中文版本：AIGC 狂潮里的垃圾与金子：好作品不是生成出来的，是工程出来的
Core argument: Generation models sample from a learned distribution, and an uncontrolled sample lands near the statistical average—which is why one-click AI content is mediocre by default, not ugly, just bland. What separates the works that stop you from the ones you scroll past isn’t prompt quality. It’s control, consistency, and process. Great AI work is never generated; it is engineered—and the engineering is the same systematic engineering film studios have used for a hundred years, with the judgment layer still owned by humans.
1. Two Worlds in One Feed
Open any content platform today and you’ll hit the same thing: AI-generated content. It all looks alike—glossy frames, overworked compositions, vaguely familiar copy, a tone that rings false in one second. LinkedIn has shipped a “seems like AI slop” report button. Snapchat announced it will stop recommending wholly AI-generated videos. The music industry is debating whether the number-one song on the charts is synthetic slop. Google Earth’s AI image tool was taken down within a day of launch.
The flood is real. That part needs no proof.
But if you scroll long enough, you’ve also seen the other kind: an AI video with an actual story, deliberate camera movement, film-grade texture. You stop. You watch the whole thing. And you think: this was made by AI?
Same tools, same era, two kinds of output. What accounts for the difference?
Most discussions stop at “your prompts weren’t good enough” and go no further. This essay intends to take the whole thing apart: why the garbage floods in by default, what the gold actually has in common, what “understanding the principles” concretely means, what a real workflow looks like, and the question underneath all of it—what AI creation actually is.
Here is the thesis up front: great work is not generated. It is engineered. AI isn’t incapable of good work. Most people simply never treat it as a craft that must be studied. This is exactly how filmmaking works—process and standards are what make good films. The essence hasn’t changed: it’s systematic engineering.
2. Why One-Click Generation Is Mediocre by Default
Let’s start with where garbage comes from. It begins with how the models work.
A mainstream generative model is, at bottom, a machine that samples from a high-dimensional distribution. It has compressed billions of images and billions of texts into a probability distribution, and every generation draws one random sample from it.
That mechanism determines three important things.
First, uncontrolled generation necessarily lands near the statistical average of the distribution. By the laws of probability, random samples cluster where the density is highest—the region that looks “most like everyone.” In other words, generation without control is inherently generic. The output isn’t ugly; it’s bland. Think of the composite face effect: average a thousand faces and you get a face that is pleasant but completely unmemorable.
Second, zero barrier to entry means supply explodes, and the quality distribution is a long tail. When generation cost approaches zero, everyone becomes a producer and output runs toward astronomical numbers. Quality always follows a long-tail distribution—the vast majority lands in the mediocre zone, and the striking work sits in the tail of the tail. For a hundred years, craft thresholds kept most people out, and what we saw of the creative world was the filtered head. Today the filters are gone and the entire unfiltered sample hits the market at once. The garbage ratio looks terrifying. Not because AI got worse—because the filter disappeared.
Third, the platform algorithms amplify the mediocre. Recommendation systems reward volume, and volume rewards generic, low-risk, everyone-looking content. So the algorithm crowns blandness itself, and with search and feeds converging on sameness, the slop completes its loop from production to distribution.
That’s why LinkedIn needed a slop button, why Snapchat had to change its feed policy, why the music industry is policing its charts. These platform moves aren’t moral posturing. They’re the forced reaction of markets whose basic experience is being polluted.
Here’s the uncomfortable part: slop is not a product of AI. It’s a product of the shortcut. People who treat creation as “press the button” will produce garbage with any tool. AI just multiplied the rate a hundredfold—where one person used to produce three mediocre pieces a day, now they produce three hundred. The quantity changed; the quality never did, because the quality switch isn’t in the tool. It’s in the user.
3. What the Gold Has in Common: Marks of Judgment and Four Layers of Consistency
Now the work that stops you. It has a shared feature: you don’t first think “was this made by AI?” You think “this is good.”
