"The thing that makes ComfyUI absorb any model is the same thing that makes it exhausting to use every day."
The Tool That Has No Ceiling
If the monolith's problem is a hardcoded shape, the obvious fix is a tool with no hardcoded shape. That tool exists: a node-graph image generator where every step — load model, encode text, sample, decode — is a separate node you wire together. New model family? Add one new node type. The internals are so modular that nothing is baked in. The ceiling is gone.
Why Modularity Lives in the Node Graph
The reason a node-graph tool can absorb any model is that its internals are genuinely separated: VAE, text encoder, backbone, sampler, and conditioning are independent units with clean interfaces. The node graph is just that separation made visible — each box on screen is one of those modular units. This is the key insight Ember steals, and the next track is entirely about it.
The Cost: A Graph Is Not a Daily Surface
Here's the trade-off. To generate one image in a node-graph tool, you compose (or load, or tweak) a graph: this node feeds that node, set this parameter here, rewire that edge there. For a power user building a custom workflow, that control is a gift. For someone who just wants to draw and occasionally say "refine this region," it's friction on every single generation. The flexibility and the daily-use cost are the same feature.
The False Choice
So the landscape offers a false choice: a simple tool with a low ceiling, or an unlimited tool with a heavy daily surface. Most people pick one and live with the downside. The forcing function is realizing you don't have to — that the modular internals (from the flexible tool) and the simple surface (from the comfortable tool) are separable, and nothing stops you from taking one of each.
What Ember Takes From Each
Ember's design is explicit theft from both: take the deep module separation that gives the node-graph tool its unlimited ceiling, and put a simple recipe-based API on top the way the monolith does — minus the monolith's hardcoded shape. Simple outside, modular inside. The node graph stays in the toolbox it came from.