The pipeline as a system
A latent-diffusion generator is not one monolithic file. A typical pipeline combines tokenizers and text encoders, a latent image representation, a learned denoiser, a scheduler/sampler, and a VAE decoder. Interfaces may package these parts into one checkpoint or expose them as separate nodes, but the conceptual stages remain distinct.
The Glossary definition explains why work happens in compressed space. This dossier focuses on how that choice organizes a real generation stack and where compatibility failures occur.
Prompt encoding and conditioning
The tokenizer maps prompt text to token IDs; one or more text encoders turn them into learned vectors. Cross-attention or joint-attention blocks make those vectors available to the denoiser. Classifier-free guidance combines conditional and less-conditioned predictions to strengthen prompt influence. Negative prompts, pooled embeddings, and size metadata depend on the model family.
Image-to-image begins from an encoded source latent plus noise. Inpainting adds a mask and often extra channels or conditioning. ControlNet, IP-Adapter, pose, depth, and reference systems add further signals; the base family must know how to receive them.
Iterative denoising and decode
The denoiser is called repeatedly at noise levels chosen by a scheduler. A sampler implements the numerical update from one latent state to the next. Steps, scheduler, seed, and guidance affect the trajectory, but none inject missing knowledge into the weights. At the end, a VAE decodes the latent to pixels; VAE precision, tiling, or mismatch can create color and detail artifacts.
ComfyUI makes this graph explicit. AUTOMATIC1111, Forge, Fooocus, and InvokeAI package similar components behind different controls. The interface can optimize execution, but it does not make an incompatible adapter or VAE compatible.
Performance and adult-AI relevance
Memory includes the denoiser, text encoders, VAE, intermediate activations, and optional controls. Higher resolution enlarges spatial tensors; additional ControlNets or hires stages add peaks. Offloading moves components through RAM and can lower VRAM at a speed cost. Quantization changes representation and requires supported loaders.
Pipeline visibility matters for reproducible adult-image and character workflows: record the exact base, VAE, adapters, controls, seed, dimensions, sampler, scheduler, and software version. Local processing can reduce exposure of sensitive inputs, but custom nodes and cloud-connected components must be audited separately.