ComfyUI generative pipeline

ComfyUI exposes the whole diffusion pipeline as a node graph rather than a prompt box, which makes it the right tool if you want to understand what each stage is contributing.

I built a workflow around the Flux model and trained a custom LoRA on it to get a consistent Ghibli-inspired illustrated style — the interesting problem being style consistency across generations, not one good image.

Running it locally on an RTX 4060 Mobile also meant real constraints on VRAM and batch size, which forced decisions about resolution and sampler that a cloud endpoint would have hidden.