Break it down and you find three things.
Story. It has a beginning, a middle, and a turn; an emotional arc; it knows when to build and when to hit. Even thirty seconds long, it has a complete narrative structure rather than a pile of flashy shots. Narrative isn’t a prose problem, it’s a structure problem: an effective opening hook, one mid-story emotional turn, a closing that lands. Every one of those is a designed decision.
Camera language. When the lens pushes in, pulls back, cuts, how the angles are chosen, how the pacing is controlled. Camera work is attention engineering—where the eye goes, where the emotion goes. Attention and emotion are the two most expensive resources in creation.
Texture. Unified lighting, composed frames, consistent detail—the same face across shots, the same hour of daylight across the scene, no extra fingers materializing. A pretty single frame is nothing; consistency across the whole piece is everything.
All three share one source: they are products of judgment. Behind every shot, every cut, every omission stands a human decision. A work is, structurally, layers of stacked judgments—narrative judgments, visual judgments, rhythm judgments, curation judgments. The reason slop is slop isn’t that the generation quality is poor. It’s that it carries no marks of judgment. Nothing was ever decided.
One engineering level deeper, the common denominator of great work collapses into a single word: consistency.
- Character consistency: the same face must not drift between shots.
- Style consistency: image, color, and light belong to one aesthetic system.
- Temporal consistency: motion, physics, and logic hold across frames.
- Content consistency: narrative, character logic, and worldview hold across the whole piece.
These four layers are precisely the hardest part of AIGC technology today, and precisely the dividing line between industrial-grade work and random generation. Notice that “cinematic quality,” translated into engineering language, simply means “all four layers of consistency pass.” The film industry spent a century solving these problems—makeup artists, continuity supervisors, lighting departments, script doctors—all in service of one sentence: controlled consistency across time. AI brought the cost of generating a single frame down to nothing, but consistency across time is still the most expensive thing in the whole chain.
4. The Principles Layer: What Users Actually Study
“Understand the principles” sounds vague. Concretely, it’s four blocks.
Block one: model nature. Know what you’re holding. Is it a diffusion model or an autoregressive model? What is it good at, bad at? Why is the default output generic? When does it hallucinate? Understanding that “generation is sampling from a distribution” tells you why control is mandatory—because without it, you get the average, guaranteed. This is lesson one for every user, and the lesson most people skip.
Block two: the control hierarchy. Control is not just a prompt. The full hierarchy looks like this: global layer—style, worldview, atmosphere—locked with style references, LoRAs, and global parameters; local layer—composition, shots, actions—handled with prompt structure, camera language, and reference images; micro layer—fixing exactly what’s wrong—with seed-locked frames, targeted regeneration, frame-by-frame correction. Real users understand that a good piece is not generated in one pass. It’s built by stacking three layers of control—global lock, local refinement, micro repair—each pass squeezing randomness out of the image.
Block three: consistency solutions. This is the deepest block. Character drift? Reference images plus a character LoRA plus fixed seeds. Style drift? Style anchoring, unified lighting parameters, one consistent generation environment. Temporal drift? Lock global settings across segments, align in latent space, unify color in post. Every one of these solutions is, in effect, a fixture—the same thing a factory uses to hold a part in place. Parts without fixtures are scrap; pieces without a consistency plan are footage, not a work.
Block four: tool boundaries. Which errors the model can’t fix and must be handled in post; which steps no model will ever replace—sound, editing rhythm, narrative structure. Know the boundaries and you stop wasting generation budget in the wrong places.
Compress all four blocks and the principles reduce to one line: know what is controllable, what is luck, and how to pin luck down when it happens. People who treat AI as a partner study control. People who treat it as a wishing machine study incantations. That difference is the difference between a director and a gambler.
5. The Workflow Layer: A Nine-Ring Engineering Pipeline
Once the principles are in hand, the next move is turning understanding into process. The difference between one-shot operation and a workflow is the most important thing in this essay to remember.
