For a long time, we watched AI tools from a distance. Not out of disdain — out of caution. The first generations were unpredictable, barely parametric, and often incompatible with a serious production workflow.
Then ComfyUI landed, with its modular graph logic. And everything changed.
AI didn't replace our artists. It removed 40% of the repetitive work that made them miserable.
In this post I'll detail how we integrated ComfyUI into our existing pipeline (Nuke + After Effects) without breaking what already worked. Three rules we set, three traps we avoided.
Rule 1: AI assists, it doesn't decide
Every AI output passes through an artist before going into comp. No direct generation in production. It's slower, but it's the only way to keep creative control and avoid the "hallucinations" that ruin a shot.
Rule 2: Strict versioning
Every ComfyUI workflow is versioned in git like code. We have a dedicated repo. Artists pull the workflow of the day, modify, commit. No JSON lost on a Dropbox.
Rule 3: Monthly benchmark
We compare AI vs manual on test shots every month. If AI wins by 3x or more, we adopt. Otherwise, we stay on the classic workflow. No religion, just numbers.
The traps
First trap: believing AI is "free". A custom workflow takes between 2 and 8 weeks to dial in and stabilize. It's an investment, not a gadget.
Second: wanting to automate everything. Some tasks are still faster by hand. Clean keying on hair + complex background, for example.
Third: neglecting training. Artists need to learn Comfy, not just "click generate". Otherwise they become dependent on the studio prompt engineer.
Result after 6 months
We gained about 30% of the time on prep steps (rotoscoping, clean plates, upscaling). Artists have more time for final comp, which is the part where their expertise really matters.
Bottom line: positive, without magic. AI is a tool like any other. Used well, it accelerates. Used poorly, it degrades.