Apply a reference style while preserving the source image

TeleStyle V2 is an image editing workflow for content-preserving style transfer. You give it a source image, an optional style reference, and a short prompt. The model then tries to keep the main content from the first image while applying the visual style, material, color, or stroke cues from the second image.

The practical difference from a plain text-to-image prompt is control. If you already have a product photo, character sketch, room image, or campaign visual, you can use that image as the anchor instead of asking a generator to recreate it from scratch. The Space also includes prompt helpers that describe the content and style references, which can make the first test easier if you are not sure how to write the prompt.

Start with the public TeleStyle V2 Space. Upload one clear content image and one visibly different style image. For a first useful test, keep the prompt narrow: ask it to transfer the style of the second image while preserving the content and characteristics of the first image. Check whether the subject, layout, and important details survive the restyle.

For local testing, use the Tele-AI model page and GitHub repo. The public workflow loads Qwen Image Edit plus TeleStyle V2 LoRA weights, so it is reproducible from open components. The caveat is hardware: the repo notes testing on an Nvidia H100 80GB GPU, so most readers should treat the Space as the first evaluation path.

This is still a GPU-heavy research workflow. Use the demo to judge content preservation first, especially on faces, fine text, small objects, or scenes where the layout matters more than the overall style.

• Restyle product, campaign, or concept images without rebuilding the scene from zero.
• Turn sketches or existing visuals into several different visual treatments.
• Compare how one source image behaves across multiple style references.
• Use a style image as a more concrete guide than a long descriptive prompt.
• Test whether a local image-editing workflow is worth setting up for heavier batches.

Try the demo

View model page