One-shot operation: have an idea, open the tool, generate, post. One person, one step, one take. It’s fast—and every output is a fresh gamble. Lucky, you get something good. Unlucky, another piece of slop. Quality is handed entirely to randomness.
A workflow breaks creation into repeatable, inspectable, improvable stages. For an AI video, a complete pipeline looks like this:
Ring one: topic and positioning. Who is this for, what does it say, what form does it take. Get the direction wrong and everything downstream is wasted.
Ring two: script. Story structure, hook, emotional arc, what the ending leaves behind. If the narrative doesn’t hold, everything after it is wasted effort.
Ring three: storyboard. Translate the script into camera language—shot sizes, camera positions, movement, seconds per shot. This is the translation layer from “writing” to “shooting,” and it’s where judgment density is highest.
Ring four: generation strategy. Which model, which style anchors, how the prompt is structured, what reference images, how batches of attempts are planned. The value of strategy is covering the most possibilities with the fewest generations.
Ring five: selection and rejection. Review every candidate against acceptance criteria: face drift, motion breakage, lighting consistency. A rejection rate above 90 percent is normal, not a malfunction. Selection is itself a stage. A workflow without a selection stage is a factory without quality control.
Ring six: post-production. Editing, music, sound design, unified color grading. Sound and rhythm often decide the final thirty percent of perceived quality.
Ring seven: review. Watch it as an audience member. Find the problems: where it drags, where it breaks, where it’s inconsistent.
Ring eight: iteration. Go back to the relevant ring with a problem list. Rewrite the script, redo the storyboard, regenerate, re-polish—until acceptance passes.
Ring nine: release and retrospect. Watch the data, listen to feedback, and write “what failed” into the next round’s strategy.
Note ring eight—iteration is the heart of the workflow. One-shot operation is a one-time bet. A workflow is a closed loop of generate—verify—rework. The point of the loop is that errors are systematically found and fixed, while open-loop errors are only published, then lost in the feed.
This is also why a workflow looks slower but costs less in total. One-shot operation saves process time, but the cost of publishing garbage is hidden: account weight decay, audience trust loss, expensive invisible rework. The workflow front-loads that cost into the selection and iteration rings—where it’s visible, controllable, and shrinks with practice. Process eliminates randomness, so quality gets a floor. With a floor, the ceiling becomes worth pursuing.
6. The Film Industry’s Legacy: Process, Standards, and Roles
At this point it has to be said plainly: the AI creation workflow is the filmmaking workflow. Not a metaphor—the same thing with a different execution layer.
A hundred years of cinema produced a complete process: script development, storyboarding, shooting, lighting, art direction, editing, scoring, review, distribution. Why? Because a century of experience proved that without process, even a genius director can’t deliver consistent quality—and with it, acceptable work can be produced at scale. Process is not the enemy of talent. Process is the insurance of talent.
The system also left a second inheritance, one that gets overlooked: role division.
- The director—judgment: narrative, staging, final calls. The one role that can never be outsourced.
- The producer—process: budget, schedule, resource organization. The AI-workflow equivalent is project management.
- The cinematographer—vision: shot size, camera position, movement. Maps to storyboarding and frame control.
- The art director—style: light, color, wardrobe. Maps to style anchoring and consistency planning.
- The editor—rhythm: shot duration, transitions, emotional pacing. Maps to post-production.
- The sound designer—texture: score, effects, mix. Sound design is routinely ignored by AI creators, and it is one of the biggest sources of “cinematic texture.”
The most useful insight of the AI era is this: these roles didn’t disappear. They became a capability checklist on one person. A solo AI creator isn’t “generating with AI.” They’re playing the whole crew alone—director today to set direction, cinematographer tomorrow to design shots, editor after that to set rhythm, producer the whole way to run the pipeline. Every role you’re missing shows up as a flaw in the work. That’s the quickest way to judge an AI creator’s level: not by their output, but by how many roles their capability checklist covers.
