Identify the kind of flicker before processing
Video deflicker can mean correcting whole-frame brightness changes, local lighting bands or unstable generated detail. Those problems need different treatment. A filter that smooths frame luminance cannot be assumed to repair a changing face, a crawling texture or an object that changes shape.
The free local tool above makes a conventional brightness preview for a short clip. It uses a 13-frame arithmetic deflicker filter, not an AI restoration model. Use it to test a global exposure problem and compare the actual result with the original before accepting a processed version.
Try a short supported clip on your device
Choose a complete MP4, MOV or WebM up to 12 MB and 15 seconds, at 720p or smaller and up to 30 fps. The preview requires even pixel dimensions and 8-bit SDR material. HDR and high-bit-depth inputs are rejected rather than silently treated as ordinary SDR.
Select Create brightness preview after choosing a clip, or load the original exposure example. The first run downloads the processing engine from jsDelivr, about 31 MB. The selected clip is read into local browser memory and is not uploaded for conversion. A cancel control stops the current operation; a smaller clip can be useful on a device with less memory.
The result is an MP4 preview with re-encoded video and AAC audio when an audio track is present. Other tracks, captions and metadata are not a preservation promise. Keep your original source. Play the result before downloading it and check whether intentional light changes have been flattened along with the unwanted variation.
Compare the original exposure fixture and filtered reference
Watch this clip and view its details
Watch this clip and view its details
The original eight-second fixture has a moving blue object, a fixed grey patch and a modest whole-frame brightness pulse. Both files are 640 by 360 at 24 fps, with 192 frames and no audio. These are authored calibration clips, not AI footage or a quality comparison between commercial products.
The downloadable processed reference uses the same named filter in a local FFmpeg build. The browser engine is a separate build, so file bytes and encoding details need not match that reference. Inspect your actual browser output rather than treating the included processed clip as the result of your own request.
In the fixed patch, encoded-luma standard deviation fell from about 6.56 to 0.47 for this controlled reference, excluding 13 frames at each end. The frame measurements and scope-limited report make that observation inspectable. It is not a perceptual score, a general restoration benchmark or a promised improvement for other footage.
Understand the filter window
The official FFmpeg filter documentation defines deflicker as smoothing temporal frame-luminance variation and exposes a window measured in frames plus an averaging mode. This tool uses deflicker=size=13:mode=am, a 13-frame arithmetic mean. The window’s time span depends on the clip’s frame rate.
ffmpeg -i input.mp4 -vf "deflicker=size=13:mode=am" -c:v libx264 -pix_fmt yuv420p -c:a aac output.mp4That command is a conventional local-processing example, not an AI model request. Review the output at the intended playback rate. More smoothing is not automatically better when the source contains a deliberate fade, a moving light or a meaningful exposure change.
Keep brightness correction separate from generated-detail repair
If the face, fabric or background changes from frame to frame, identify that instability separately. A stable frame mean can coexist with a changing pattern, and a lower brightness-variation measurement does not prove that identity or texture is correct.
The CVPR 2023 blind-deflickering research project studies a neural approach for captured, processed and generated videos. It is a separate research system from this browser filter. We did not install or execute its model, and its research examples are not the source of our calibration clip.
For an unstable generated take, a revised reference, shorter operation or replacement scene may be more appropriate than repeatedly filtering it. Use the action-scene review guide when the defect is causal order, body continuity or screen geography. Those are creative and temporal-structure problems, not just brightness measurements.
Review source and result at the same size
Compare the source and processed export at a matched display size. Inspect the first and last moments, a representative moving section and any intentional lighting change. Check duration, frame rate, framing and sound in the actual returned file.
The local video comparison guide is useful for reviewing two files side by side. Upscaling, interpolation and deflickering are separate operations. A larger output or smoother cadence does not establish that the original exposure or detail was preserved.
Use the blank finishing ledger to record the observed problem, selected filter or model, settings, actual output and acceptance. Keep unsuccessful trials and the original source so a later pass can be reviewed without treating the previous result as ground truth.
Return to the scene source when the defect is structural
For a new shot, use Clout AI Cinema or Image and Video to work from accepted references and inspect a short take before a longer sequence. Check the available operation and quote. This page’s local brightness filter does not establish a dedicated AI deflicker integration inside Clout.
Open the dedicated source fixture and filtered reference players to practice the comparison. A useful finishing decision names what improved, what changed unintentionally and which actual file was accepted.
