Hy Image 3.5 Preview: what Tencent shipped
Tencent's Hy Image 3.5 Preview generates and edits in one model; its API guide lists 4K output and 20 references. No weights or report. Rows to use today.
Every guide, comparison and workflow we’ve published on Image Editing.
37 articles — page 1 of 2
Tencent's Hy Image 3.5 Preview generates and edits in one model; its API guide lists 4K output and 20 references. No weights or report. Rows to use today.
Qwen-Image-2.1, released 20 September 2026, generates native RGBA with a true alpha channel up to 2K from a prompt, replacing generate-then-background-remove.
Qwen-Image-2.1 composes up to 10 reference images into one scene by prefix KV cache reuse. Long prompts overload the encoder and add unsolicited Chinese text.
Grok 2.0 is region edit and compositing. GPT Image 2 is instruction-following stills. Pick the lever, not the arena rank.
O1 Image draws a new still. O1 Image Edit changes a supplied still. Same 1 credit, different contracts.
Shipped 7–8 Aug 2026 and #2 on stills arenas: edit one region, composite several stills, and pick Quality Mode on purpose.
HiDream O1 Image Edit is an edit. Point at the object. A new text-to-image throws away the still you approved.
HunyuanImage 3.0 Instruct Edit is an edit. Point at the object. A new text-to-image throws away the still you approved.
Kling Image O1 is an edit. Point at the object. A new text-to-image throws away the still you approved.
Mai Image 2.5 Edit is an edit. Point at the object. A new text-to-image throws away the still you approved.
Nano Banana 2 is an edit. Point at the object. A new text-to-image throws away the still you approved.
Seedream 4.5 is an edit. Point at the object. A new text-to-image throws away the still you approved.
Edit endpoint vs generate. Infographics and dense type are an edit of a layout, not a photoreal hero.
SeedVR Upscale is an edit. Point at the object. A new text-to-image throws away the still you approved.
Photo cutouts go through generate_image_from_image (and sometimes remove_background). Erasing a rejected still, or erasing before the board is locked, spends an image edit on a tile you will not use.
Background text a model got wrong is not repairable, but it is not necessary either. A soft-focus pass that reads as optics, and where it stops working.
Maskless element re-render regenerates one named object. A decision table for when it beats masked inpainting and when the mask is still the safer tool.
Fused fingers are a small-region problem, not a whole-image one. A masked inpaint loop that lands in two or three tries, plus the depth-guided fallback.
Contact with a product is where hands fail. A grip-region mask-and-inpaint recipe, then a composite fallback that keeps the real SKU geometry intact.
Do not re-roll a good image for one misspelled word. Quote the exact string, describe the font, set a mask margin, and know when to composite type instead.
Remodelers and flooring retailers lose deals in the imagination gap. A four-variant render-to-video build from one room photo, plus the sourcing rules.
Reve 2.1 holds first on Artificial Analysis image editing at 1262 Elo. What an editing score actually measures, and how to turn a rank into a revision workflow.
Generative upscalers re-draw letterforms and will respell a word that was already right. Mask text out of the pass, or drop denoise on text-bearing images.
Instruction-based edit models regenerate the whole frame. Scope the change, then composite the region back so the background stays put.