Head to head: Heygen v5 Digital Twin vs MiniMax H3 Reference to Video
Heygen v5 Digital Twin vs MiniMax H3 Reference to Video
By Ryan Merket · Published
This matchup tests whether a digital-twin system or a reference-to-video model can better execute demanding prompts involving subject continuity, camera movement, and evolving environments. The results expose a decisive gap between merely stable footage and genuine prompt fulfillment.
MiniMax H3 Reference to Video wins decisively, scoring **32.3 to Heygen v5 Digital Twin’s 5.5**. It swept all 12 task verdicts, with no ties, and the aggregate result is statistically clear at **limited confidence**. The crucial distinction is that Heygen’s output was often temporally stable—but consistently the wrong video. Across the rainy-street walk, relay sprint, cathedral glide, and neon velodrome prompts, it repeatedly produced a seated woman in an indoor interview scene. Clean continuity cannot compensate for ignoring the requested subject, setting, action, and camera direction. MiniMax actually executed the assignments. It preserved the yellow-coated man and umbrella through the rain sequence, tracked the runner with bib 217, advanced coherently through the candlelit cathedral without visible cuts, and delivered the cyclist with the requested cyan-magenta lighting progression. Fine details occasionally slipped—the baton or mismatched spikes were not always visible, and some camera beats were imperfect—but these were misses within otherwise faithful videos, not wholesale prompt failures. **Final call: MiniMax H3 Reference to Video is the clear winner. The 12–0 sweep reflects a fundamental difference on these tasks here: it generated the requested scenes, while Heygen largely generated the same unrelated interview.**
Temporal consistency
A man in a yellow raincoat walking toward camera down a rainy street; his face, coat, and umbrella must stay perfectly consistent with no morphing or flicker from the first frame to the last, steady tracking shot, 16:9.
OpenAI: GPT-5.6 Sol Pro: Model B closely fulfills the rainy-street tracking-shot prompt while keeping the man's face, yellow coat, and umbrella highly consistent across all sampled frames. Model A is temporally stable and visually clean but depicts a seated woman indoors without a raincoat, umbrella, rainy street, or walking motion. (Second judge pass, order swapped — scores are the average of both: Model B closely follows the rainy-street prompt and keeps the man's face, yellow coat, umbrella, and centered tracking com | Anthropic: Claude Fable 5.1: Model A completely ignores the prompt, sho…
Relay Baton Close Track
A single continuous 16:9 shot: intimate close-up tracking beside a women's relay runner on a faded municipal track, her sharp copper undercut, tiny crescent scar above the left eyebrow, teal compression sleeves, mismatched orange and graphite spikes, and bib number 217 all staying perfectly identical as she sprints the curve holding a striped yellow baton, while the camera glides shoulder-high from her front-left to a parallel side profile and slightly behind her, catching sweat, breath, and fabric motion with no morphing or flicker; overcast late-afternoon stadium light, tense focused mood.
OpenAI: GPT-5.6 Sol Pro: Model B closely realizes the runner, track, bib 217, copper undercut, teal sleeve, baton, and requested front-left-to-behind tracking progression with strong visual continuity, though some fine details are unclear. Model A is a temporally stable interview scene but is almost entirely unrelated to the requested relay sprint. (Second judge pass, order swapped — scores are the average of both: Model B closely matches the runner, track, copper undercut, teal sleeve, yellow baton, bib 217, overcast m | Anthropic: Claude Fable 5.1: Model A shows a woman on a couch talking t…
Single continuous shot
One unbroken take gliding slowly through a candlelit cathedral from the entrance toward the altar, no cuts, jumps, or transitions, dust and warm light in the air, 16:9.
OpenAI: GPT-5.6 Sol Pro: Model A depicts a static interview on a sofa and does not follow the requested cathedral scene or forward glide. Model B strongly matches the prompt with a coherent, apparently unbroken advance from the cathedral entrance toward the altar, warm candlelight, atmospheric haze, and consistent Gothic architecture. (Second judge pass, order swapped — scores are the average of both: Model B closely fulfills the requested continuous forward glide through a warmly candlelit cathedral, with coherent spat | Anthropic: Claude Fable 5.1: Model A shows a woman talking on a couch,…
Velodrome Neon Wake
A single continuous 16:9 shot: a lone keirin cyclist in a matte plum skinsuit and mirrored visor pedals slowly along the inside rail of a nearly empty velodrome tunnel entrance, and the camera dollies backward in a low smooth close tracking shot as he rolls toward it and then slightly past, while the lighting changes visibly and believably over the clip—dim gray ambient light first, then bands of cyan and magenta neon signage above the tunnel sputter on and stabilize, followed by a sweep of warm motorcycle headlight glow washing across the concrete wall and his bike frame—creating a moody, anticipatory atmosphere with natural reflections and shadows evolving continuously.
OpenAI: GPT-5.6 Sol Pro: Model A is a temporally stable interview shot but bears virtually no resemblance to the requested cycling scene. Model B closely captures the lone plum-suited cyclist, smooth tracking composition, staged cyan-magenta illumination, and warm headlight sweep with convincing evolving reflections, though the tunnel entrance and inside-rail positioning are not exact. (Second judge pass, order swapped — scores are the average of both: Model B closely realizes the cyclist, velodrome, low tracking compos | Anthropic: Claude Fable 5.1: Model A shows a woman talking on a couch w…
Matchup powered by OpenRouter.