The model is not the pipeline: a local generative UGC workflow

Generative video has reached the point where producing a convincing few seconds of footage is relatively easy. Producing thirty seconds that can actually run as an advertisement is another problem entirely.

I wanted to see how far a locally run video workflow could be pushed toward realistic UGC: the kind of casual product video that looks like it was recorded by an actual creator rather than produced as a polished commercial. The goal was not to make AI-looking video. It was to make ordinary-looking video that could plausibly pass as real.

A useful UGC advertisement has to solve several problems at once. The person has to stay consistent. The performance has to feel natural. Speech has to stay coherent. The product has to stay recognizable, down to the label text. And when something does go wrong, the footage still has to be editable into something usable.

At the same time, generating everything at maximum resolution is computationally expensive. So instead of asking a single model to produce a finished commercial, I approached it as a multi-stage production pipeline.

Seven stages, each responsible for a smaller problem. The first four decide whether there is anything worth finishing.

01 · Start with a controlled character

The process begins with a prerendered image of the person who will appear in the video. For the first test I built a specific UGC persona: a 28-year-old woman with a messy bun, light freckles, an oversized cream hoodie and gold hoop earrings, filmed in a bright apartment kitchen. The reference establishes the visual identity before video generation begins.

The objective is deliberately different from advertising photography. I don't want perfect lighting, perfect composition or a commercial-model performance. I want slightly imperfect, handheld, conversational footage.

02 · Generate the performance

The character reference goes into MiniMax H3 with a production prompt rather than a scene description. It specifies character and appearance, environment, lighting, camera behavior, performance style, product interaction, dialogue, timing, gestures and blocking, audio characteristics, and continuity requirements.

The video is structured into timed sections so the model has a defined progression through the performance: zero to seven seconds establishes the creator and introduces the product, seven to eighteen demonstrates it and communicates the benefit, eighteen to thirty gives the recommendation and the call to action.

The prompt also asks explicitly for the things that make UGC feel like UGC: handheld movement, micro-expressions, natural breathing, conversational delivery, room tone, small imperfections. The intention is to give the model creative constraints, not simply a visual description.

03 · Generate low, select high

One of the most important decisions in the workflow is not generating everything at final resolution. The initial generation runs low, and at that stage I'm only asking whether the performance works, whether the character holds together, whether the speech works, whether the movement looks natural, and whether the shot actually feels like UGC.

If the answer is no, there's little value in spending additional compute to upscale it. Creative iteration and expensive finishing stay separate, and the expensive half stays downstream of the decision.

04 · Iterate instead of fighting the model

A practical lesson from testing H3 is that not every failure requires changing the workflow. The same basic setup can produce substantially different results from another seed, and minor language or speech problems can often be resolved simply by generating again.

That changes the question from "how do I force this generation to work" to "is this generation worth keeping, or is another sample likely to solve the problem faster." As generation quality improves, that distinction becomes more important, not less.

The remaining three stages only ever run on a take that has already earned them.

05 · Upscale after the creative decision

Once the underlying video works, it gets upscaled. The objective isn't simply to make the footage larger: the finishing process has to retain the small details that matter in advertising, particularly product appearance and label information.

06 · Treat the product as its own problem

Generative video can produce a convincing person while simultaneously getting details on a physical product slightly wrong, particularly small typography and packaging details. Rather than requiring the video model to solve everything at once, I use a reference-to-image stage to integrate the product.

Character reference → H3 video generation → low-resolution evaluation → seed selection or regeneration → upscale → product reference-to-image pass → second upscale → editing and quality control

The important idea is decomposition. Each stage is responsible for a smaller problem.

07 · Edit around the failures

The final stage isn't generative at all. It's editing.

A generated video doesn't have to be perfect to be useful. A minor speech artifact might only occupy a fraction of a second. An awkward movement can become a cut. A strange transition can disappear underneath a change in framing. Rather than regenerating an entire thirty-second sequence because of one imperfect moment, conventional editing handles localized failure.

This is where the process stops looking like AI video generation and starts looking like AI-assisted video production.

The remaining issues aren't fundamental failures of the concept. They're the localized problems that still require iteration: occasional language or speech hitches, minor generative artifacts, product and detail fidelity, selecting the strongest generation from multiple seeds, and further refining the finishing and editing process.

That's working well enough to keep pushing toward portfolio-quality output. It is not yet finished.

What isn't here yet is measurement. Generation time, how many seeds it takes to reach one usable take, and the hardware this runs on are the figures that would turn a described workflow into a costed one. Recording them is the next thing on the list, and this piece gets less hedged when they land.

The biggest lesson so far is that the model isn't the production pipeline. A generative model is one component. The useful system is everything around it: reference preparation, prompting, generation, evaluation, iteration, upscaling, product integration, editing.

The goal isn't to make a model produce a perfect advertisement in one shot. The goal is to build a workflow in which imperfect generations can be evaluated, corrected, recombined and finished efficiently. That makes the limitations of the model part of the production design rather than simply a reason the project fails.

I'm interested in what happens as local generative models continue to improve. If a model can produce convincing performances locally, and the surrounding workflow can handle consistency, product fidelity, resolution and editing, the economics of producing large volumes of advertising creative begin to change. This project is an exploration of that possibility.

The interesting question isn't whether AI can make a video. It's whether we can build a system that makes usable creative, repeatedly.