CloudSignal AI

Blog

Generative vision: the new image models and what they change for CV teams

By Ashwin Rajendraprasad Computer Vision

The wave that gave us chatbots also gave us image models that generate and edit photoreal pictures from a sentence. For a computer-vision team the question is whether any of it helps you ship. In specific places the answer is yes, and it brings new failure modes you have to manage.

Two families, don’t confuse them

Start by separating two kinds of model that are easy to lump together. Diffusion models are generative: they synthesize and edit pixels. The models you already run in a CV stack are discriminative: they detect, segment, and classify, reading pixels rather than making them. The self-supervised backbones and open-vocabulary segmenters in your pipeline are on the discriminative side of that line.

Mechanically, a diffusion model learns to reverse a noising process: start from random noise and denoise, step by step, toward an image that matches a text prompt. The important framing for a CV team is that adopting one adds a tool alongside your perception stack. Generation and perception solve opposite problems.

Confusing the two is where teams get into trouble. A perception model is accountable to ground truth: it is right or wrong about something that actually exists in the frame. A generative model has no such anchor; it produces something plausible, and plausibility is not correctness. Hold that one distinction and everything downstream gets easier to reason about.

What 2025’s models actually do

The capabilities that matter are the practical ones. Today’s models do text-to-image generation, instruction-based editing (“make the background white”), inpainting (fill a masked region), and outpainting (extend beyond the frame). The newer trick is keeping an object consistent across a series of edits, the thing that turns one-shot novelty into a workflow you can build on.

Outpainting earns a specific mention for anyone juggling fixed-aspect assets: extending an image beyond its original frame lets you re-target one product shot to a wide banner, a square tile, and a vertical story without a reshoot. Instruction-based editing, meanwhile, collapses a chain of manual masking steps into a sentence. That’s powerful, and precisely the point where an unreviewed output can slip a warped edge or a hallucinated detail into your catalog.

For a product team, inpainting and consistent editing are usually more valuable than raw text-to-image. They work on assets you already own, so every run starts from a known image.

Which model is which

Name the models plainly and without hype. Google’s Gemini 2.5 Flash Image, nicknamed “Nano Banana”, is an image generation and editing model released in 2025. The higher-end Nano Banana Pro is Gemini 3 Pro Image, from Google DeepMind. To keep the two straight: Nano Banana is Gemini 2.5 Flash Image; Nano Banana Pro is Gemini 3 Pro Image. Around these sits a broader open diffusion ecosystem.

Update, July 2026: the family has grown since this was published. Google added Nano Banana 2 (Gemini 3.1 Flash Image) in February 2026 and Nano Banana 2 Lite (gemini-3.1-flash-lite-image) in June 2026.

I’m deliberately not ranking them on quality here. Benchmark claims need a source, and the useful question for a CV team is where generation belongs in the pipeline.

Where it earns its place in a CV pipeline

Three uses are real and defensible. First, synthetic and augmented training data: generating or augmenting images to expand a training set is an established technique for CV, with the standing caveat that it’s a supplement to real data, never a replacement. In practice that means reaching for it where real data is thin (rare classes, awkward lighting, conditions that are expensive or unsafe to stage) and to balance a skewed dataset before training, rather than to manufacture the bulk of what the model learns from. Second, editing for e-commerce: background removal, virtual staging, and general product visualization, where you’re transforming assets rather than inventing facts. Third, product visualization and try-on, where placing or restyling a captured object is the whole point.

In each case the generative model is doing production work (filling a gap in a dataset, cleaning up an asset) rather than making a perception decision. That distinction is what keeps the use honest.

The new failure modes

Generation introduces problems that discriminative models don’t have, and you have to design for them. Hallucinated detail: the model produces plausible texture or structure that was never there, convincing and fabricated at once. Warped text and edges: a classic diffusion tell that will sink an e-commerce image if it slips through. Provenance and watermarking: invisible watermarking such as Google’s SynthID exists to mark AI-generated images, and you should track what in your pipeline is generated versus captured. Rights and consent: for both training inputs and generated outputs.

There’s one failure mode that bites CV teams specifically. Over-training on generated data risks distribution shift and, in the extreme, model collapse, where the model drifts toward its own synthetic outputs and away from reality. Synthesize to fill gaps; don’t let synthetic data come to dominate the distribution you train on.

All of these are manageable with a gate. Track which assets are generated versus captured, keep provenance metadata attached, and put a review step in front of anything that reaches a customer or a training run. Skipping that check is where the cost shows up.

CloudSignal lens: where generation belongs

Our position is straightforward. Generative vision belongs inside the CV workflow (augmenting data, editing assets, visualizing products) with a human and an eval in the loop. Its outputs are guesses about what a picture could look like, and a system that can’t tell you which of its pixels it invented shouldn’t be making decisions on them.

Used that way, these models are a real productivity gain. Skip the eval and what you ship is a liability with good production values.

If you’re standing up a synthetic-data or editing pipeline and want it to survive an eval rather than just a demo, that’s a build conversation worth having early. It connects directly to the 3D product-visualization work on the rendering side.

Sources / further reading

Written by Ashwin Rajendraprasad for CloudSignal AI.