# Hy Image 3.5 Preview

> Source: https://aiwiki.ai/wiki/hy_image_3_5_preview
> Updated: 2026-10-04
> Fact-checked: 2026-10-04
> Categories: Chinese AI, Generative AI, Image Generation
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Hy Image 3.5 Preview." aiwiki.ai, 4 Oct 2026. https://aiwiki.ai/wiki/hy_image_3_5_preview
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Hy Image 3.5 Preview** is [Tencent](https://aiwiki.ai/wiki/tencent)'s preview model for [AI image generation](https://aiwiki.ai/wiki/ai_image_generation) and editing, announced on September 22, 2026. It accepts text and reference images and belongs to the [Hunyuan](https://aiwiki.ai/wiki/hunyuan) image-model family. Tencent provides hosted access through Tencent Cloud; [ComfyUI](https://aiwiki.ai/wiki/comfyui) also exposes the model through partner API nodes.[1][2][5]

The preview has a distinct API identifier, `hy-image-v3.5-preview`. Its release name should not be confused with [Hunyuan Image 3.0](https://aiwiki.ai/wiki/hunyuan_image_3), the earlier model used as the comparison baseline in Tencent's announcement.[1][3]

## Capabilities

Tencent's examples cover Chinese and English text, posters, infographics, UI layouts, photography, and stylized artwork. The company emphasizes preserving unchanged people and products during edits.[2]

| Task | Documented use |
| --- | --- |
| Text-to-image | Generate an image from a written instruction.[1][3] |
| Reference-based generation | Include images alongside the text instruction.[3] |
| Multi-turn editing | Carry earlier generation results and instructions into a later request.[3] |
| Visual design | Tencent's examples include text-heavy layouts and product or portrait editing.[2] |

## Release and evaluation

Tencent's September 22 announcement described output up to 2K and a price of US$0.024 per image through its international cloud API.[1] The accompanying release article described a blind GSB human evaluation involving more than 100 professional designers and reported an overall win-rate improvement of more than 30% over the preceding model.[2]

This is Tencent's reported comparison, not an independent benchmark. The release article does not give the underlying win and tie counts, absolute win rates, or a calculation that would justify interpreting the figure as a 30-percentage-point increase. It also does not establish equivalent gains on every task.[2]

ComfyUI merged its text-to-image and image-edit partner-node integration on September 22, 2026.[6]

## Output sizes and reference limits

Limits are interface-specific: the launch describes five references and 2K output, while TokenHub documents larger context limits and explicit 4K requests. ComfyUI identifies its 4K output as upscaled from 2K, not native 4K generation.[2][3][5]

| Interface or setting | Documented limit or behavior |
| --- | --- |
| Tencent launch description | Up to five reference images; output up to 2K.[2] |
| TokenHub input limit | 100k tokens.[3] |
| TokenHub image input | Up to 20 images in context, each at most 20 MB; PNG, JPEG, or JPG through a URL or Base64.[3] |
| TokenHub automatic area tiers | `generate_max_pixels` supports 1K, 1.5K, and 2K tiers.[3] |
| TokenHub explicit size | `size` supports `4096x4096`; total area cannot exceed 16,777,216 pixels.[3] |
| ComfyUI edit node | One to five reference images.[5][6] |
| ComfyUI 4K option | A 2K render upscaled by the model.[5][6] |

TokenHub's 20-image context allowance does not increase ComfyUI's five-reference limit.[3][6]

## API access

TokenHub accepts synchronous `model` and `messages` requests authenticated with a bearer API key. The last user message is the current instruction; earlier messages provide context.[3]

For multi-turn editing, append returned `assembled_history` and the next instruction to earlier messages. Generated-image links expire after 12 hours.[3]

| Region | Documented generation endpoint |
| --- | --- |
| Mainland China | `POST https://tokenhub.tencentmaas.com/v1/wand/hunyuan-image/v35-generation`[3] |
| International | `POST https://tokenhub-intl.tencentcloudmaas.com/v1/wand/hunyuan-image/v35-generation`[3] |

