AI Image-to-Video Arena Leaderboard 2026

Full Leaderboard — 31 Models Ranked

Every ranking below comes from blind head-to-head comparisons run by real users on the Arena platform. No curated cherry-picks, no marketing demos. I've linked each model to its official documentation so you can test them directly.

The xAI Disruption

Nobody saw this coming. When I last updated this leaderboard three weeks ago, Google held both #1 and #2 without contest. There was no public whisper of xAI entering the image-to-video space. Then Grok Imagine Video appeared — not one variant, but two — and the 720p model went straight to the top of blind comparisons.

I've been running Grok against my standard test suite, and what jumps out immediately is temporal coherence. Feed it a portrait and the subject doesn't morph mid-animation. Hair physics stays consistent frame to frame. Eye direction tracks naturally through head turns. I tested one of my hardest inputs — a medium shot of someone turning their head while wind catches their scarf — and Grok held every detail through the entire clip. Most models lose the scarf pattern or distort the face during the turn. Grok handled it with a stability I've only seen from Veo's best renders.

The strategic play here tells you a lot about xAI's approach. They shipped two resolution tiers simultaneously: 720p at #1 and 480p at #4. The 480p variant has already accumulated substantial Arena comparisons and holds its own near the very top. This means xAI's motion architecture is fundamentally strong — the quality shows up before resolution scaling even enters the picture. If they push to native 1080p while maintaining this level of temporal fidelity, Google's audio integration becomes the only remaining differentiator keeping Veo in the conversation for the crown.

Google: Dethroned But Not Defeated

Losing the #1 spot doesn't mean Google lost the war. They still command seven of 31 positions — more than any other organization. Veo 3.1 Audio at #2 and Veo 3.1 Fast Audio at #3 remain formidable. The Veo 3 Audio variants hold #7 and #8. The non-audio Veo 3 engines sit at #13 and #15. And the aging Veo 2 clings on at #27.

Google's enduring advantage is a capability no competitor has replicated: synchronized audio generation. When I animate a café scene with Veo 3.1, I hear espresso machines hissing, cups clinking, ambient conversation — all timed precisely to the visual motion. A beach photograph gets crashing waves matched to the foam cycle. A forest path gets birdsong that shifts with the virtual camera's position. This isn't post-production audio layered on top; it's co-generated in the same forward pass as the video. In my experience, matching audio elevates perceived quality dramatically — your brain trusts motion more when it hears it.

But Veo 2 sitting at #27 tells a sobering story about deprecation speed. Twelve months ago, Veo 2 was the gold standard for I2V. Now it's outranked by twenty-six models, including several from companies that didn't have video products a year ago. Each generation in this space ages in months, not years, and Google's own newer models have made Veo 2 feel like legacy infrastructure. This rapid internal cannibalization is both Google's greatest strength and its most expensive commitment — they have to keep shipping just to stay ahead of themselves.

The Eastern Powerhouse

If you only track the top three, you're missing the structural story. Chinese AI companies collectively hold seventeen of 31 positions on this board — more than half the entire leaderboard. This isn't a niche presence. It's ecosystem-level dominance of the mid-to-upper tier, and it has direct implications for anyone building a production pipeline around image-to-video generation.

Shengshu: The Generational Leap

Vidu Q3 Pro at #5 is the model I'd tell you to pay closest attention to. Shengshu's Q2 generation — Q2 Turbo and Q2 Pro — sits at #16 and #20. Respectable, but unremarkable. The jump to Q3 is not incremental; it's architectural. In my testing, Q3 Pro handles multi-subject scenes with a precision its predecessors couldn't match. Two people walking in opposite directions? The Q2 models would start merging their outlines by frame 30. Q3 Pro keeps them distinct through the entire sequence. For portrait animation, it preserves skin textures and micro-expressions in a way that feels organic rather than synthetic. If Shengshu maintains this rate of generational improvement, a Q4 model could challenge the top three by late 2026.

Bytedance: The Camera Specialist

Seedance v1.5 Pro at #9 has become my go-to for complex camera choreography — dolly shots, orbital pans, crane-to-handheld transitions. When the animation demands intentional camera movement rather than a static frame that drifts, Seedance delivers. Seedance v1 Pro at #11 remains a reliable workhorse for standard animation tasks, and v1 Lite at #25 is the choice when speed matters more than peak quality. Bytedance's three-tier strategy gives you a complete pipeline: Lite for experimentation, v1 Pro for solid output, v1.5 Pro for the hero shot.

