Developing a Non-AI Image Generator: An Optimization Approach
The current trend of generative AI relies on massive datasets and extreme RAM consumption. We are exploring a different path: creating a generative system without neural networks.
The Concept: Instead of training a model on billions of images, we are focusing on modular composition. By defining specific foreground elements (characters, cars, buildings) rendered from common angles, we can programmatically stitch and composite these assets to create complex scenes. This significantly reduces the storage and computational overhead.
Technical Challenges & Solutions:
Data Management: Initially, we tried per-iteration file requests, which proved disastrous for disk I/O. We moved to an indexing system, but it remained a bottleneck.
Current Solution: We are now loading entire compressed modules into memory. By keeping the footprint in the low-megabyte range rather than gigabytes, we have achieved much faster performance.
How it works (Early Stage): We are currently focusing on specialized blending techniques:
- Patch Blending
- Block Blending
- Dot Diffusion
- Live preview
- Main prompt, variant and negative
- Random seed
- 40mb zip data
- Tag Discrimination
