Conditioning a diffusion model on external inputs involves modifying the model architecture or training process to incorporate additional data that guides the generation process. This is typically done by embedding the external input into the model’s layers or modifying the diffusion process to depend on the input. For example, text-to-image models like Stable Diffusion use text prompts as external inputs by encoding them into embeddings and integrating them into the model’s cross-attention layers. These embeddings influence the noise prediction step at each denoising iteration, steering the output toward the desired content described in the text.
One common method is to use a conditional encoder that processes the external input (e.g., text, class labels, or images) into a latent representation. This representation is then concatenated with the noisy input or injected into the model’s layers. For instance, in class-conditional image generation, a label embedding is combined with the timestep embedding and fed into the model’s residual blocks. Similarly, in audio generation, a spectrogram or MIDI data could be encoded and used to condition the model to produce music matching specific patterns. The key is ensuring the model learns to associate the external input with the corresponding output during training by exposing it to paired data (e.g., images and their text descriptions).
Another approach involves modifying the diffusion process itself. Techniques like classifier guidance use a pretrained classifier to compute gradients during sampling, which adjust the denoising steps to align with the external input. For example, if conditioning on an object class, the classifier’s gradients push the generated image toward higher confidence for that class. More recently, methods like ControlNet allow fine-grained spatial conditioning by training auxiliary networks that process inputs like edge maps or segmentation masks. These networks output feature maps that are fused with the main diffusion model’s features, enabling precise control over composition. Developers can implement these strategies using frameworks like PyTorch by extending existing diffusion architectures to include conditioning mechanisms in their forward passes.