Training Open-Source Image Models on China-Developed Accelerators: Arbitrary Resolution and Chinese Text Rendering

    Export limits on high-end accelerators and a newly released open-source image model point to the same shift: generative training is no longer tied to a single source of premium compute. This article explains how an open-source image model trained end to end on China-developed accelerators combines an autoregressive backbone with a diffusion decoder, renders Chinese text through explicit glyph encoding, and generates native arbitrary aspect ratios—plus the systems work and deployment choices behind that result.

    Background: China-developed accelerators reach training scale

    Recent changes in export policy for high-end accelerators coincided with the release of an open-source image generation model. Taken together, the signal is clear: generative AI training no longer has to depend on a single source of premium compute. China’s accelerator ecosystem is moving from inference toward training, and it is starting to produce open-source results.

    Until recently, China-developed AI chips were used mainly to speed up inference. Training—especially large-scale pretraining—was still widely assumed to require top-tier accelerators. That assumption is easy to understand. Pretraining can involve hundreds of millions to hundreds of billions of parameters and thousands of accelerators. It places extreme demands on matrix operators, high-speed interconnects, distributed scheduling, checkpoint resume, and memory management. Instability in any one of those layers sharply cuts training efficiency. This model, by contrast, runs the full pipeline on a China-developed accelerator cluster: data preprocessing, large-scale pretraining, supervised fine-tuning, and reinforcement learning. That means the stack now supports production training, not only inference demos.

    For developers and enterprise teams, this is a procurement signal as well as a technical one. A complete model-production pipeline can be built on China-developed chips, which reduces supply-chain concentration and supports data residency and compliance requirements.

    Architecture: how the autoregressive backbone and diffusion decoder split the work

    Image generation still involves a trade-off. Diffusion models are strong at texture, lighting, and fine detail, but weaker at semantics, long-prompt following, and text layout. Autoregressive language models are strong at meaning, logic, and content planning, but generating images pixel by pixel is inefficient and poor at detail. Combining the two is now a major direction in multimodal generation.

    This open-source model uses a hybrid design: an autoregressive backbone plus a diffusion decoder. The backbone handles semantics and global composition. Like a director, it reads the subject, scene, style, and the spatial relationship of text in the prompt, then plans the overall layout. The diffusion decoder turns that layout into a high-quality image. Like an illustrator, it paints clear texture, accurate lighting, and readable text block by block. Together they form a “understand first, then generate” pipeline.

    A dedicated glyph-encoding module is the key to text rendering. Conventional diffusion models often treat text as ordinary visual texture. Their internal representation has no notion of character structure, so Chinese characters frequently look plausible while the strokes are wrong. Glyph encoding represents each character’s strokes, radicals, and structure explicitly, so the model knows how many strokes a character should have and whether it is left–right or top–bottom. That sharply reduces wrong and missing characters. The approach matches a broader industry direction: inject a language model’s reasoning into the image generator so the model “thinks it through” before it draws.

    Direct Chinese posters: fixing models that forget the characters

    Chinese poster generation is one of the hardest text-to-image tasks. Chinese has a large character set and complex structure, type styles vary widely, and commercial posters usually stack a headline, subhead, body copy, and decorative type. Older models either draw unreadable symbols or produce severe artifacts where text overlaps graphics. Even models that render English text reasonably well often fail on Chinese.

    The notable point of this release is that text rendering is not an OCR post-process bolted on afterward. Character structure is modeled inside the network. On public evaluations of complex visual text and long-text rendering, it reported the best open-source results. On that basis, it can be treated as commercially usable across the three requirements of a Chinese poster: accurate text, sensible layout, and visual quality.

    Complex visual text covers artistic type, outlined type, gradients, 3D lettering, and text interwoven with graphics. Long-text rendering means full sentences or paragraphs on a poster, not a few isolated characters. Together, those capabilities let a user supply ad copy and a style description and receive a complete poster draft. That cuts design cost for social covers, e-commerce hero images, offline print materials, and explanatory illustrations.

    Any aspect ratio: native generation instead of a second crop

    Traditional text-to-image models are usually trained on one or a few aspect ratios, such as 4:3, 16:9, and 1:1, and only switch among those sizes. Real jobs are less tidy. A poster may be a vertical 1240×1600, a feed ad 3:4, or a long image 1:3. If the model cannot emit those ratios, developers generate a standard size and crop it, which drops subjects and cuts off text.

