Tencent previews Hy Image 3.5 with unified generation and editing

Tencent’s Hunyuan team has released a preview of Hy Image 3.5, its latest hosted image-generation model. Tencent claims a 30% human-evaluation win-rate improvement over Hy Image 3.0, alongside text-to-image and image-to-image support through one endpoint, output up to 2K, and greater consistency across related generations.

One model ID, two workflows

The preview exposes both generation modes under hy-image-v3.5-preview. Applications that previously routed new-image requests and reference-based edits to separate models can consolidate that logic, although production teams still need to verify request schemas, supported parameters, response formats, and error behavior in Tencent Cloud’s current documentation.

Tencent describes the output ceiling as 2K. Developers working with fixed layouts should confirm the accepted width, height, and aspect-ratio combinations before changing rendering pipelines or storage estimates.

Reference-heavy jobs get favorable billing

At $0.024 per generated image, 1,000 outputs cost $24 at list price. Tencent bills generated outputs and lists uploaded reference images at no charge, which gives teams predictable input costs when a request includes several product, character, style, or layout references.

The preview is also available through Miora, a design agent, and OnSolo. Tencent says both partner tools will provide free access for two weeks. Those interfaces can support visual evaluation, but API-specific testing such as latency, concurrency, parameter coverage, and failure handling still requires Tencent Cloud access.

The 3.0 foundation remains visible

Tencent presents Hy Image 3.5 as an iteration on Hunyuan Image 3.0. The earlier model has 80 billion total parameters and uses a Mixture of Experts architecture with 64 experts and 13 billion active parameters. In plain terms, each request activates a subset of the network, reducing the computation required compared with using every parameter for every inference.

For reference-based editing, Tencent previously offered a dedicated 3.0-Instruct variant. That model analyzed an input image, identified regions to change or preserve, and used Tencent’s MixGRPO training method to improve instruction following and consistency in untouched regions.

Version 3.5 consolidates those generation and editing paths into one hosted model. The deployment options remain distinct: Hunyuan Image 3.0 is available as open-source code, and 3.5 is currently distributed through Tencent’s paid cloud API and partner applications.

Best-fit workloads

• Product and campaign assets: Generate coordinated 2K images where visual consistency across a batch affects usability.
• Reference-driven editing: Supply multiple source images without adding per-reference charges.
• Design platforms: Use one backend for prompt-based creation and guided edits.
• Character and brand systems: Test whether repeated references preserve identity, clothing, colors, and composition across outputs.
• Chinese and bilingual graphics: Evaluate Hunyuan’s established focus on Chinese-language rendering alongside English text accuracy.

The 30% claim needs a benchmark

Tencent reports a 30% win-rate improvement in human evaluation against Hy Image 3.0. That vendor-reported figure does not indicate a 30-percentage-point gain, and its value depends on the prompt set, sample size, rater instructions, output selection process, and evaluation criteria.

Independent comparisons against current versions of Seedream, Imagen, GPT-image, and Midjourney will provide a clearer competitive picture. Internal testing should use production prompts and reference images, with results scored across several dimensions:

• Prompt and edit-instruction adherence
• Identity and object consistency across a batch
• Text accuracy in Chinese and English
• Preservation of regions excluded from an edit
• Latency, timeout rate, and concurrency behavior
• Safety-filter behavior and rejected-request handling
• Effective cost after retries and discarded outputs

A practical migration checklist

1. Run the same prompts, seeds, aspect ratios, and references through 3.0 and 3.5 where the interfaces permit comparable settings.
2. Separate visual-quality testing in Miora or OnSolo from API tests involving latency, quotas, authentication, and error recovery.
3. Confirm regional availability, rate limits, data-retention terms, content policies, and commercial-use conditions.
4. Measure consistency across full batches instead of selecting a single favorable output.
5. Keep the existing 3.0 path available until the preview meets production thresholds and Tencent publishes a stable release path.