Creating a Cinemagraph from a Still Image Using Region Masks and Motion Brushes

Creating a cinemagraph from a still image: which models support region masking, the source-image rules, and diagnosis of mask boundary artefacts that break the illusion.

Creating A Cinemagraph From A Still Image Using Region Masks And Motion Brushes

You have a still image, a portrait with wind-blown hair, a landscape with a waterfall, a product shot with steam rising from a cup, and you want exactly one element to move while everything else stays frozen. That is an image-to-video cinemagraph region mask workflow, and it requires a model that can apply motion to a selected area only. As of 2026-09-26, four commercial models offer this: Pika 2.0, Runway Gen-3 Alpha, Kling 1.6, and Kuaishou's KwaiCut. Open-weights models do not expose region masks natively, though you can approximate the effect with frame-by-frame compositing in ComfyUI. The central problem is not getting motion into one area; it is keeping the mask boundary invisible so the moving region does not shimmer, slide, or flicker against the static background. This page names the models that can do it, gives the exact settings, and diagnoses the three failure modes that break the illusion.

AI Cinemagraph Generation: Models With Region-Mask Support

Only two of the models listed in the research expose a true region-mask interface where you paint the moving area directly. Pika 2.0 calls this feature Pikadditions and places it in the scene-ingredient panel. Kling 1.6 calls it Motion Brush, available on the klingai.com web application. Both require a PNG or JPEG source image at the model's native aspect ratio, 1920×1080 px for both, and produce clips of 4 to 10 seconds. Runway Gen-3 Alpha offers Motion Brush in its Director Mode, but the mask is a soft bounding box rather than a per-pixel selection, which increases the risk of background sliding near the edges. Luma Dream Machine does not support region masking; its motion control is limited to whole-frame camera moves such as pan, tilt, zoom, and roll.

Open-weights models such as Stable Video Diffusion, Wan 2.1, and LTX-Video do not ship with region-mask controls. You can simulate the effect by generating a full-frame animated clip, compositing it with a static copy of the source image in an FFmpeg overlay, and masking the boundary by hand. This workflow is time-intensive and produces seam artefacts unless you blur the mask edge in a video editor.

Selective Animation Image-To-Video: Pika 2.0 Settings And Prompt

Pika 2.0 is the most direct tool for selective animation image-to-video because its region mask is a pixel-level paint tool rather than a motion-strength slider. Upload your source image in landscape 1920×1080 px. Click Pikadditions and paint the area you want to animate, waterfall, hair, steam, flag, fabric. Set the motion strength slider to a value between 30 and 60 on an unnumbered scale; the default is 40. A motion strength above 70 causes the mask boundary to tear, pulling pixels from the static region into the animated area. Set the frame count to 4 seconds for a single motion loop or 8 seconds if the element completes a cycle such as a swinging pendulum. The prompt should name the moving element and the type of motion: "flowing water cascading downward, gentle ripple on surface, wind moves the mist leftward." Do not describe the background or the overall scene; the model anchors those to the source image. Pika 2.0 produces audio optionally via Pikaffects sound effects; disable it unless you need ambient noise, because the audio generation adds about 2 seconds of processing time per clip.

For a waterfall shot, the correct prompt is "water flowing down, splashing at the base, mist drifting right." The model generates 25 frames of 1080p video. If the waterfall appears to stutter or skip, reduce motion strength to 20 and regenerate. If the rock beside the waterfall begins to shimmer, the mask has leaked into the static region, repaint the mask with a tighter brush.

Model Region-Mask Capabilities As Of 26 September 2026

Motion Brush Image-To-Video: Kling 1.6 Settings And Prompt

Kling 1.6 is the second model with native region-mask support. Its Motion Brush tool works as a combined bounding box and paint interface: you draw a rectangle around the element to animate, then paint the specific area within that box. The model interprets the paint stroke as the region to move and the box as a soft boundary to reduce flicker near the mask edge. Upload your source image at 1920×1080 px. Set the clip length to 10 seconds, Kling 1.6 maintains motion coherence longer than Pika 2.0, typically 8 usable seconds before subject drift becomes visible. Set the motion strength to a value of 3 on a 1-5 scale for water; use 4 for fast movement such as wind-blown fabric. The camera controls are pan, tilt, zoom, and roll; leave them at zero for a cinemagraph. The start-and-end-frame feature is available but counterproductive for a cinemagraph, it forces the element to return to its exact start position, which produces a mechanical loop. Leave end frame empty.

The prompt for Kling 1.6 must name the motion direction: "hair blowing in a strong wind from right to left, strands separate and wave." Kling 1.6 responds best to prompts that specify speed and direction. Do not use the word "cinemagraph" in the prompt; the model has not been fine-tuned on the term. If the mask boundary causes a visible seam, the hair appears to detach from the scalp, reduce the bounding box size and regenerate. If the background behind the hair begins to warp, the motion strength is too high; reduce to 2.

