Tencent: Hy Image 3.5 Preview

HY Image 3.5 Preview is Tencent’s unified image-generation and editing model. It accepts a written brief with optional reference images and supports continuing the same visual task across several turns. The result is an image for a defined scene, composition or edit. Tencent’s examples include product imagery, campaign posters, interface-design visuals and material concepts. Choose whether the task needs a new image, a composition informed by references or a focused change to an existing result.

The producer identifies the edition as hy-image-v3.5-preview and documents a 100K-token prompt limit. Its input is a chronological conversation containing text and images, with the final user message supplying the current instruction. Earlier messages provide context for iterative editing. This differs from HY Image 3.0’s single-prompt workflow and its separate input limits. For a continuous design task, retain the relevant brief, reference roles and chosen image so the next instruction can build on the intended visual state.

Prepare PNG or JPEG/JPG reference images. Reference inputs can contain up to 20 images, each no larger than 20 MB in the documented workflow. An image can be supplied as a referenced or encoded asset. If the conversation carries too many images, the producer guide describes trimming older images first. Keep the active reference set focused rather than relying on every historical image to remain in context. The input-resizing control preserves smaller images and scales larger ones proportionally toward a pixel-area target; its default target is 1,048,576 pixels. Preserve sufficient visible detail for the product, material or layout that needs to guide the result.

Output size has two distinct preparation paths. With automatic sizing, generate_max_pixels sets a target area: 1,048,576 pixels for the 1K tier, 2,359,296 for the default 1.5K tier, or 4,194,304 for 2K. The producer FAQ says an area-target setting supplied alone defaults to a square ratio. When both size controls are omitted, the model uses the brief and the default 1.5K tier. Those tiers do not create an automatic 4K setting. Use an area tier when a square image suits the task. To let the brief guide proportions at the default 1.5K tier, omit both size controls; request explicit dimensions when the layout requires a specific portrait or wide size.

For explicit dimensions, the producer permits positive integer sides from 256 to 8192 pixels, with a total area no larger than 16,777,216 pixels. A 4096 × 4096 square meets that limit. The side range and area ceiling apply together, so an 8192-pixel edge does not permit an 8192 × 8192 image. Request the required width and height explicitly when the final layout needs a specific size, rather than assuming a target-area tier determines the same dimensions. Keep crop-sensitive details and the required empty space visible in the prompt.

The model supports multi-turn editing through returned conversation history. Tencent distinguishes the final delivered image from the intermediate image used as context for the next edit, and its examples retain the assistant’s reasoning, tool request and generated-image result when continuing. For a product composition, first establish the subject and framing; then ask for a lighting change or a new background while specifying what should retain its appearance. Keep the selected result and its history together so the next turn works on that image rather than a discarded variation. In Tencent’s documented workflow, both the final and history image links expire after 12 hours, so preserve the image assets before relying on them for a later edit.

The documented controls include a seed, a custom watermark footnote of up to 16 characters and optional external-search augmentation, which is off by default. The seed is a non-negative integer; zero or omission lets the producer choose randomly. These controls serve different purposes from the brief itself. For a reference-based product visual, describe the product’s appearance, camera angle, surface, lighting and intended layout, then use the reference to guide the identifying details. For a poster, give the hierarchy and exact wording instead of leaving its message to be invented.

Tencent’s practical guidance recommends checking text and layout, especially where product specifications, organisation names or other factual wording matter. Inspect the completed image for the subject, proportions, material, crop and lettering. Retain the final asset with its brief and settings, and use a concrete instruction tied to a visible issue for the next edit.