## ComfyUI integration

The integration has separate `Tencent HY Image: Text to Image` and `Tencent HY Image: Edit` nodes. Both call Tencent's hosted service through Comfy API. Users must sign in, but these nodes require neither a local GPU nor a model download.[5]

The edit node identifies references in input order with names such as `@Image1` and `@Image2`. Its controls include aspect ratio, output resolution, seed, reference detail, and an optional AI-generated watermark. Custom dimensions must be multiples of 16 and stay within the 4096x4096 area ceiling. The integration's source code validates reference numbering and rejects batches exceeding five images.[5][6]

ComfyUI recommends specifying both the requested change and what should remain unchanged. Higher `reference_detail` processes larger reference images but takes longer. These controls do not guarantee that every edit preserves identity or layout perfectly.[5]

## Pricing

Tencent's launch article listed RMB0.15 per output image at 2K or below in China and US$0.024 internationally, with no input charge.[2] The international TokenHub pricing table retrieved on October 5, 2026 distinguishes output tiers:[4]

| Output tier | Listed token usage per image | Estimated international cost |
| --- | --- | --- |
| 1K | 15,000 | US$0.024[4] |
| 2K | 15,000 | US$0.024[4] |
| 4K | 20,000 | US$0.032[4] |

The table uses US$1.60 per million tokens. Tencent describes these figures as reference costs and says final charges depend on the console and invoice. They are Tencent Cloud rates, not necessarily the price charged by another integration.[4]

## Relationship to Hunyuan Image 3.0

HunyuanImage 3.0's technical report describes an autoregressive multimodal mixture-of-experts model with more than 80 billion total parameters and 13 billion active parameters per token, together with published code and weights. Those disclosures concern 3.0, not the 3.5 preview.[7]

The 3.5 preview's documented access paths are hosted services. ComfyUI's API-node integration is not a downloadable 3.5 checkpoint or evidence that the model inherits 3.0's weight license.[5][6][7]

## References

1. Tencent Hy. [Hy Image3.5 preview is live](https://x.com/TencentHunyuan/status/2102226552310419473). Official announcement, September 22, 2026.
2. Tencent Hunyuan. [Hy Image3.5 preview release: a cost-effective model for professional creation](https://hunyuan.tencent.com/research/hy-image35-preview?langVersion=zh). Chinese-language release article, September 22, 2026. Accessed October 5, 2026.
3. Tencent Cloud. [Hy image generation API guide](https://cloud.tencent.com/document/product/1823/135745), updated September 24, 2026; [international guide](https://intl.cloud.tencent.com/document/product/1300/83708). Model identifiers, requests, image constraints, size controls, and multi-turn history. Accessed October 5, 2026.
4. Tencent Cloud. [Model pricing](https://intl.cloud.tencent.com/document/product/1300/78937), image-generation table. Accessed October 5, 2026.
5. Comfy Organization. [Tencent Hunyuan Image 3.5 Preview in ComfyUI: Text to Image and Edit Workflows](https://docs.comfy.org/tutorials/partner-nodes/tencent/hy-image-3-5). Official integration documentation. Accessed October 5, 2026.
6. Comfy Organization. [Hunyuan Image 3.5 partner-node integration, pull request #16462](https://github.com/Comfy-Org/ComfyUI/pull/16462), merged September 22, 2026; [node source at the merge commit](https://github.com/Comfy-Org/ComfyUI/blob/b09760de30799bdb9c603a9c3e99792c4aefb47a/comfy_api_nodes/nodes_hunyuan_image.py).
7. Tencent Hunyuan Foundation Model Team. [HunyuanImage 3.0 Technical Report](https://arxiv.org/abs/2509.23951), arXiv:2509.23951, submitted September 28, 2025. Used to distinguish the preceding model, not as a technical report for the 3.5 preview.