Still Image With Moving Element: Preventing Mask Boundary Artefacts

The illusion of a still image with moving element breaks when the mask boundary is visible. Three failure modes appear, and each has a specific cause and fix.

Mask Boundary Artefact

The edge between the moving region and the static background shimmers, slides, or shows a thin black line. This happens because the model's temporal attention layers try to smooth the transition between the two states, creating a frame-by-frame jitter. Fix: in Pika 2.0, paint the mask at least 20 pixels inside the moving element's edge, never paint to the exact boundary. In Kling 1.6, use the bounding box to create a 30-pixel buffer around the painted area. In Runway Gen-3 Alpha, the soft bounding box cannot be tightened, so use a slightly smaller box than the element and accept that the edges will be softer.

Flicker In The Static Region

The background that should remain frozen shows subtle frame-to-frame brightness shifts. This is flicker caused by the latent-space decoder producing slightly different luminance values across frames. Fix: preprocess the source image to reduce local contrast. A flat, evenly lit background flickers less than one with high-contrast edges. If flicker persists, post-process the clip in FFmpeg: extract frames, apply a temporal median filter to the background region only, then re-encode. The command for frame extraction is ffmpeg -i clip.mp4 frames/frame_%04d.png; after filtering, encode at 24 fps with H.264.

Subject Drift

The animated element drifts across the frame over the clip duration, pulling the static region with it. Subject drift is worse in Runway Gen-3 Alpha than in Pika 2.0 or Kling 1.6 because its soft bounding box leaks motion into the surrounding pixels. Fix: keep the clip short, 4 seconds maximum in Gen-3 Alpha. For longer clips, generate multiple 4-second segments in Pika 2.0 and concatenate them in FFmpeg with a dissolve transition between segments.

All three models produce clips at 24 fps. The native frame rate for Pika 2.0 and Kling 1.6 is 24 fps; for Runway Gen-3 Alpha it is 30 fps. If you need a 24 fps output for a film project, re-encode in FFmpeg: ffmpeg -i input.mp4 -r 24 output.mp4. Do not use frame interpolation on a cinemagraph, it introduces motion where you want stillness.

Cinemagraph Workflow: Source Image Requirements

The source image determines whether the result looks like a deliberate creative choice or a failed animation. A cinemagraph works only when the moving element can plausibly exist in isolation within the composition. A person blinking, a car driving past, a bird taking flight, these require the entire context to move with them, and the mask cannot hide it. The reliable subjects for an image-to-video cinemagraph region mask workflow are flowing water, steam, hair, smoke, flags, curtains, rotating fans, swaying grass, and falling leaves. Each of these can animate for 4-10 seconds without needing the background to acknowledge the motion.

Avoid faces. A smiling face with moving hair looks unnatural because the hair masks a tight area around the forehead, and the skin underneath needs to be perfectly static. The model's temporal consistency on skin is poor, it introduces micro-movements that read as flicker. Avoid reflective surfaces. A waterfall with a mirror-like pool at the base will show the reflection moving independently of the actual water, creating a false parallax effect. Avoid thin vertical lines such as grass blades against a pale sky, the mask boundary artefact is most visible where high-contrast edges meet the static background.

The source image must be at the model's native resolution or higher. Pika 2.0 and Kling 1.6 accept 1920×1080 px input; if your image is smaller, the model upscales it, softening the static background. For open-weights models such as Stable Video Diffusion, the minimum resolution is 1024×576 px; a smaller image causes aspect-ratio distortion because the model pads or crops to fit. The aspect ratio must match the model's training aspect ratio, 16:9 for landscape, 9:16 for portrait. A square or ultra-wide image forces a naive resize that stretches the subject.

Commercial-Use Licence And Watermarking For Cinemagraphs

A cinemagraph generated for a client project, a broadcast ad, or a social-media campaign requires a commercial-use licence from the model vendor. Pika 2.0 permits commercial use on paid plans, with a metadata-only watermark that satisfies YouTube Creator Studio labels and the EU AI Act Article 50(2) obligation. Kling 1.6 similarly permits commercial use on paid plans, with metadata watermarking. Runway Gen-3 Alpha permits commercial use on its paid tier, but the visible watermark on the free tier cannot be removed without upgrading. Luma Dream Machine places a visible watermark on all outputs, including paid plans, and does not permit commercial use under its standard terms as of 26 September 2026, you need a separate enterprise agreement.

The EU AI Act Article 50(2) requires providers to mark AI-generated outputs in a machine-readable format. Both Pika 2.0 and Kling 1.6 embed metadata that satisfies this requirement. YouTube's AI-generated content disclosure, effective March 2024, requires creators to label synthetic content; a metadata-only watermark does not exempt you from declaring it in the Creator Studio. Meta and TikTok have similar policies effective May 2024. The US Copyright Office's 2025 report states that prompts alone do not provide sufficient human control for copyright registration; a cinemagraph where you painted the mask and wrote the prompt likely qualifies as a human-authored work if you selected the source image and adjusted the mask boundary, but the Office determines this case by case.

