The author's concluding evaluation of Generative AI rendering with control nets, reference image and prompt is that speed is the real benefit and control is the cost.
You can explore about 10 visual directions in the time one traditional render takes. Rough layout to fine render takes minutes.
Rerun with a new seed or an adjusted prompt. Tighten the control passes if the geometry drifts (Control-pass choice as a freedom dial (outline vs depth vs blend)).
The author judges the quality impressive for rapid prototyping and previs. He frames it as a fast exploration tool, not a replacement for traditional final rendering. Shot length is a further constraint (Shot-length limit and batch chaining in the AI renderer).
Stefan's Blender + ComfyUI SDXL pipeline (Blender to ComfyUI SDXL image-to-image render pipeline) gives a concrete speed figure for this trade-off. He claims about 10 seconds per image, or 3-4 seconds on a very good GPU. He calls that roughly 100 times faster than a ray-traced render and says it is free and local. The cost is control. Lighting and materials come from the model, not the scene. Colours and floor detail can drift, as seen in his demo. Control is recovered through Flat, shadowless, high-contrast base render, the depth and canny ControlNets (Multi-ControlNet stack with per-net strength (CR Apply Multi-ControlNet)) and Denoise tuning for image-to-image.