This LoRA is an experiment in translating a 2D animated style into photorealism. Using cartoon image to image generation, I utilized targeted prompt engineering to generate photorealistic variations of the character while maintaining facial stability and feature consistency.
From that generation run, I curated a dataset of roughly 60 images to train this model on Marge Simpson of "The Simpsons (1989)" who is canonically in her mid 30s. This was a multi step process of training a beta model with a some images for facial consistency then expanding training content to the full 60 images. This specific version is trained at ~2,000 steps. Because the source material is a highly stylized 2D cartoon, the measure of success and interpretation of a "realistic" counterpart is inherently subjective. I welcome your feedback and test images!
Description
v2.0 is the first release version as v1.0 had inconsistent body proportions.
Comments (4)
can confused by your strategy with this approach, was this some kind reinforcement training experiment to maintain this photorealistic visual aesthetic of marge simpson? I assume then its positive and high quality synthetic images of successful results were used to train this lora. The strategy of using synthetic images for lora training isn't bad as long there's quality review oversite in the selection of images to used.
I used a combination of custom cartoon-style character LoRAs at low weights (0.3–0.5 strength) alongside detailed prompting and other style LoRAs to push the outputs toward realism. From these, I selected the best 20–60 images to train an initial realistic LoRA. Then, as needed, I generated a new batch of images and selected those that most closely matched the target body shape and features to train the final LoRA.
For example, when generating the initial images to achieve very pale skin for the Morticia Adams LoRA, prompting alone caused too much inconsistency. Once I incorporated a skin-tone slider LoRA, however, I was able to generate enough consistent synthetic images to successfully train the LoRA.
@badbta thanks for the elaborate explanation to your process. whenever i had to use synthetic data for training, it was for unpopular, underrated and obscure characters with very limited images and screenshots of them online or in their original media source. Impressive how well krea2 trains and learns when given a good amount high quality images whether authentic or synthetic.
@obinna7713 Very true, the model is impressive.