In September, Chinese open-source image generation models fired off a rapid series of moves: SenseTime, Ant, and Shusheng couldn't sit still

This past September, Chinese AI image generation saw a change that wasn't flashy but was significant: open source. Within just one week, SenseTime, Ant, and the Shusheng (Intern) team at Shanghai AI Laboratory released open-source image generation models, putting capabilities that had previously been held tightly by closed-source giants onto the table. The impact of this may be far greater than the launch of any single new model.

What each of the three players brings to the table

SenseTime SenseNova U1.5 Lite was open-sourced on September 8. It is an 8B image generation model. Its selling point is direct 4K output, and the official line is that it is "on par with closed-source large models." The 8B size is crucial: it means the GPU in an average developer's machine can handle it, without immediately needing 12GB or 24GB of VRAM.

Ant's 6B unified image model was open-sourced on September 6. It takes a more "hassle-free" approach: text generation and instruction-based editing are combined in a single model. In other words, you can ask it to create an image from scratch, or give it an image and tell it to "replace the background" or "brighten the colors," without switching between two models. Even rarer, Ant has fully disclosed its training recipe. There are plenty of open-source models, but not many lay out the training details.

Shusheng InternLumina-U2 was released on September 3. It is a large vision model focused on unified image understanding and generation, and its weights will be open. Its positioning leans more toward "one model that can both understand images and create them," which gives it an edge over pure generation models in scenarios where the model must first understand a reference image and then modify it accordingly.

Why open source deserves its own discussion

Many people may think open-source models are far removed from ordinary users—after all, they are not training models themselves. But the real chain of impact is this: open-source models become the foundation for all kinds of free tools.

If an independent developer or a small team cannot afford big-company APIs but wants to add image generation to their product, what do they do? The answer is to pull an open-source model from Hugging Face and run it locally or on a cheap rented GPU. Sizes like SenseTime's 8B and Ant's 6B fall exactly into the range that "can fit into a consumer-grade machine." The more open-source models there are—and the smaller and stronger they are—the more free image generation setups ordinary people can build.

Seen from another angle, this is also a response from Chinese models to established open-source paths like Stable Diffusion and Flux. In the past, open-source image generation by default meant the Stable Diffusion ecosystem. Now Chinese players are starting to push in, bringing local advantages such as Chinese language understanding and Chinese text rendering, which is a real incremental gain for Chinese-speaking users.

After open source, what are they competing on?

Open source is not the finish line, but the starting point. Once a model is open-sourced, whether it can actually be used depends on three things.

First, the ecosystem. Weights alone are not enough; a model needs supporting inference tools, LoRA training scripts, and WebUI support. The reason Stable Diffusion remains many people's default choice today is not any single version, but the entire toolchain around it.

Second, Chinese-language capability. The biggest differentiator for Chinese models lies in Chinese prompt understanding and Chinese text rendering. This is an area where overseas open-source models have long fallen short, and it is also where Chinese open-source models should hold their ground hardest.

Third, the path to commercialization. How open-source models make money has always been a hard problem. For players like SenseTime, Ant, and Shusheng, open-sourcing is most likely not charity, but a way to nurture an ecosystem and then monetize through enterprise services, cloud APIs, and customization. For users, this means more and more free options, but whether "free" can be equated with "continuous maintenance" remains to be seen over time.

One more piece of background worth adding: this wave of open-sourcing happens to land at a time when the overseas open-source ecosystem is in a bit of a lull. Stable Diffusion has not had a major version update in a long time, and Flux has yet to release complete open-source weights. By open-sourcing 6B and 8B models that "can fit into a home machine" at this moment, Chinese models are effectively laying the foundation for Chinese open-source image generation while the other side is catching its breath.

What it means for everyday creators

For actual people, the change brought by these open-source models is that the cost of image generation continues to fall.

Previously, if you wanted to reliably use AI image generation for serious work, you either subscribed to a tool or paid per-image API fees. As open-source models have multiplied, a batch of free, locally run options has emerged. Although these options generally have a learning curve—you need to set up the environment and tune parameters—the community is gradually smoothing down the barrier.

My assessment is that, in the short term, everyday creators will still find web-based tools the easiest. But if you are building a small product or app and need to embed image generation, the density and size of this wave of open-source models are worth a fresh look. At the very least, the claim that "China has no good open-source image generation models" no longer holds from September onward.