AI Slop Is a Craft Problem, Not a Tech Problem

We learned this the long way

We are an animation and visual effects studio. Before any of the current tooling existed, our people were on set gathering plates and lens data, sitting in suites building composites in layers, keyframing performance by hand, and grading in ACES for masters that had to survive a cinema screen.

That is not nostalgia. It is the reason we can say anything useful about AI slop.

Photorealism is unforgiving. A frame either holds or it does not, and the difference is usually a few pixels of edge treatment, a half-stop of light that should not be there, a contact shadow that never got the attention it needed. You learn to see those things by being wrong about them for years, in front of supervisors who were wrong about them for years before you.

That accumulated seeing is the whole job. Everything else is machinery.

What slop actually is

Slop is not a rendering failure. Most of what gets called slop is technically clean. The lighting is plausible, the camera move is smooth, nothing is visibly broken, and it still feels like nothing.

It feels like nothing because nobody decided anything.

Somebody described what they wanted, a system returned something adjacent to it, and that was accepted because it looked approximately right. No one asked whether the light was motivated. No one asked what the shot was for. No one rejected thirty versions before finding the one. The output is the first plausible answer rather than the right one, and audiences can feel the difference even when they cannot name it.

This is why slop is a craft problem. You cannot direct what you cannot evaluate. A studio without the years behind it does not know that the shadow is wrong, so it ships the shadow. The tool did not fail. The judgement was never there to begin with.

How we actually use it

We use AI heavily, and we are specific about where.

It runs on the labour that never carried authorship. Roto and paint. Cleanup. Matchmove. The passes that consume artist-weeks and that no audience has ever felt the presence of a human being inside. Removing that work from a schedule does not remove anything from the screen. It returns time to the shots where taste is visible.

It runs on iteration. Our directors can look at thirty versions of a sequence in the time a traditional boards-and-animatic pass produced one. That is not thirty answers. That is twenty-nine rejections and one decision, made faster and with more of the field seen.

It does not run on the decisions. Not what the shot means, not where the camera goes, not whether a performance is true, not which of the thirty. Not the grade's intent. Not the final look at any frame carrying a client's name.

An artist reviews every frame that comes back from an automated pass. The pipeline proposes. The supervisor decides. That order is the studio.

Amplified, not replaced

The phrase we use is that creativity is amplified by intelligence, and the order in it is deliberate.

Amplification requires something to amplify. Point these tools at a studio with twenty years of knowing why a shot works and they compress the schedule without touching the standard. Point them at no studio at all and they produce exactly what they are asked for, which is the problem, because the brief is never the answer. The brief is the starting position. Everything of value happens in the distance between what was asked for and what the work turned out to need.

That distance is craft. It cannot be prompted, because it is made of judgement about things nobody thought to specify.

Why we can show this instead of claiming it

Every studio now says "human-led, AI-assisted." It costs nothing to say and cannot be checked, which is why it has stopped meaning anything.

We can be specific because the tooling is ours and it is public. Crucible cooks Houdini digital assets lazily at Karma render time, so an effect stays live and procedural right up to the render and an artist keeps adjusting while looking at the finished frame. houdini_usd_gsplat brings 3D Gaussian Splatting into Houdini Solaris as first-class USD primitives, so photographic capture composes with everything else in the scene.

Both are MIT licensed and on GitHub. Read them. What they do is remove waiting. Neither of them makes a single creative decision, because that is not what we built them for.

This is the same standard we would apply to anyone: not do you use AI, but show me what it does, and show me who decided.

If you are commissioning work

Five questions worth asking any studio, ours included:

1. What did you make before these tools existed? Craft is the foundation. There is no version of this where the foundation is optional.
2. Which steps are automated, by name? A studio that answers in tools and tasks has thought about it. "We leverage AI across the workflow" is not an answer.
3. Who reviews the output, and when? There should be a named role, not a process.
4. Is your tooling built, licensed, or prompted? All three are legitimate. The answer tells you how much control they have when the work needs to change.
5. Show me the shot before and after the automated pass. If the automation is doing invisible labour, this is a comfortable request.

The short version

The tools are not the risk. The absence of a studio behind them is.

We came up making frames that had to hold on the largest screens available, we still work to that standard, and we use every tool we can build to get there with more attempts and less waiting. What we do not do is hand over the decisions, because the decisions are the work.

Creativity amplified by intelligence. In that order, always.

If you are making something that has to hold up, talk to us — or look at the visual effects and animation and film and commercial production work first.