    A practical feature of this model is an improved tokenizer strategy that natively supports any width and height from 1024×1024 through 2048×2048. A user can request a nonstandard size such as 1240×1600 without first drawing a square and then stretching it. Composition is planned for the target ratio during generation, not repaired afterward.

    In practice, that capability usually rests on three changes. Positional encoding and attention no longer assume a fixed grid. Training data samples many aspect ratios so the model learns composition rules for each. The decoder adjusts generation to the target size. For developers, the generation API can take width and height directly, which removes a large amount of image post-processing.

    Training optimization: systems engineering on China-developed accelerators

    Running the full pipeline on local accelerators is not a matter of moving model code from one machine to another. Large-scale pretraining has to solve operator support, communication cost, data movement, and fault recovery. The training suite covers data preprocessing, pretraining, supervised fine-tuning, and reinforcement learning, with targeted optimizations at several layers.

    • Dynamic-graph multi-stage pipeline dispatch. Micro-batch computation is split into fine-grained pipeline stages so each device has work in every time slice, reducing idle time spent waiting on a layer. That removes the bottleneck of traditional layer-by-layer dispatch.
    • Multi-stream parallelism. In large distributed runs, gradient sync and model parallelism create heavy communication. Placing communication on a separate stream lets it overlap with compute, hiding network latency and shortening step time.
    • Fused operators. Operators are adapted to the memory-access behavior of the target chips. AdamW, exponential moving average, gradient clipping, and RMS normalization are fused into fewer kernels, cutting launch overhead and memory traffic.
    • Preprocessing moved onto the training cluster. Cleaning, resizing, and augmentation run inside the training cluster instead of shuttling huge sample sets between a CPU cluster and the compute cluster, which lowers the I/O bottleneck.

    These changes are not a new model architecture. They are existing algorithms tuned to run efficiently on China-developed accelerators. They show systems engineering: end-to-end training only lands when data flow, compute flow, and communication flow stay in step.

    Evaluation and open source: why public weights matter over time

    At release, the model reported the best open-source scores on public benchmarks for complex visual text and long-text rendering. The result is useful because it is not a lead on a single metric. It covers text accuracy, layout quality, and image quality at the same time. Long-text rendering, in particular, decides whether a Chinese poster is usable.

    Open source also compounds the ecosystem. With weights and inference code public, research teams can reproduce results, inspect model details, and try different training strategies. Enterprises can deploy the model on their own compute, fine-tune it on private data, or embed it in an existing workflow. Compared with a closed API, the open release gives developers more control and makes it easier to optimize for special cases.

    An open model is not a universal out-of-the-box product. Production use still needs some fine-tuning on business data: industry ad posters, rare characters and phrases, and layout preferences. That matters especially for Chinese.

    From an API call to production

    Most application teams have two fast paths. One is a hosted cloud API billed per image: lower cost and faster to ship. The other is downloading the weights and deploying them, which fits data-sensitive workloads, strict latency needs, or deep customization. The two can run in sequence: validate the product with the API, then move to a private deployment once the result holds.

    The sketch below shows how to request an arbitrary size at generation time:

    1. # Illustrative pseudocode: generate a 1240x1600 Chinese poster
    2. response = generate_image(
    3. prompt="Tech summit key visual. Chinese title: Intelligent Computing Ahead; subtitle: A new era of multimodal generation",
    4. width=1240,
    5. height=1600,
    6. layout="poster",
    7. text_style="business"
    8. )

    Production also needs supporting services. Outputs usually go to object storage. Async jobs need a message queue and a scheduler. The inference service needs a container platform that can scale. Monitoring, alerts, and logs track failure rate, latency, and content quality. At high concurrency, teams can preload weights, batch prompts, and schedule work across an accelerator pool.

    Outlook

    A multimodal model trained end to end on China-developed accelerators is only a start. More large and multimodal models are likely to be pretrained on these clusters, and training frameworks will adapt more cleanly to the chips. The hard problems remain: long-running stability at cluster scale, automatic failure recovery, mixed scheduling across accelerator models, and access to high-quality Chinese image–text data. Those factors decide whether the ecosystem moves from “usable” to “good to use.”

    For developers and enterprises, this is a practical window to build a China-developed accelerator stack. Validating training and inference early, and accumulating experience in preprocessing, fine-tuning, and serving, leaves more room in later capacity planning. Chinese posters and arbitrary-resolution generation are the visible outputs. The more important direction is the underlying pattern: generation that plans before it draws.