And “process” in the AI era takes the form of the nine rings above. AI lowered the single-point cost—“shooting” went to nearly free. But the system cost hasn’t dropped a cent: script, storyboard, judgment, taste, review—all the effort that must be spent, is still spent. A director used to spend money with an entire crew. A solo creator now saves money with a full workflow. The money saved is execution money. The money that can’t be saved is judgment money.
7. The Essence: Three Layers of Systematic Engineering
Now the whole framework can be closed. AI creation as systematic engineering sits on three layers, each supporting the next:
The technology layer—tool principles. How the model works, what control dimensions exist, how consistency is solved, where the boundaries are. This layer answers “can it be made?” People who study it treat AI as a tool. People who don’t treat it as a lottery ticket.
The process layer—workflow. The nine-ring closed loop from topic to retrospect, upgrading one-shot operation into controllable engineering. This layer answers “can it be made reliably?” People with process have a quality floor. People without it have luck.
The judgment layer—taste and curation. Whether the story is worth telling, whether the shot should go this way, what stays and what goes, whether it’s good. This layer answers “is it worth making?” It is the least outsourceable part of the entire engineering effort. AI can replace execution. It cannot replace judgment.
The relationship between the three layers: technology sets the ceiling, process sets the floor, and judgment decides where the work lands between them. Missing any layer, and the work collapses at that point. Studying only technology produces no work; building process without taste produces competent mediocrity; having taste without tools leaves you forever “imagining it perfectly and making nothing.”
This is also the complete answer to the panic that “AI will replace creators”: what gets replaced is the person standing only on the technology layer, treating execution as the whole of creation. The genuinely valuable part of the creator’s identity—judgment—was never in AI’s range. The people panicking are the ones who mistook execution for creation.
8. A Roadmap for Ordinary People
If this analysis is useful to you, here’s a road you can actually walk: three stages.
Stage one: use. Get familiar with the mainstream tools. Internalize that “generation is sampling from a distribution.” Learn basic prompting, seeds, reference images. The goal isn’t output; it’s learning the tool’s temperament. The most common mistake is skipping this stage to chase “making videos”—and ending up treating slot pulls as creation and luck as skill.
Stage two: control. Build the three-layer control concept: global layer locks style, local layer locks frames, micro layer fixes flaws. Start managing all four consistencies—character, style, temporal, content. The sign of this stage: you can point at an image and say “this is wrong because of that, and changing these parameters fixes it.”
Stage three: design. Step out of the tools and into engineering. Build your own nine-ring workflow. Go down the film crew’s role list and fix your missing capabilities. Make taste and judgment your main practice. The sign of this stage: output is consistently above the acceptance line, and every iteration raises the line.
Each stage has different study material: stage one is model fundamentals and tool documentation; stage two is control techniques and consistency case studies; stage three is narrative structure, storyboard language, and audiovisual rhythm—that is, take the film school curriculum and make it your major.
Through all three stages you fight one enemy: the temptation of one-click generation. It is always there, always fast, always bland. The only defense is tasting once what a work engineered feels like. After that, you can’t go back.
Conclusion: One-Click Is the Most Expensive Illusion
Back to the opening.
The flood is real. The gold is real. The difference is not the AI; it’s the user. It’s not the generation; it’s the engineering. AI did not change the essence of creation. It changed the cost of execution—turning “could be made” into “could be pressed.” It did not turn “could be imagined” or “could be recognized” into freebies.
So my judgment on the AIGC era comes down to two sentences.
First: one-click is the most expensive illusion. It looks free, and it charges you the rest of your creative career.
Second: systematic engineering is the cheapest shortcut. People who study principles, build process, and train judgment will produce good work in any era. The AI era just runs the experiment faster.
The tools are new. The engineering is old. People who know engineering never run out of work.