Gemini Omni 1.2 Flash vs InternLumina-U2: Rented Seconds Against Downloaded Weights

Put Gemini Omni 1.2 Flash beside InternLumina-U2 and the comparison that jumps out — image quality, generation speed, benchmark scores — is the least interesting one available, because the two models are not competing for the same budget line. Gemini Omni 1.2 Flash is a metered API that charges about $0.07 per second of 720p video, went generally available on 5 November 2026, and returns a file you own the rights to use but never the weights. InternLumina-U2 is a 16-billion-parameter sparse diffusion language model from Shanghai AI Laboratory, Apache-2.0, seven shards per hardware tree, sitting on Hugging Face where anyone with the disk can download it. One is a service. The other is an asset. The useful question is not which generates better pictures; it is which of those two things your project can actually absorb.

That framing matters because the obvious query — "which one is better at image generation" — has a misleading answer. InternLumina-U2 does generate images, and on its own vendor-reported benchmarks it does so competitively. But it also understands video and 3D in the same backbone, which is a capability Gemini Omni does not have at all, while Gemini Omni generates video with synchronized audio, which InternLumina-U2 cannot do. Judged on the overlap, they are both image-and-video-adjacent systems. Judged on the union, they barely touch.

The overlap, precisely

InternLumina-U2 handles text QA, text-to-image generation, image understanding, image editing, video understanding and 3D understanding inside one model. The "understanding" qualifier is doing real work in that list: it reads video and reads 3D, it does not produce either. What it produces is images and text.

Gemini Omni 1.2 Flash produces video. It takes text, images and video in, and returns video with native audio, three to fifteen seconds at 24 frames per second, 360p through 4K where the top two tiers are upscaled. It does not produce images as a deliverable — every frame is a frame of a clip — and it does not do still-image editing in the sense a designer means it.

So the overlap is exactly one thing: generating an image from a prompt. InternLumina-U2 does it as one task among six. Gemini Omni 1.2 Flash does it only as the first frame of motion. If that is your entire use case, the comparison is real. If it is not, you are choosing between a still-image researcher's toolkit and a video production endpoint.

• Modality in — InternLumina-U2 takes text, image, video and 3D; Gemini Omni 1.2 Flash takes text, image and video.

• Modality out — InternLumina-U2 returns text and images; Gemini Omni 1.2 Flash returns video with synchronized audio.

• Architecture — InternLumina-U2 is a 16B-total, roughly 1B-active MoE diffusion language model on the LLaDA-2.0 backbone with Apple AToken eight-codebook discrete visual tokens; Gemini Omni 1.2 Flash's architecture is undisclosed.

• Access — Apache-2.0 weights on Hugging Face, with separate Ascend and NVIDIA trees; a closed hosted API.

• Price — no per-call price, because there is no hosted endpoint; about $0.07 per second of 720p video output, with no free tier.

• Context — 131,072 tokens in, 65,536 out on the Omni side; the InternLumina card does not publish an equivalent window.

What the benchmarks do and do not say

InternLumina-U2's numbers come from its own paper and project page (arXiv 2610.20818, submitted 14 October 2026) and nobody has reproduced them. The headline generation figure is a GenEval score of 89.3 under the authors' own PHV sampling protocol, alongside DPG-Bench 87.79, UniGenBench 86.11 and TIIF-Bench 88.57 on short prompts and 87.08 on long ones. The paper claims leading performance among unified models in visual understanding, image generation and image editing, and says the model beats FLUX.1-dev and edges Qwen-Image on GenEval, UniGenBench and TIIF while activating roughly fourteen times fewer parameters.

There is a discrepancy worth knowing before you quote any of it. The project page's results table lists InternLumina-U2 at GenEval 0.81, with LLaDA2.0-Uni at 0.89 and InternVL-U at 0.85 — which puts the newer model behind its own predecessor on that row. The 89.3 figure corresponds to the optimized sampling protocol described in the paper, not the default evaluation. Two numbers, two protocols, same model. The paper is the more favourable reading and the project page is the more conservative one, and neither has been independently checked.

Gemini Omni 1.2 Flash has no published benchmark scores from Google at all. The 5 November launch post is entirely qualitative — longer clips, better reference adherence, more natural dialogue, fewer jump cuts — and explicitly says work continues on "fast sports action, rapid script changes and complex physics." That is an unusually candid list of known weaknesses, and it is the only quality statement Google has made. Third-party numbers exist in the wild but they mostly describe earlier Omni versions: Megaton's comparison page is headed "Gemini Omni Flash" while its links point to the preview slug, lists audio output as "No" (wrong for 1.2, which generates audio by default), and reports zero rated videos and zero measurements behind its scores. One hands-on reviewer, Segmind's Rohit Rao, did test 1.2 directly across twenty generations at a cost of $16.17, and his conclusion is the most useful independent data point available: calls ran 28 to 136 seconds, most in the 40 to 90 second band, the model was the most instruction-exact he had tested, and about one instruction in four was skipped. He also found it is not a cinematic photoreal-people model.

