Team research project · January 2026
Conditioning mechanisms in diffusion transformers
A controlled comparison of adaptive layer normalisation and cross-attention for text-conditioned image generation.
Conceptual illustration; not an experimental result.
Research question
Text can enter a diffusion transformer in different ways. This experiment compares adaptive layer normalisation (adaLN) and cross-attention under the same constrained training setting.
Methodology
The project uses CIFAR-10 images paired with synthetic captions. Both conditioning mechanisms are implemented and examined after a single training epoch, making training constraints an essential part of the interpretation.
Findings
Under this short training budget, the comparison did not establish strong semantic differences between the generated outputs. The report offers a qualitative examination rather than a quantitative image-quality benchmark.
Limitations and further work
Longer training, richer captions and quantitative image metrics would be needed to make stronger claims about conditioning quality. This project is an exploration of architecture and experimental design, not a production image generator.
Related research
Dreaming agents