HunyuanImage 3.0-Instruct is the instruction-tuned variant of Tencent's open-source image model — the same native multimodal autoregressive base (80B parameters, about 13B activated, 64-expert mixture) with a visual reasoning layer on top. Before it draws, it runs a short chain-of-thought "think" step, then edits an existing image from a plain-language instruction, applies style transfer, or fuses several references into one coherent composite. That makes it a genuine editor, not just a generator.
Every prompt below is written as an instruction over an attached image (or images). The pattern that matters: name the change, then say what to keep identical — that clause is what protects the pixels you care about and gives a clean, targeted edit. For fusion, describe which element comes from which reference. Keep the Hunyuan Image cheat sheet open for modifiers, browse the best Hunyuan Image prompts for generation, and see Hunyuan text & poster prompts for typography work.
Background swaps & replacement
Replace the setting while protecting the subject. Name the new background, then pin the subject, pose, and the light on them so the model only rebuilds the scene behind.
1. Studio-to-cafe background swap
Replace the plain grey studio background behind the person with a softly blurred warm cafe interior at golden hour. Keep the subject, their pose, clothing, hair, and the studio lighting on them exactly the same. Match the new background's warm color temperature and add a subtle rim on the shoulders so they sit naturally in the scene, with clean edges around the hair. Photorealistic result.Settings: Hunyuan Image 3.0-Instruct · edit · 1 reference image · keep subject identical
Why it works: naming exactly what to keep and how to blend the new light gives a clean composite instead of a re-drawn person.
2. Product background replacement
Replace the cluttered background behind this bottle with a seamless soft sage-green studio backdrop. Keep the bottle, its label, cap, and the existing highlights and reflections on the glass exactly the same. Add a gentle gradient falloff behind it and a soft contact shadow beneath so it reads as a clean product shot. Do not alter the bottle's shape, position, or color.Settings: Hunyuan Image 3.0-Instruct · edit · 1 reference image · keep product identical
Best for: e-commerce listings that need a consistent catalog background without a reshoot.
3. Sky replacement at golden hour
Replace the flat overcast sky in this landscape with a warm golden-hour sky, soft gradient from amber near the horizon to pale blue above with a few wispy clouds. Keep the foreground, buildings, water, and their positions exactly the same, but relight the scene subtly so the warm sky color reflects on the tops of surfaces and the shadows lengthen naturally. Photorealistic, no halos along the skyline.Settings: Hunyuan Image 3.0-Instruct · edit · 1 reference image · keep foreground identical
Why it works: asking it to carry the sky's color into the scene lighting is what stops a pasted-on look.
4. Solid-color catalog background
Cut the shoe out of its current background and place it on a solid pure-white catalog background. Keep the shoe, its angle, materials, laces, and existing lighting exactly the same. Add a soft, natural drop shadow directly beneath it for grounding, keep the edges crisp and clean with no leftover fringe, and center the product with even margins. Flat, even e-commerce studio look.Settings: Hunyuan Image 3.0-Instruct · edit · 1 reference image · keep product identical
Best for: marketplace listings that require a uniform white background across a whole range.
Retouch & cleanup
Fix flaws while keeping the photo honest. Ask for natural results and name what must not change — texture, identity, color — so the edit stays subtle.
5. Natural skin retouch
Gently retouch this portrait: soften temporary blemishes and reduce shine on the forehead and nose, even out the skin tone slightly. Keep natural skin texture, pores, freckles, and fine lines visible — do not smooth the skin into plastic. Keep the person's face, expression, features, hair, and the lighting exactly the same. The result should look like a light professional retouch, not a filter.Settings: Hunyuan Image 3.0-Instruct · edit · 1 reference image · keep identity identical
Why it works: explicitly preserving pores and freckles keeps the retouch believable instead of over-smoothed.
6. Remove stray flyaway hair
Remove the stray flyaway hairs that cross the face and the ones drifting against the background in this portrait. Follow the natural hairline and keep the overall hairstyle, volume, and part exactly the same. Do not change the face, expression, skin, clothing, or background beyond removing those loose strands. The cleanup should be invisible, as if the hair simply sat neatly.Settings: Hunyuan Image 3.0-Instruct · edit · 1 reference image · keep hairstyle identical
Best for: headshots and beauty shots where a few loose strands are the only distraction.