So the honest scoreboard has InternLumina-U2 with detailed, unreproduced, internally inconsistent vendor numbers and Gemini Omni 1.2 Flash with almost no numbers at all and one small independent test. Neither is a basis for a confident quality verdict.

The cost comparison that people get wrong

InternLumina-U2 has no price, and that is not the same as being free.

The weights are 67.6 GB across two trees of seven shards, roughly 33.8 GB each, so downloading one hardware variant is a 34 GB pull before you have a working environment. The repository requires FlashAttention for the AToken path and, for the 3D pipeline, Blender 4.5. Training ran on 1,024 Ascend 910B NPUs across 1.05 trillion tokens, which tells you the model was built for a cluster even though inference is far lighter — 1B active parameters is a modest serving load, and the paper's PHV sampling strategy claims a 4.45× end-to-end latency reduction at native 1024×1024.

What you actually spend is a GPU, an engineer-week to get the environment right, and whatever your storage and egress cost. What you get is unlimited generation at zero marginal cost, a model you can fine-tune, and no vendor that can deprecate the endpoint under you.

Omni's cost is legible to the cent. A fifteen-second 720p clip is about $1.05 in output tokens. A 360p draft is about $0.35. A 4K upscale of that fifteen seconds is about $3.14. You pay per output second, you pay again on every retry, and conversational editing means retries are the normal workflow rather than the exception. A ten-shot ad with three edits per shot is not a fifteen-dollar job; it is closer to sixty.

Those two cost structures are not convertible. Comparing "free after capital" to "seven cents a second" produces a number that depends entirely on utilisation, and at low volume the self-hosted model is dramatically more expensive per image. At high volume it is dramatically cheaper. The crossover is arithmetic you have to run with your own GPU price, not a fact anyone can hand you.

What you cannot get from either

If you need a talking-head clip with lip-synced dialogue, InternLumina-U2 cannot help and there is no workaround short of a separate video generator. If you need a 3D asset understanding pipeline — reading a mesh or a scene description alongside images — Gemini Omni 1.2 Flash cannot help and there is no workaround short of a separate vision model.

There is also a paperwork difference that matters more than it sounds. InternLumina-U2 is Apache-2.0, which means commercial use, modification and redistribution with attribution, and the licence text is on the repository. Gemini Omni 1.2 Flash is a hosted service whose terms govern what you may do with output, every clip carries a SynthID watermark whether you want one or not, and editing or extending uploaded video is unavailable in the EEA, Switzerland and the UK. For a team in London building a product on top of conversational video editing, that regional restriction is disqualifying in a way no benchmark comparison would surface.

Where OrcaRouter sits, honestly

Neither model is on our catalogue. google/gemini-omni-1.2-flash returns 404 on OrcaRouter's model pages, and InternLumina-U2 has no hosted endpoint anywhere — its card says as much, and only a community Space uses it. So no call to our API reaches either one, and this piece would rather state that than imply a host we do not have.

What the gateway does cover is the layer both workloads sit inside. A self-hosted InternLumina-U2 pipeline still needs a text model to write the prompts, score the outputs and route failures, and that is a call we serve across more than 200 models on one key with automatic failover. On the Omni side, because OrcaRouter passes the provider's list price through at 0% markup, a Google price cut lands on our side the same day rather than at the next billing cycle — and the Omni line has already moved its video rate once this year, from $17.50 to $12.00 per million output tokens. If you are comparing a self-hosted build against a rented one, the rented side moving under you is exactly the variable a pass-through bill makes visible.

How to decide

Build on InternLumina-U2 if you need images, you need 3D or video understanding in the same model, you have GPU capacity you are already paying for, and the licence terms matter to your product. Accept that you are running a research artifact: the benchmarks are the authors' own, the GenEval table disagrees with itself, and the tooling expects a specific stack.

Buy Gemini Omni 1.2 Flash if you need motion with sound, you want a working endpoint today, and your volume is low enough that per-second billing is cheaper than an idle GPU. Accept that you have no quality benchmarks, no text-output mode, no negative prompts, a regional restriction on the editing features, and a watermark on every frame.

The mistake to avoid is treating this as a bake-off. It is a procurement decision between a download and a meter, and the model quality question — the one both vendors would like you to argue about — is the one with the least evidence on either side.