Posted in:
Marc Berman
Google unveiled Gemini Omni at I/O 2026 on May 19, eight days before this post. Gemini Omni Flash, the first model in the family, started rolling out the same day through the Gemini app, Google Flow, and at no cost on YouTube Shorts and YouTube Create. API access for developers and enterprises ships in the coming weeks.
This is not a “Veo 3.2.” It’s a new product category — and it makes most of the AI video generator comparisons published earlier this year feel outdated overnight.
Three things separate Gemini Omni from the previous generation of video models, including Google’s own Veo 3.1.
The first is native multimodal input. Gemini Omni accepts text, images, audio, and video in the same generation call, then produces a roughly 10-second video clip with synchronized audio. Veo 3.1, by comparison, has separate text-to-video and image-to-video paths that production teams stitch together by hand. Gemini Omni doesn’t make you stitch.
The second is conversational editing that preserves state. You can ask Gemini Omni to revise its own output — change the lighting, add a character, alter the camera move — and the model carries character identity, scene composition, and physics across turns. That replaces the prompt-and-pray loop that the current generation of video models still forces on you.
The third is physics as a model property, not a render trick. Demis Hassabis described Gemini Omni publicly as a “world model” — trained against a representation of how gravity, motion, and fluid dynamics actually behave. Early reviews are flagging the physics handling as the standout improvement, with fewer of the floating-water and broken-anatomy artifacts that still embarrass earlier video models on physically complicated prompts.
Gemini Omni doesn’t replace the other top models. It joins them at the top of a shortlist that grew this week.
The Artificial Analysis Video Arena (with audio) currently ranks HappyHorse-1.0 at Elo 1213, Dreamina Seedance 2.0 720p at 1212, Kling 3.0 Omni 1080p (Pro) at 1103, Kling 3.0 1080p (Pro) at 1096, and Veo 3.1 at 1095. Gemini Omni is too new to have a stable arena ranking yet, but the early signals put it in that same bracket.
The models worth paying attention to in mid-2026 are:
Nobody who ships real video work uses just one of these. The cost of forcing the wrong model onto a shot is higher than the cost of switching.
Side note for anyone testing this week: LoraAI is running HappyHorse at 20% off through its current promo window, which is the cheap way to throw real briefs at the current Video Arena #1 without paying full per-second list.
The Artificial Analysis Image Arena is quieter. OpenAI’s GPT Image 2 (high) leads at Elo 1338, with GPT Image 1.5 (high) at 1267, Nano Banana 2 (Gemini 3.1 Flash Image Preview) at 1264, and Nano Banana Pro (Gemini 3 Pro Image) at 1219. The closed top three are within 70 Elo of each other — none is dominant on every prompt.
Black Forest Labs’ FLUX.2 [dev], released November 25, 2025, is the strongest open-weight option at Elo 1159: a 32-billion-parameter rectified flow transformer with multi-reference conditioning across up to 10 images. ByteDance’s Seedream 5.0 Lite, released late January 2026, is the unique one in the top tier — it ships with web-connected retrieval, which means it can pull current information into the generation rather than only working from training data.
Same promo window on the image side: LoraAI has GPT Image 2 generation uncapped right now, which is the cleanest way to run a serious head-to-head on your own prompts without the per-image meter eating the comparison.
Here’s what no model launch this year fixes: none of these frontier models knows what your specific character, product, or brand looks like. They draw plausible versions of similar things. For one-off prompts that’s fine. For producing on-brand assets at volume, it’s the gap that keeps content teams stuck in re-touching cycles.
That gap is closed by LoRA training. The original technique came from Edward Hu and colleagues’ 2021 paper (arXiv:2106.09685), which showed Low-Rank Adaptation can reduce trainable parameters by 10,000x compared to full fine-tuning, with no quality loss. Applied to diffusion-based image and video models, the same approach lets you train a small adapter file on a curated reference set and slot it into a compatible base model.
The mistakes that burn most teams’ credits are the ones nobody covers properly:
This is editorial-discipline work, not engineering work, which is why companies that hand it to engineering tend to get mediocre results. The LoRA training pipeline that fits inside a creative workflow matters more than the raw training step.
Stop committing to one model vendor for video. The list that matters this month — Gemini Omni Flash, Veo 3.1, Kling 3.0, Seedance 2.0, HappyHorse — does not live inside any one provider’s account.
The setup that fits this picture is an integrated platform that runs the actual top models under one credit balance, with LoRA training built into the same workflow. LoraAI is one of those. It has Gemini Omni alongside Veo 3.1, Kling V3, Wan, HappyHorse, PixVerse, and Seedance on the video side, and GPT Image 2, Nano Banana Pro, Seedream 5.0, Flux 2, and Qwen Image on the image side — all under one credit balance, with LoRA training on Flux, Kontext, Wan, and Nano Banana base models. LoRAs trained on the platform appear directly in the generation UI without an export step.
You can try LoraAI on 50 free credits at signup, no card required.
See more
©2026 Programming Insider.
Leave a Reply