Animators often draw characters with three fingers and a thumb for practical and artistic reasons:
Simplicity and Efficiency: Drawing one less finger saves time and effort, making the animation process more efficient and cost-effective.
Visual Clarity: Fewer fingers mean simpler, clearer hand movements, which are easier for the audience to understand and follow.
Tradition and Influence: Early animation pioneers set a precedent with this design choice, and it has become a traditional practice in the industry.
Anthropomorphism: Characters with three fingers and a thumb strike a balance between human and non-human traits, making them more relatable and unique.
Cultural Considerations: Some cultures have specific associations with the number of fingers, influencing animators to adopt this design to be culturally sensitive.
By drawing characters with three fingers and a thumb, animators achieve practical, cost-effective, and visually appealing results while maintaining tradition and cultural sensitivity.
Fast forward to today, and generative AI art has come of age
There are base models and LoRAs (Low-Rank Adaptations) for every occasion, but most, if not all struggle to render a hand with four fingers and a thumb accurately.
The Theory (it was just a passing thought)
My theory is that these base models and LoRAs are trained on a mixture of photographs, screen grabs, digital drawings, etc., meaning that some cartoon images with only four fingers have possibly been used alongside genuine images of hands with five fingers. This mix could confuse the AI training data, leading to poor hand rendering. Furthermore various mixes of models have flooded the market, but with initial bad hand data intact they are potentially compounding the problem.
Has anyone trained a base model from scratch using only photos or only cartoons? Are the results better or worse at rendering hands accurately?
Discussion is open on whether it's worth exploring this further. Enlighten me and the community with your thoughts on hand generation and potential ways to improve it or even if you think my theory bears any creedence!