Model comparisons usually get written as leaderboards, which is the least useful shape for this decision. There is no single best video model for a UGC ad, because the ad is not one job. Rendering a person speaking a scripted line is a different problem from rendering a product being held, which is different again from testing whether a hook holds attention for three seconds. The models diverge sharply on which of those they are good at, and they diverge by roughly 20x on price.
So this is organized as a decision rather than a ranking: four axes that determine the choice, a real cost table, and then the specific situations where each model is the right call.
Or let the pipeline pick
siriusly.ai runs these models behind one interface - chunk planning, audio, captions and merge included - so switching model is a dropdown rather than a rewrite.
Try it freeThe Four Axes That Decide It
1. Does it generate its own speech?
This is the biggest architectural fork, and it splits the field in two.
Self-voiced models generate the speech audio and the matching lip movement in the same pass. Because both come out of one process, they cannot drift apart - there is no sync step to get wrong, ever. The cost is control: you get the voice the model produces for that character, and re-rolling for a different delivery means re-rendering the video too.
Silent models render video only. You generate speech separately with a dedicated text-to-speech system, which gives you a real voice library, voice cloning, per-language casting and precise control over pacing - then you have to get the lips to match, either by conditioning the generation on the audio or by a separate lip-sync pass. That extra step is where most of the visible failures in this format come from.
For a talking-head UGC ad, self-voiced is the safer default. For anything where the voice is a brand asset or has to be cast per market, silent plus dedicated TTS is worth the extra plumbing.
2. What is the duration ceiling?
Every model caps a single generation, usually between 4 and 10 seconds. A 30-second reel is therefore always several generations stitched together, and the ceiling determines how many joins you have to hide. Fewer, longer chunks means fewer opportunities for the character to drift. The technique for holding continuity across them is last-frame chaining, and it works on all of these models.
3. How faithful is it to a reference image?
UGC production is image-to-video, not text-to-video. You have a creator and you have a product, and both need to survive the render recognizably. Models vary a lot here, and in a specific way worth knowing: most are noticeably better at holding a face than at holding a product. Faces are heavily represented in training data and have strong structural priors; a specific bottle with specific label text does not. Any model will quietly redesign packaging text given the chance.
4. What does a second of output cost?
Because you are producing dozens of variants rather than one hero asset, per-second cost compounds directly into how many hooks you can afford to test. This is the axis that most changes behaviour once you look at the real numbers.
The Cost Table
Published or measured per-second rates for a second of finished output, as of August 2026. Provider pricing changes often and varies by routing, so treat these as ratios worth reasoning about rather than a live rate card - check your provider before budgeting.
| Model | Audio | 480p | 720p | 1080p |
|---|---|---|---|---|
| Grok Imagine 1.5 | Silent | - | $0.023 | $0.04 |
| Seedance 2.0 Mini | Silent | $0.038 | $0.082 | - |
| Veo 3.1 Lite | Native | - | $0.05 | $0.08 |
| Flux 3 Video (draft) | Silent | - | $0.06 | - |
| Seedance 2.0 Fast | Silent | $0.061 | $0.131 | - |
| Gemini Omni | Native | - | $0.066 | $0.066 |
| Seedance 2.0 | Silent | $0.074 | $0.159 | $0.408 |
| Veo 3.1 Fast | Native | - | $0.10 | $0.12 |
| Minimax H3 | Silent | - | $0.113 | - |
| Seedance 2.5 | Silent | $0.111 | $0.251 | $0.615 |
| Flux 3 Video | Silent | - | $0.17 | $0.29 |
| Veo 3.1 | Native | - | $0.40 | $0.40 |
The spread from cheapest to most expensive is about 17x at 720p. Translated into what matters, a 30-second reel costs about $0.70 on Grok Imagine and about $12.00 on Veo 3.1 - which means the same budget buys either one hero render or seventeen hook tests.
Model by Model
Gemini Omni
Generates speech and picture together in one pass, which makes it the most reliable option for a talking head: lip sync is structurally guaranteed rather than achieved. Its hard ceiling is 10 seconds per generation, and it only accepts four discrete durations (4, 6, 8 or 10 seconds) - ask for 30 and it quietly returns 10. At around $0.066 per second it is also unusually cheap for a native-audio model, and its 720p and 1080p tiers are priced identically, so there is no reason to render it at 720p.
Use it for: the default talking-head UGC reel, in volume.
Avoid it for: anything needing a specific cloned voice, or long unbroken action.
Veo 3.1
The quality benchmark, with native audio, strong prompt adherence and the most convincing motion in the group. It is also six times the price of Omni at $0.40 per second, which puts a 30-second reel at around $12. The Fast tier at $0.10 and the Lite tier at $0.05 bring it into testing range and are genuinely useful - Lite in particular is one of the cheapest native-audio options available at all, though it gives up meaningful quality to get there.
Use it for: hero creative, the variant you scale after testing, anything with complex camera movement.
Avoid it for: the first thirty hook tests.
Seedance 2.5
Silent, and very good at holding a reference image, which makes it the strongest option when the product has to stay recognizable. It offers a real resolution ladder - roughly $0.111 at 480p, $0.251 at 720p and $0.615 at 1080p - so you can test cheap and finish expensive within one model, which keeps the look consistent between the test and the scaled version. Because it is silent, it needs a separate speech pipeline.