KlingAI: Four Tiers, One Ecosystem

Kling 2.6 Pro (#10), Kling 2.5 Turbo 1080p (#12), v2.1 Master (#17), v2.1 Standard (#19) — four models spanning different price and capability tiers. Kling 2.6 Pro is the standout for character animation: fluid body motion with face consistency that I haven't seen matched outside the top four. Kling 2.5 Turbo 1080p is notable for native high resolution in a fast rendering tier — when your delivery format demands pixel count and you can't afford an upscale step, this model saves time and money.

MiniMax, Alibaba, Tencent, and Luma AI

MiniMax's Hailuo family occupies four spots (#14, #18, #21, #23) spanning pro through fast tiers — the iteration machine I rely on for rapid drafting before committing an expensive render elsewhere. Alibaba's Wan 2.5 I2V at #6 remains the best option when artistic style preservation is non-negotiable: feed it a watercolor painting and it animates it as watercolor, not as a photorealistic reinterpretation. Tencent's Hunyuan Video 1.5 at #24 rounds out the Chinese roster with quiet, steady improvement each cycle.

Luma AI's Ray 3 at #22 deserves special mention for 3D-aware animation. Feed it a product shot or architectural render and it infers depth, generating camera motion that respects three-dimensional structure — parallax on foreground objects, correct occlusion on backgrounds. For e-commerce product videos and real estate visualization, Ray 3 is a specialist worth knowing. Their older Ray 2 at #29 shows how far the generational gap has widened even within a single company.

The Open-Source Signal

LTX-2-19b from Lightricks at #28 is the most significant entry on this list for a specific audience: teams that cannot send proprietary images to external APIs. Available on HuggingFace with open weights, this 19-billion parameter model runs on-premise. The quality gap between LTX-2 and the top 10 is real — you'll notice it in fine detail and temporal stability. But for workflows where data privacy is non-negotiable — medical imagery, unreleased product designs, classified architectural plans — LTX-2 is currently the strongest open-weight option for image-to-video generation.

The broader trajectory matters here. Wan v2.2 at #26 is also openly available. As more capable models release their weights, the floor for what's achievable without a cloud API keeps rising. I estimate open-source image-to-video is roughly where open-source language models were in mid-2024 — about twelve months behind the frontier, but closing fast. By late 2026, I expect open-weight I2V models to rival mid-tier commercial offerings, fundamentally changing the build-versus-buy calculus for enterprise teams.

My Recommendations by Use Case

What Comes Next

Having tracked this space month over month, here's where I see the landscape heading through the rest of 2026.

Audio co-generation goes mainstream. Google pioneered it with Veo 3, and the perceptual quality gap it creates is too large for competitors to ignore. I expect at least two other providers — likely xAI and Bytedance — to ship integrated audio by Q4. Once that happens, silent animation will feel like an artifact from an earlier era, the way static thumbnails feel now compared to animated previews.

Resolution escalation accelerates. Most top models currently max out at 720p. Kling 2.5 Turbo already pushes native 1080p. By end of year, 1080p will be standard for pro tiers and we'll see the first 4K previews from at least one lab. The compute cost will be punishing, but demand from broadcast and advertising workflows is undeniable.

xAI scales up aggressively. Two models in three weeks — with the 720p variant claiming #1 on arrival — signals serious investment. I'd expect higher resolution variants and possibly audio integration from Grok before summer. If they maintain this motion quality at 1080p, they become the clear frontrunner.

Runway needs a Gen5 moment. Runway Gen4 Turbo at #30 is a difficult position for the company that essentially created the commercial AI video category. Their creative tooling and user experience remain best-in-class, but the underlying model needs a generational leap. If Gen5 doesn't ship by mid-2026 with top-10 quality, Runway risks becoming the company that defined the market and then watched everyone else win it.

Open-source narrows the gap. LTX-2 proved open weights can produce viable image-to-video results today. The next wave — possibly a Wan 3 or LTX-3 — will push into territory that rivals mid-tier commercial models. For enterprise teams building proprietary pipelines without external API dependencies, this is the trend that matters most.

The missing players. Meta, Apple, and Amazon remain conspicuously absent from this leaderboard. Meta's video research publications suggest capability that could compete at the top tier, but they haven't shipped a public-facing I2V product. The moment Meta enters — especially if they release an open-weight model, as they did with Llama for language — the entire competitive landscape reshuffles overnight.