Exposure Fusion (Blend Bracketed Photos)
Blend dark, normal and bright exposures into one photo with detail in shadows and highlights.
How it works
Upload a ZIP of 2–5 bracketed photos of the same scene. The photos are aligned automatically and blended with Mertens exposure fusion: every pixel is weighted by contrast, saturation and good exposure, and a Laplacian pyramid hides the seams. No HDR file or tone mapping is needed, and the result is an ordinary 8-bit PNG. Photo Lab in Lumina runs the same engine on your device for larger photos.
- Weights (Mertens, Kautz & Van Reeth 2007): contrast C = |3 × 3 Laplacian| of the grey image, saturation S = standard deviation of R, G and B, well-exposedness E = exp(−Σ(v − 0.5)² / 2σ²) over the three channels; W = (C + 0.001)^ωc · (S + 0.001)^ωs · E^ωe. The 0.001 floor (not in the paper) keeps grey photos and flat areas from falling back to a plain average.
- Blending: the weights are normalised per pixel, each photo is split into a Laplacian pyramid (5-tap binomial filter, Burt & Adelson 1983) and every level is multiplied by the Gaussian pyramid of its weight map. The summed pyramid is collapsed and clipped to 0–255. The pyramid stops about two levels before 1 pixel.
- Alignment reuses Lumina Astro Lab's registration: Harris corners of a histogram-equalised grey image (at most 1600 pixels per side) take the place of stars in the triangle-matching RANSAC; each photo is matched with its neighbour towards the middle photo, refined by Lucas–Kanade intensity matching with gain and offset, and resampled bicubically. Pixels outside a photo get weight 0; the crop keeps the largest rectangle every photo covers.
- Clipping columns report the share of pixels with a channel at 255 (highlights) or all channels at 0 (shadows) in every input and the result.
Sources
- Mertens, Kautz & Van Reeth (2007): Exposure Fusion, Pacific Graphics 2007 (IEEE)
- Burt & Adelson (1983): The Laplacian Pyramid as a Compact Image Code (IEEE Transactions on Communications)
- Baker & Matthews (2004): Lucas-Kanade 20 Years On: A Unifying Framework (International Journal of Computer Vision)
- Valdes et al. (1995): FOCAS automatic catalog matching (triangle matching) (PASP)
Part of Lumina
-
Lumina Photo editor & image toolkit
Bring out what you saw. A free photo studio, right in your browser.
Related tools
- Focus Stacking (Combine Photos Focused at Different Distances) – Merge photos focused near and far into one image that is sharp throughout.
- Astro Image Stacker (align & stack) – Calibrate, align and stack astro frames with sigma clipping; download a 32-bit FITS.
- Photo Tone Curves & Profile Processor – Apply levels, RGB curves and white balance, or reuse a Lumina photo profile.
- Photo Collage & Image Joiner – Make a photo grid or join photos horizontally and vertically.
How to use Exposure Fusion to blend bracketed photos
Merge a dark, a normal and a bright shot of the same scene into one photo with detail in the shadows and in the sky, without an HDR file.
What it is for
A camera cannot hold a bright window and a dark room in one exposure. Bracketing solves that: you take the same scene two to five times at different exposures. This tool blends those shots into one ordinary photo. Every pixel comes mostly from the photos where that spot is detailed, colourful and close to mid-grey, so the sky comes from the dark shot and the shadows from the bright one.
It is exposure fusion, not HDR merging. There is no radiance map, no 32-bit file and no tone-mapping step, which is why the result usually looks natural instead of "HDR-ish". Upload a ZIP of 2–5 PNG, JPEG or WebP photos of the same size (12 megapixels together). The middle photo of the ZIP is the reference for alignment. Larger sets run in Photo Lab in Lumina, on your own device.
Worked example
The first example ZIP holds three synthetic brackets at −2, 0 and +2 EV of a scene lit 64 times brighter on the right than on the left. The bright frames are hand-held: photo 1 is shifted and turned by 0.4°, photo 3 by −0.3°. With the default settings the tool reported:
- Alignment: photo 1 matched on 70 corners with an RMS of 0.343 px, offset 3.594 / −3.065 px and 0.41° of rotation; photo 3 on 49 corners, −3.689 / 2.418 px and −0.3°.
- Crop: the result is 376 × 251 px instead of 384 × 256, because the turned frames leave thin empty borders.
- Clipping: 16.64 % of the 0 EV photo and 47.24 % of the +2 EV photo have a channel at 255. The fused result has 0 %.
- Share of pixels: 34.05 %, 31.63 % and 34.32 % — each exposure leads in about a third of the picture.
The second example switches alignment off and raises the well-exposedness weight to 2 with σ 0.15. The output keeps the full 384 × 256 px and every photo is blended where it lies, so fine edges soften wherever the shifted frames disagree. That is why alignment is on by default.
How to read the result
Download fused PNG is the photo itself, 8-bit and without camera metadata. The metrics show the size and how much of the result is still clipped. The alignment table lists each photo's offset, rotation and scale relative to the reference, how many corners agreed and the fit's RMS in pixels; "reference" marks the middle photo. Share of pixels says which photo carried the most weight where. Clipping per photo compares highlights at 255 and pure black in every input and in the result. A warning appears when a photo could not be aligned; it is then blended as it is.
Tips and common mistakes
- Shoot the bracket from a tripod or in a fast burst. Alignment fixes shift, rotation and small scale changes, not parallax from moving the camera sideways.
- Anything that moves between frames — people, waves, branches — can appear twice. There is no ghost removal.
- Two stops between exposures is a good default; three shots cover most backlit scenes.
- Lower Well-exposedness width (σ) for a punchier, more local look; raise Contrast weight to prefer textured areas.
- Use the same white balance for all shots, otherwise colours can shift across the picture.
FAQ
Is the result an HDR image?
No. Exposure fusion blends the photos directly into a normal 8-bit picture. It never creates a high-dynamic-range file, so it needs no tone mapping and displays the same everywhere.
Why is my result a few pixels smaller?
Aligned photos do not cover the whole frame. Crop to the area every photo covers keeps the largest rectangle that every photo fills. Turn it off to keep the full size, with edges that show only some of the photos.
Can I use RAW files?
Convert them to PNG or high-quality JPEG first, for example with the HEIC & RAW converter, using the same settings for every frame. The tool decodes PNG, JPEG and WebP to 8-bit sRGB.
Are my photos kept?
Download links expire after about two hours. Photo Lab in Lumina runs the same blend in your browser without any upload.
Guide last updated:
For scripts and AI agents
- JSON API:
POST https://luminaexposurefusion.usefulweb.app/api/v1/calculate(same calculation as this page) - Interface description: OpenAPI 3.1 · AI guide: llms.txt
- MCP server (Streamable HTTP):
https://luminaexposurefusion.usefulweb.app/mcp· tool namelumina_exposure_fusion
Example request
curl -s -X POST https://luminaexposurefusion.usefulweb.app/api/v1/calculate \
-H 'content-type: application/json' \
-d '{…see OpenAPI examples…}'