 3500 steps, this can be helped by reducing the strength of the LoRA in your generations.
For more info on LoRA training check out the channel and video below:
https://www.youtube.com/@Aitrepreneur
Future:
Analog Film Portraits v1.1: Training on a larger data set with better tagging. Less aggressive training.
Neon Lights v1.0: Photographers will often use color gels or colored lights to achieve a “Neon” look to their images. SD is really, really bad at this, often over exposing or adding random neon necklaces/clothing. I’ve already trained a beta version which does a great job of recreating the effect. See an example image here: https://unsplash.com/photos/3TLl_97HNJo
Analog Film landscapes v0.5: Will train a LoRA on non-portrait images of various film stocks to give more flexibility to image generation.
Rembrandt Lighting v0.5: Fairly hard to achieve this kind of lighting in SD with prompting. Will train on a medium sized data set to try to recreate this famous lighting style of more dramatic portraits.
Slide Film v0.5: Slide film has a very obvious color style and grain structure that has always been difficult to replicate digitally. Will beta test some LoRA training as the weeks move on.
Street Photography v0.5: Mostly made using wide angle lenses, and on a variety of subjects. Will attempt to replicate.
Studio Fashion photography v0.5: Fashion photographers shoot in a specific style / angle / focal length, with fairly soft lighting attempting to highlight the clothing rather than the face. Will attempt to train on various styles.
Why this Film LoRA?:
SD is trained on a LOT of images, a cursory glance through the LAION database for “Portrait photo” reveals most images are older, lower quality, and over exposed. Most of the tagging is also fairly poor when it comes to style. Searching for “Film Photo” will also just return some really bad examples, with logos, and posters thrown in. SD doesn’t really have a good data set for certain photographic styles, especially professional level photography. I’m attempting to help SD produce high quality photography in specific styles that will help photographers interested in image generation achieve the look they want.
Feedback is very welcome, please leave some comments letting me know what works and what doesn’t for you. This will be super helpful in training future LoRAs.