ComfyUI vs Diffusers: Mellon's experimental workflows

Learn why Mateo moved to Diffusers and how Mellon adapts model graphs, with an SD 3.5 example using int8 quantization and weighted prompts.

Player not loading? Watch on YouTube

Mateo explains why he moved his experiments from ComfyUI to Diffusers. He still recommends ComfyUI for general tinkering, but says its structure increasingly obstructed his own development work. Diffusers suits him because he finds it easier to customize, despite frustrations with its overall design.

His flexible node interface, Mellon, is a proof of concept rather than software ready for general use. The demonstration pairs it with Modular Diffusers, which builds pipelines from reusable blocks. Switching between Z image, SDXL and Flux 2 changes the graph's inputs and settings. Another example loads a workflow from Hugging Face into a single node whose fields come from a JSON definition.

A separate experiment customizes the standard SD 3.5 pipeline. Mateo loads components separately for int8 quantization, weights scene and style prompts, averages their tensors and adds noise to embeddings. He also sets guidance to five before disabling it at 35% of denoising. These are choices in his experimental workflow, not general recommendations.

The planned Mellon rewrite demonstrates immediate updates for browser-side nodes and connections to multiple servers with separate dependency environments. For readers exploring local AI workflows, the video explains the design and experiments rather than providing a complete setup walkthrough. It closes with Mateo's concerns about reviewing coding-agent contributions to open source projects and the industry's reliance on cloud services.