7. Dust and scratch cleanup
Clean up this scanned old photograph: remove the dust specks, scratches, and small creases across the surface, and reduce the yellowed staining in the corners. Keep the original composition, the people, their faces and clothing, and the vintage tonality exactly the same — do not colorize it or sharpen it into a modern look. The result should read as the same photo, just cleanly restored.Settings: Hunyuan Image 3.0-Instruct · edit · 1 reference image · keep composition identical
Why it works: telling it not to colorize or over-sharpen keeps the restoration faithful to the original.
8. Even out uneven lighting
Balance the lighting in this interior photo: lift the dark shadowed corner on the left and tame the blown-out highlight near the window, so the room reads with even, natural exposure. Keep the furniture, layout, colors, and every object in the same position, and keep the warm daylight mood. Do not add or remove anything — only correct the exposure and recover detail in the shadows and highlights.Settings: Hunyuan Image 3.0-Instruct · edit · 1 reference image · keep layout identical
Best for: real-estate and interior shots taken in tricky mixed light.
Add & remove objects
Add or delete a single element and let the rest stand. Describe the object precisely, say where it goes, and lock everything else so the edit stays local.
9. Remove a background bystander
Remove the person walking in the background on the right side of this photo, and reconstruct the wall and pavement behind them so the fill matches the surrounding texture, perspective, and lighting seamlessly. Keep the main subject in the foreground, their pose, and everything else in the frame exactly the same. The result should look as if that bystander was never there.Settings: Hunyuan Image 3.0-Instruct · edit · 1 reference image · keep subject identical
Why it works: asking it to match texture, perspective, and lighting in the fill is what makes the removal invisible.
10. Add a reflection under product
Add a subtle, realistic mirror reflection of this bottle on the glossy surface directly beneath it, fading out softly toward the bottom and matching the bottle's colors and the scene's lighting. Keep the bottle, its label, the background, and the existing shadows exactly the same. The reflection should sit at the correct angle for the surface and feel physically plausible, not a flipped copy.Settings: Hunyuan Image 3.0-Instruct · edit · 1 reference image · keep product identical
Best for: cosmetics and drinks shots where a clean reflection adds a premium feel.
11. Add steam to a coffee cup
Add gentle wisps of rising steam above the coffee in this cup, soft and semi-transparent, catching the warm light and drifting slightly to one side. Keep the cup, the coffee, the table, the props, and the lighting exactly the same. The steam should look natural and freshly poured, subtle rather than heavy, and fade out as it rises.Settings: Hunyuan Image 3.0-Instruct · edit · 1 reference image · keep scene identical
Why it works: steam is a small, believable add that signals "hot and fresh" without touching the rest of the shot.
12. Remove a logo from clothing
Remove the printed brand logo from the front of the person's t-shirt and reconstruct the fabric behind it, matching the shirt's exact color, weave texture, folds, and the lighting across it. Keep the person, their pose, face, the shirt's shape, and everything else in the frame exactly the same. The area should read as plain fabric with no trace of the logo.Settings: Hunyuan Image 3.0-Instruct · edit · 1 reference image · keep subject identical
Best for: clearing trademarks from stock and lifestyle shots before publishing.
Restyle & style transfer
Change the medium, keep the picture. Pin the composition and subject positions so the restyle re-paints the same scene instead of inventing a new one.
13. Repaint photo as ink-wash
Repaint this photograph as a traditional Chinese ink-wash painting. Keep the overall composition, the mountain silhouette, and the position of the boat on the lake. Translate the tones into soft washes of grey and black with expressive brush strokes and visible rice-paper texture, and add generous negative space in the sky. Keep it clearly recognizable as the same scene, just rendered in a different medium.Settings: Hunyuan Image 3.0-Instruct · edit · 1 reference image · preserve composition
Why it works: style transfer keeps the structure while changing the medium — pinning the composition stops it drifting into a different picture.
14. Convert portrait to oil painting
Restyle this portrait as a classical oil painting with visible brushwork, rich impasto texture, and warm gallery lighting. Keep the person's likeness, pose, expression, and the composition exactly the same, so it stays recognizably them. Render the skin, hair, and clothing in painterly strokes with deep shadows and soft highlights, in the manner of a formal 19th-century portrait.Settings: Hunyuan Image 3.0-Instruct · edit · 1 reference image · preserve likeness
Best for: gift-style portraits and stylized avatars that must still look like the person.
