What is the flicker-aware gradient exposure feature for low light photography?
Artificial lighting flickers at 50 or 100 cycles per second depending on the country’s power grid frequency, far too fast for a person to notice by eye, but well within range of a camera’s exposure timing. When a camera’s exposure length lands awkwardly against that flicker cycle, the resulting photo or video frame shows visible banding, alternating light and dark stripes across the image.
How does Xiaomi solve this problem?
One of Xiaomi’s innovations aim to solve this by capturing several frames at different exposure lengths in a burst, then choosing two adjacent ones. One frame uses a short exposure, shorter than the light source’s flicker cycle.
The other uses a long exposure, timed to an exact multiple of that cycle, which the patent calls a safe shutter time, a duration long enough to average out the flicker completely. The short frame freezes motion but carries the banding. The long frame is flicker-free but blurs anything that moved. A neural network separates the frame into motion areas and static areas. The system then removes flicker stripes from the short frame using the long frame as a brightness reference, and stitches the final image together: the de-flickered short-exposure data covers the moving subject, and the clean long-exposure data covers everything that held still.
Why might Xiaomi need this innovation for its next product?
The Xiaomi 17 Ultra earned strong marks for handling low light overall, but banding under artificial indoor lighting is a specific and recognisable failure mode that shows up in videos or LED fixtures running on mains power. It is not a subtle defect. Once someone notices the striping in a video, it is hard to unsee, and it becomes a recurring complaint in any camera comparison that includes a scene lit this way.
Will Xiaomi’s IP translate to the 18 Ultra?
Most likely, yes.
Every layer required for this feature already exists on Xiaomi’s current hardware. It needs a burst capture sequence, which the sensor and ISP already support. It needs a neural network to separate motion from static regions, which runs on the same NPU capacity Xiaomi already uses for other computational photography features. There is no sensor redesign, no new component to source, and no external supplier to wait on.
However, there is no public signal yet that this specific mechanism is scheduled for the 18 Ultra, so the prediction rests on the fact that nothing stands in its way rather than on a leak confirming it directly.
What is the residual noise map denoising architecture?
Multi-frame denoising typically works by feeding several captured frames into a large neural network, which learns to average out noise while preserving detail. The computational cost of that approach scales with how many frames go in, and running a large network across many full-resolution frames is expensive in terms of processing time and battery.
Xiaomi’s patent changes the order of operations. Multiple overlapping frames are still captured and fused together, but that fusion step happens through conventional temporal averaging with registration and warping, not through the neural network. Alongside the fused image, the system also generates something called a residual noise map, a record of exactly how much noise remains in each region of the image after fusion, based on the statistical properties of the original per-frame noise. Only the fused image and this residual map, not the full stack of original frames, get passed into a single, smaller neural network for final cleanup. Because the network already knows where noise remains and how much, it does not need to reconstruct that information from scratch, so it can be considerably smaller while still cleaning the image effectively.
Why might Xiaomi need this?
The 17 Ultra already performs well on noise in low light. DXOMARK specifically noted this as a strength. That is worth being honest about upfront: some version of effective multi-frame denoising is very likely already running on the 17 Ultra, whether or not it uses exactly this residual map architecture. What this patent predicts is not a new capability appearing out of nowhere, but a more efficient way of achieving the same result, using a smaller neural network and less processing time to reach the same or better denoising quality.
Will Xiaomi’s IP translate to the 18 Ultra?
Very likely, though the honest framing matters here. If the 17 Ultra’s current denoising already uses a heavier neural network processing full frame stacks, the 18 Ultra’s improvement would come from adopting this leaner architecture, faster processing, lower power draw per shot, or the ability to run denoising at higher resolution without the computational cost that a full-stack approach would require. It is a refinement to how an existing strength is achieved, not the introduction of a new one. Oppo and TCL’s comparable hybrid approach shipping in flagship devices confirms the general architecture pattern, frame fusion plus a lighter secondary network, is commercially viable at this product tier.
How will this impact low light photography if it ships?
Most users will not notice a difference in the output itself, because the noise performance was already competitive. What changes is speed and consistency. Burst shooting, video recording, and multi-shot night modes become less taxing on the processor and battery, which can translate into faster shot-to-shot time or the ability to apply this level of denoising to video, where computational budget per frame is far tighter than for a single still photo.
What is the artifact probability weighted fusion technique?
When a camera combines several long-exposure frames to build a cleaner, less noisy final image, any camera shake or subject movement between those frames creates a mismatch, a ghost or smear where the same object appears in slightly different positions across the stack. Most systems detect this by comparing frames and applying a fixed brightness difference threshold: if two frames differ by more than a set amount at a given pixel, that pixel is flagged as a motion artefact and treated differently during fusion. The problem with a fixed threshold is that it cannot tell the difference between actual motion and ordinary sensor noise in a flat, low-detail area, like a plain wall or a clear sky, so it sometimes misclassifies quiet, noisy regions as motion and produces visible errors where none should exist.
Xiaomi’s patent replaces the fixed threshold with a probability calculation. Using a noise variance coefficient specific to that camera’s sensor, calibrated in advance, the system builds a map of how much random noise to expect at each pixel given its brightness level. It compares this expected noise level against the actual difference between frames, and feeds both values into a probability density function, a statistical formula that outputs the likelihood a given pixel is showing real motion rather than ordinary noise. Pixels judged more likely to be artefacts get less weight during the final fusion, pixels judged likely to be clean noise get full weight. An adjustable correction factor lets the system be more conservative on cameras with weaker sensors and less conservative on cameras with cleaner ones.
Why might Xiaomi need this?
Long-exposure multi-frame fusion is central to how the 17 Ultra and its LOFIC sensor achieve their dynamic range and noise performance. Any weakness in how the system distinguishes real motion from noise shows up as faint ghosting or smearing artefacts in low light, exactly the kind of flaw that a demanding reviewer or a pixel-peeping comparison would catch, even if a casual viewer would not.
Will Xiaomi’s IP translate to the 18 Ultra?
Very likely, for the same structural reason as the flicker feature. This is a post-capture software step. It needs a calibrated noise profile for the sensor being used, which Xiaomi would generate as part of standard sensor qualification regardless, and a probability calculation that runs on existing processing hardware. There is no sensor change, no external component, and no dependency on any supplier’s roadmap. The main open question is not whether Xiaomi can implement it, but whether it becomes a visible, marketed improvement in the 18 Ultra or a quiet accuracy refinement inside an existing fusion pipeline that never gets a name on a spec sheet.
How will this impact low light photography if it ships?
Long-exposure night shots and multi-frame HDR captures would show cleaner detail in flat, low-texture regions, walls, sky, out-of-focus backgrounds, precisely the areas where a fixed-threshold system is most likely to misfire. The improvement is subtle in any single photo but becomes visible in side-by-side comparisons against a phone using a cruder threshold method, particularly in scenes with a mix of a still background and a slightly moving foreground subject, a hand holding a drink, hair moving in a breeze, where current systems most often show faint doubling or smear artefacts.