Kalshi's marketing team had a problem. Actually, they had 300 to 400 problems—all of them video clips generated for an NBA Finals ad campaign earlier this year. Only 15 were usable.
The sports betting company had spent roughly $2,000 running prompts through Google's Veo 3 model, but that figure barely captured the real expense: hours of iteration, endless regeneration, the creative equivalent of throwing darts blindfolded. Generate, discard, regenerate. Industry insiders have a phrase for this workflow—"prompt and pray"—and it's become the defining frustration of the generative video boom.
A small startup in Gwangju, South Korea thinks it has the answer. Or at least, an answer.
HOMI AI just secured a $1 million grant from Korea's Tech Incubator Program in August, arriving at a peculiar moment for the generative video market. The sector is racing toward a projected $20.7 billion valuation by 2034, up from roughly $2 billion last year. But the conversation—at least among the professionals actually trying to use these tools—has shifted. It's no longer just about how realistic a clip looks. It's about whether you can fix it when something goes wrong.
When Generation Isn't Enough
The AI video arms race has produced genuinely impressive results. OpenAI opened Sora to wider access last December. Google iterated Veo to version 3.1 with native audio. Runway raised $308 million at a $3 billion valuation this past April; Luma AI closed a $900 million round led by Saudi-backed Humain, reaching an estimated $4 billion valuation.
Speed and fidelity have improved dramatically. What hasn't improved: the ability to make surgical edits without starting over.
According to Adobe's MAX 2025 survey, 86% of creators now use generative AI tools, with 55% specifically for editing and enhancement. That second number matters more than it might seem. It suggests a production reality that benchmark demos don't capture. Professionals don't just need one-shot generation. They need to adjust a character's expression in frame 47 without regenerating the entire 10-second sequence. They need to lock visual consistency across shots—the same face, the same lighting, the same impossible-to-define aesthetic coherence. They need motion trajectory controls with the precision of traditional VFX work.
HOMI AI, a team of fewer than 10 people, is betting that precision control is the gap in the market. CEO Taeyoon (Ted) Kim frames it bluntly: moving beyond "an image toy" to something closer to a professional creative platform. The company previously raised seed funding from Hustle Fund on a SAFE with a $7 million post-money cap, and claims acceptance of a Transformer-Diffusion framework based on Bridge Diffusion theory at ECCV 2024. Whether that academic pedigree translates to commercial traction remains to be seen.
The Workflow Problem
This shift toward "after-generation" editing isn't really about technical novelty. It's about workflow compatibility—a decidedly unglamorous concept that happens to determine whether anyone actually uses your product.
Adobe's Firefly Video Model now includes Prompt-to-Edit functionality. Users can adjust camera motion, add or remove objects, extend clips—all directly within Premiere Pro's timeline, without leaving their NLE. Runway's References system lets creators maintain character and scene consistency across multiple shots. Its Director Mode offers fine-grained camera controls that feel less like AI generation and more like traditional direction.
These aren't experimental features anymore. They're becoming table stakes.
The academic community has followed suit, perhaps more enthusiastically than usual. Papers like Video-P2P, AnyV2V, EditCtrl, and MotionV2V explore cross-attention control, first-frame injection, trajectory-based editing, and what researchers call "disentangled local-global modifications." At ECCV 2024, works on DragVideo—interactive drag-style video editing—and various bridge-diffusion approaches signaled a field-wide pivot toward granular, non-destructive manipulation.
Then there's the regulatory backdrop, which is tightening faster than the technology is improving. The EU AI Act's transparency obligations take effect in August 2026, requiring marking and labeling of synthetic content. YouTube and TikTok have rolled out C2PA-based disclosure labels. California, Washington, and New Jersey have enacted state-level deepfake and election-content laws. Provenance and auditability—features that post-generation editing tools can more easily support—are no longer optional add-ons. They're requirements.
The Quality-Cost Squeeze
The gap between hype and production value is narrowing. Sometimes painfully.
Svedka's "primarily AI-generated" Super Bowl LX spot in 2026 drew attention for all the wrong reasons—months of prep work and custom training for mascot motion still left visible quality compromises. Samsung's AI-heavy social video ads faced scrutiny over artifacts and disclosure subtleties. Mondelez announced plans to cut production costs roughly in half using AI-generated TV ads, which sounds impressive until you consider that the quality bar for broadcast work remains unforgiving.
Meanwhile, companies solving for control—not just generation—are drawing capital and enterprise traction.
Adobe's Generative Extend feature, which lets editors lengthen clips inside the timeline, has been described as "transformative" by production studios like Versus Creative. Pika's suite of post-generation features—add, swap, twist frames and scenes—positions the platform as a surgical editing tool rather than a pure generator.
HOMI AI's approach, centered on Bridge Diffusion and what Kim calls "professional-grade precision," fits cleanly into this emerging category. The company's funding strategy is worth noting: stacking Korea's non-dilutive government support (TIPS grants range from KRW 1.0 billion to 1.2 billion per startup) with seed venture capital allows for sustained R&D on control mechanisms before heavy commercialization. With U.S. expansion plans and headquarters listed in San Francisco alongside its Gwangju base, the company is making a deliberate bet that controllability—not raw generation speed—will be the differentiator as the market matures.
The competitive landscape is crowded, though not always in obvious ways. Open-weight challengers like Genmo's Mochi-1 (Apache 2.0 license) and Alibaba's Wan 2.x are chasing accessibility and cost efficiency. Kuaishou's Kling model claims up to 2-minute 1080p clips at 30fps with strong physical realism. But benchmark leaderboards—VBench, Artificial Analysis, Video Arena ELO—vary so widely in methodology that they often reflect generation quality over edit precision. Which metric matters more depends entirely on who's asking.
What Comes Next
The trajectory seems clear, at least from where the market sits today. Post-generation editing will become standard infrastructure, not a premium feature. "Prompt-to-Edit" capabilities, keyframe control, and object-level operations will be expected across leading platforms. NLE integrations will expand beyond Adobe into Resolve, Final Cut, and the specialized agency workflows that don't make headlines but move enormous budgets.
Control quality—not just clip quality—will differentiate models. Identity locking, motion trajectory control, and region-specific edits are emerging as central buying criteria for professionals. Research and product roadmaps align: every major platform is investing in controllability as the second act of the generative video story.
For investors tracking the space, the signal is less about raw compute scale and more about workflow fit. Sequoia's AI Ascent 2025 report noted improving app retention and verticalization as key trends; controllability in creative tools maps directly to sustained usage, which is a polite way of saying people stop churning. Gartner's Hype Cycle 2025 placed generative AI squarely in the disillusionment phase—the companies that survive will solve production pain points, not just generate impressive demos that win Twitter likes.
HOMI AI's $1 million grant and academic credentials position it as a bet on this thesis. Whether a small team can compete against Adobe's distribution reach, Runway's capital base, and Google's infrastructure remains an open question. The history of creative software suggests scale matters less than solving a real workflow problem, but the history of AI startups suggests the opposite.
What's not in question is the underlying shift. Precision, auditability, and surgical control are no longer nice-to-haves. They're the price of entry for professional adoption.
The "prompt and pray" era is ending—not because the technology suddenly became perfect, but because professionals need to ship work, and shipping requires fixing what's wrong without regenerating what's right. That's a more mundane milestone than the ones AI companies tend to celebrate. It also happens to be the one that matters.