15. Anime restyle of a photo
Convert this photo of two friends on a street into clean cel-shaded anime art. Keep the composition, their poses, hairstyles, outfits, and the background layout the same, and keep their expressions recognizable. Use bold clean linework, flat vibrant colors, simple shading, and a bright anime sky. It should feel like the same moment redrawn as a frame from an anime, not a new scene.Settings: Hunyuan Image 3.0-Instruct · edit · 1 reference image · preserve composition
Why it works: naming poses, outfits, and layout as fixed keeps the anime version tied to the original frame.
16. Match a reference art style
Restyle image 1 to match the art style of image 2: adopt image 2's palette, brushwork, and level of texture, while keeping image 1's subject, composition, and the positions of every element exactly the same. Do not copy image 2's content — only borrow its look and feel. The result should read as image 1's scene painted in image 2's style.Settings: Hunyuan Image 3.0-Instruct · fusion · 2 reference images · preserve composition of image 1
Best for: keeping a consistent house style across a set of images by pointing at one reference.
Multi-image fusion
Combine references into one coherent composite. Say plainly which element comes from which image, and which reference sets the lighting and perspective everything must match.
17. Product into a lifestyle scene
Combine the references into one photorealistic scene: place the perfume bottle from image 1 on the marble tabletop from image 2, matching the perspective and the soft window light of image 2. Add a natural contact shadow beneath the bottle. Keep the bottle's shape, label, cap, and color exactly as in image 1. Premium product-photography look, with the bottle sitting believably in the space.Settings: Hunyuan Image 3.0-Instruct · fusion · 2 reference images
Why it works: naming which image supplies the light and perspective tells the model what everything else must conform to.
18. Face onto a new outfit
Fuse the references: keep the person's face, hairstyle, and skin tone from image 1 exactly the same, and dress them in the outfit shown in image 2, fitting it naturally to their body and pose. Match the lighting and color of image 1 so the clothing looks worn in that scene, not pasted on. Keep the background of image 1. The face must stay clearly the same person.Settings: Hunyuan Image 3.0-Instruct · fusion · 2 reference images · keep face from image 1
Best for: fashion mockups that pair a consistent model with different garments.
19. Three-reference brand composite
Build one clean brand image from the references: take the sneaker from image 1, place it on the textured concrete surface from image 2, and set them against the gradient backdrop color from image 3. Match a single consistent studio light across all three, add a soft contact shadow, and keep the sneaker's shape, colors, and materials exactly as in image 1. Cohesive, high-contrast commercial look.Settings: Hunyuan Image 3.0-Instruct · fusion · 3 reference images · keep product from image 1
Why it works: assigning each reference one job — product, surface, backdrop — keeps a three-image composite from turning muddy.
20. Object plus texture fusion
Fuse the references into one image: keep the shape and form of the handbag from image 1, but apply the leather texture and color from image 2 to its surface, following the bag's contours, seams, and folds so the material wraps naturally. Keep the bag's hardware, straps, and the studio lighting exactly the same. The result should look like the same bag manufactured in the new material.Settings: Hunyuan Image 3.0-Instruct · fusion · 2 reference images · keep form from image 1
Best for: material and colorway previews for products without a physical sample.
Text & product edits
Hunyuan renders best-in-class English and Chinese type, so text swaps are a real strength. Quote the old and new copy, keep strings short, and lock the font, color, and position.
21. Swap a headline on a poster
On this poster, change the headline that currently reads "SUMMER SALE" to read "WINTER SALE" instead. Keep the same font, weight, size, color, letter spacing, and position, and keep every other element — the artwork, the smaller text, the layout, and the background — exactly the same. The new headline must be sharp, correctly spelled, and sit in the identical spot as the old one.Settings: Hunyuan Image 3.0-Instruct · edit · 1 reference image · keep layout identical
Why it works: quoting both strings and pinning the type styling lets the model rewrite only that headline and nothing else.
22. Recolor a packaging label
Recolor the label on this coffee bag from terracotta to deep forest green, keeping the label's layout, the text "EMBER ROAST", the icon, and every printed detail exactly the same — only the background color of the label changes. Keep the bag's shape, the studio lighting, and the reflections the same, and make sure the type stays legible and correctly spelled against the new color.Settings: Hunyuan Image 3.0-Instruct · edit · 1 reference image · keep text identical
Best for: testing packaging colorways before committing to a print run.