Use it for: product-forward shots, and any pipeline built on cloned or cast voices.
Avoid it for: a lip-sync-critical shot with no audio conditioning step.
Seedance 2.0 (and Fast / Mini)
The previous generation, at roughly two thirds the price of 2.5 across the ladder, plus Fast and Mini cuts that go substantially cheaper again - Mini lands near $0.038 at 480p. The quality gap to 2.5 is real but not proportional to the price gap, which makes this family the value pick for volume work. Fast and Mini top out at 720p.
Use it for: high-volume variant generation, hook testing, anything where you will re-render the winner anyway.
Avoid it for: the final asset.
Flux 3 Video
Silent, with strong aesthetic control and a distinctive look, at $0.17 per second at 720p and $0.29 at 1080p. Its draft mode is the interesting part: a flat $0.06 per second regardless of resolution, intended for previewing composition before committing. Using draft mode to check framing and motion before a full render is a genuinely good workflow and one more models should offer.
Use it for: stylized or design-led creative where the look is the point.
Avoid it for: straight naturalistic talking heads, where cheaper models are indistinguishable.
Minimax H3
Silent, reference-to-video, around $0.113 per second at 720p, with a higher tier available. Notably strong at multi-reference composition - holding a person and a product together in one shot, which is precisely the demonstration beat most UGC ads need and most models handle worst.
Use it for: shots where a specific person must hold a specific product.
Avoid it for: long dialogue-driven sequences.
Grok Imagine 1.5
The cheapest usable option in the group at roughly $0.023 per second at 720p, which is about a third of Omni and a seventeenth of Veo. Quality is correspondingly lower and it is silent. At this price the right mental model is not "a cheap final render" but "an almost free test": you can render forty hook variants for the cost of one Veo reel, find the two that hold attention, and re-render those properly.
Use it for: mass hook testing, storyboard-level previews.
Avoid it for: anything a customer sees.
The Two-Tier Workflow
The cost spread makes a single-model pipeline hard to justify. The pattern that gets the most out of a budget is explicitly two-tiered:
- Test tier. Render only the first 5 to 8 seconds of each variant - the hook, and nothing else - at 480p on the cheapest model that produces a coherent person. Twenty hook variants at 6 seconds on a budget model runs roughly $4 to $5 in total.
- Read the retention curve, not the conversions. At this length you are measuring one thing: does the first three seconds hold. That is the only question the test tier can answer, and it is the question that decides most of the creative's fate anyway.
- Production tier. Take the two or three surviving hooks, render the full reel at 1080p on a quality model, and localize from there.
Worked example - twenty hooks down to two finished reels
| Stage | What renders | Cost |
|---|---|---|
| Test | 20 hooks x 6s at 480p, budget model | ~$4.60 |
| Production | 2 winners x 30s at 1080p, quality model | ~$24.00 |
| Total for two tested, finished reels | ~$28.60 | |
Rendering all twenty at production quality instead would have cost about $240 to learn the same thing.
Note the asymmetry: the test tier is cheap enough to be rounding error, so the constraint on how many hooks you test is how many you can write, not what they cost. That is the actual change these price points enable, and it is why creative volume stopped being a budget question.
Common Mistakes
Picking one model for the whole pipeline. You either overpay for tests or under-deliver on the final asset. There is no rate that is right for both.
Fix: two tiers, cheap for discovery and expensive for the winner. Keep the creator reference identical across both so the winner looks like what you tested.
Rendering at 1080p by default. On most models 1080p costs 2 to 2.5x the 720p rate, and the platform re-encodes your upload regardless.
Fix: 720p for everything except final assets, and check first whether your model prices the two tiers identically - Gemini Omni does, so 720p there saves nothing.
Expecting a silent model's lips to match audio you generate separately. Nothing in the render knows about your audio track unless you condition it on one.
Fix: either use a self-voiced model, or supply the audio to a model that conditions on it. Bolting on a lip-sync pass afterwards is the option that most visibly reads as synthetic.
Asking any model to render legible product packaging text. It will produce near-miss gibberish and the render is wasted.
Fix: keep the label out of focus, or composite the real product asset over the generated shot in post.
FAQs
What is the single best model for a UGC ad?
If forced to one: a self-voiced model in the Gemini Omni price band, because guaranteed lip sync removes the failure mode that most reliably ruins a talking-head ad, and the price allows enough variants to actually find a winner. But the two-tier workflow beats any single choice, and the gap is large.
Are these prices going to hold?
Almost certainly not. Per-second video generation pricing has fallen steadily and the mid-tier is competitive enough that rates move within months. The ratios have been more stable than the absolute numbers, so treat the structure of the table as the durable part.
Does resolution matter if the platform re-encodes everything?
Less than people assume for the video, more than they assume for the captions. Platforms re-encode aggressively, so a 1080p master mostly buys you headroom rather than visible detail. Where it does show is hard-edged burned-in text, which degrades before anything else - covered in the captions post.
How do I compare models fairly myself?
Fix everything except the model: the same reference image, the same prompt, the same duration, the same resolution. Then judge on three specific things rather than a general impression - identity fidelity against the reference, whether the product survives recognizably, and whether motion holds through the whole clip instead of only the first two seconds.
Which models does siriusly.ai run?
The families above, selectable per run, with chunk planning, speech, captions and merge handled the same way regardless of which one you pick - so switching model to trade cost against quality does not mean rebuilding anything around it.