ComfyUI

ComfyUI (OSS) (comfy-org, 135.5k★, GPL-3.0, Python) — “the most powerful and modular diffusion model GUI, API and backend with a graph/nodes interface. The fastest local inference engine in the world.” The concept worth isolating: generation as a dataflow graph — instead of a form with sliders, a pipeline of nodes (load checkpoint → encode prompt → sample → decode → save), each node a composable, inspectable, reusable step.

Why the graph wins

• The workflow is the artifact: a ComfyUI workflow is a JSON graph — versionable, shareable, diffable. Compare with prompt-in-a-form tools where the recipe is unrecoverable from the output. The file format doubles as the sharing format (workflow JSONs circulate like recipes).
• Subgraphs — a packaged chunk of a flow (“blueprint”) reusable as one node; the composition unit between node and full workflow.
• Every model family plugs in: image (SD/FLUX/Ming/Qwen-Image), video, audio, 3D — the Small Vision Language Models generation side runs through it (Ming-Image ships first-party ComfyUI packaging; Comfy-Org republishes weights).
• API/backend, not just GUI — the same graph runs headless; workflows become services.

The ecosystem’s real problem (and the 2026-09 fix)

The known pain isn’t quality or speed — it’s assembly: “玩 ComfyUI 最耗时间的从来不是出图，是搭线” (the time sink is never generating, it’s wiring). Shared workflow JSONs arrive with missing nodes and mismatched versions. The @Ryrenz-reported fix: Comfy-Org open-sourced workflow_templates (OSS) (1k★+ in days) — the official workflow templates + subgraph blueprints with model metadata, node version info, and thumbnails, importable as runnable starting points, plus an Astro static site cataloging them in 11 languages. “搭线的时间省下来，才是真正开始创作” — time saved from wiring is when creation actually starts.

Local angle

• 3090-class: ComfyUI is the natural serving layer for Ming-Image-0.1-design (int8) and Qwen-Image-2.1 — the Small Vision Language Models A/B (design-niche vs unified editing) would run through ComfyUI workflows, making the comparison itself versionable graphs.
• Same instinct as the rest of the local-AI stack: llama.cpp for text, Parakeet for STT, ComfyUI for pixels — each the open, scriptable, local-first layer of its modality.

Related: Small Vision Language Models · Local LLM inference · Inference engines (the text-side analogue of the “which engine” question)