23. Add a bilingual caption to an image
Add a two-line caption across the lower third of this photo: the English line "SPRING FESTIVAL" in a bold clean sans-serif, and directly below it the Chinese line "春节" in a matching weight, both centered and in white with a subtle shadow for legibility. Keep the underlying photo, its colors, and composition exactly the same. Both lines must be sharp and correctly rendered in each language.Settings: Hunyuan Image 3.0-Instruct · edit · 1 reference image · keep photo identical
Why it works: bilingual English-Chinese type is exactly where Hunyuan pulls ahead — the caption stays legible in both scripts. More in Hunyuan text & poster prompts.
24. Fix a misspelled sign
The shop sign in this photo reads "CAFE MOARNING" with a typo. Correct it to read "CAFE MORNING", keeping the exact same font, color, size, spacing, and position on the sign. Keep the sign board, the storefront, and the rest of the scene exactly the same. The corrected text must match the original lettering style so the fix is seamless and correctly spelled.Settings: Hunyuan Image 3.0-Instruct · edit · 1 reference image · keep scene identical
Best for: cleaning up typos in signage and mockups without rebuilding the shot. For the full toolkit, keep the Hunyuan Image cheat sheet handy.
Frequently Asked Questions
What is HunyuanImage 3.0-Instruct?
Instruct is the instruction-tuned variant of Hunyuan Image 3.0, Tencent's open-source native-multimodal image model. On top of text-to-image generation it adds visual reasoning — a chain-of-thought "think" step before it draws — plus editing an existing image from a plain-language instruction, style transfer, and multi-image fusion that combines several references into one coherent composite. For edits you feed the image, give a clear instruction, and say what to keep identical.
How do I edit part of an image without changing the rest?
Name the exact change and then spell out what must stay identical. Instead of "make it look nicer," write "replace the background with a blurred cafe, keep the subject, their pose, clothing, and the lighting on them exactly the same." The keep-identical clause is doing the real work — it tells the model which pixels to protect while it rebuilds the rest, so you get a targeted edit instead of a re-drawn picture.
How many reference images can I use for fusion?
Keep it small — two or three references usually give the cleanest composite, and that covers most real jobs like putting a product into a scene or combining a subject, a background, and a texture. The exact maximum depends on the host you use (the official Tencent portal, Hugging Face, GitHub weights, or an API like Tencent Cloud or WaveSpeed), so check its limits. In the prompt, describe which element comes from which reference — "the bottle from image 1 on the surface from image 2."
Does Instruct keep faces consistent when editing?
It holds identity well when you tell it to. Add an explicit instruction such as "keep the person's face, expression, and skin tone exactly the same" so the model treats the face as fixed and edits only around it. For heavy changes — a new outfit or a new background — that clause is what stops the face from drifting. If a face does shift, tighten the keep-identical wording and regenerate rather than adding more edits.
How does style transfer work in Hunyuan Instruct?
Feed the source image and describe the target medium — ink-wash, oil painting, flat anime, watercolor — while pinning the structure you want preserved: composition, the main subject's position, the horizon. Style transfer changes the medium, not the layout, so the keep-composition clause stops it drifting into a different picture. You can also attach a style reference and ask it to match that reference's palette and brushwork.
Can Instruct swap or fix text inside an image?
Yes, and it is a real strength — Hunyuan renders best-in-class bilingual English and Chinese type. Put the old and new copy in double quotes: "change the headline that reads 'SUMMER SALE' to 'WINTER SALE', keep the font, size, color, and position the same." Keep each string short, name what stays fixed, and the model rewrites just that text while leaving the rest of the layout intact.
What resolution can Instruct output?
Up to 2048x2048 (2K). There is no native 4K, and there are no inline flags like --ar — resolution and aspect ratio are interface or API settings you pick in whatever host you use. For edits, working at 2K keeps detail like hair edges and fine text crisp; if you need a larger deliverable, upscale the 2K result afterward in a separate tool.
Where can I access Hunyuan Image 3.0-Instruct?
Hunyuan Image 3.0 is Tencent's open-source model under an Apache 2.0 license — an 80B-parameter (about 13B activated) 64-expert mixture-of-experts, and the largest open-source image model at launch. You can run the weights yourself from Hugging Face or GitHub (Tencent-Hunyuan/HunyuanImage-3.0), use the official portal at hunyuan.tencent.com, or reach it through APIs like Tencent Cloud and WaveSpeed AI. Instruct availability varies by host, so check which variant each one exposes.