Runway Gen-3 Alpha Cinemagraph Workflow

Runway Gen-3 Alpha is the third model that can produce a cinemagraph, but its region-mask implementation is less precise than Pika 2.0 or Kling 1.6. The Motion Brush in Director Mode places a rectangular bounding box over the area to animate. You cannot paint a freeform shape, so the mask always includes some static background. Gen-3 Alpha outputs at 1280×768 px, lower resolution than the competitors, and the effective detail is closer to 720p upscaled, a gap of roughly one resolution tier. The motion strength slider in Gen-3 Alpha is unnumbered; the default produces visible background sliding after 3 seconds. Set the clip length to 5 seconds maximum; at 10 seconds, subject drift is severe enough that the animated element has moved 15-20% of the frame width away from its start position.

The prompt for Gen-3 Alpha should name the motion type and the scene context: "waterfall, water flowing down with mist rising, camera still." The camera control panel offers pan, tilt, zoom, and roll; for a cinemagraph, set all camera controls to zero. Gen-3 Alpha's motion coherence degrades faster than Pika 2.0's; the third failure mode from the research, subject drift, is the primary issue. If the waterfall begins to slide leftward across the frame after 3 seconds, the only fix is to generate a shorter clip or switch to Pika 2.0.

Luma Dream Machine: Not For Cinemagraphs

Luma Dream Machine does not support region masking. Its motion control is limited to whole-frame camera moves: pan, tilt, zoom, roll, and a motion strength slider. A cinemagraph requires selective animation of one region while the frame remains static, but Dream Machine animates the entire frame. You cannot isolate a waterfall without the surrounding rocks and trees moving. If you try to simulate a cinemagraph by reducing motion strength to its minimum, the output is near-static but still shows flicker in the background because the model's latent decoder produces inconsistent frame-to-frame output. Dream Machine is useful for smooth camera moves on a landscape, a slow zoom into a mountain, but it is the wrong tool for the cinemagraph workflow described on this page. Use Pika 2.0 or Kling 1.6 instead.

Open-Weights Models: Stable Video Diffusion, Wan 2.1, LTX-Video

If you need to run generation locally on a consumer GPU, the open-weights models Stable Video Diffusion, Wan 2.1, and LTX-Video are available, but none offers a native region mask. You can approximate a cinemagraph in ComfyUI by generating a full-frame 25-frame clip from SVD, extracting the frames with FFmpeg, compositing each frame over the static source image with a hand-drawn mask in an image editor, and re-encoding. This workflow takes 90-120 minutes per clip and produces seam artefacts unless you apply a Gaussian blur to the mask edge. The result is never as clean as a Pika 2.0 generation because the temporal consistency of the two layers, the animated region and the static background, does not match. SVD produces 1024×576 px output at 25 frames; LTX-Video produces 768×512 px at 121 frames. Wan 2.1 outputs 1280×720 px at 5 seconds. All three require a VRAM of 8-16 GB depending on the model size; Wan 2.1's 14B parameter variant needs 24 GB. The commercial-use licence matters: SVD and LTX-Video prohibit commercial use; Wan 2.1's Apache 2.0 licence permits it.

For the tinkerer who still wants to try: in ComfyUI, load the SVD model from Hugging Face Diffusers, set the motion_bucket_id parameter to 127 (mid-range motion), set fps_id to 6, and generate. Extract the frames, composite with the source image in GIMP, and encode with FFmpeg at 24 fps. The mask boundary artefact will be visible where the two layers meet, you cannot avoid it without a native region mask.

Post-Processing: FFmpeg Loop Encoding For Social Media

A cinemagraph is usually deployed as a short looping video for social media. The native clip from Pika 2.0 or Kling 1.6 is 4-10 seconds, which is too long for a tight loop. Cut it to the exact number of frames that makes the motion look continuous. For a waterfall, identify the frame where the water returns to a similar shape as the first frame, typically frame 12 of 24. Use FFmpeg to extract that segment and encode it as a seamless loop: ffmpeg -i input.mp4 -vf "select='between(n,X,Y)',setpts=N/FRAME_RATE/TB" -an loop_output.mp4, where X is the start frame and Y is the end frame. Set the output frame rate to 24 fps. For platforms that require a maximum resolution of 1080p, the native 1920×1080 px output from Pika 2.0 and Kling 1.6 is already compliant; Runway Gen-3 Alpha's 1280×768 px needs upscaling, which introduces softening.

Aspect-ratio constraints apply: a 9:16 portrait source image for TikTok must be generated in portrait mode. Pika 2.0 and Kling 1.6 accept 1080×1920 px input; Runway Gen-3 Alpha outputs 768×1280 px in portrait mode. Do not crop a landscape generation to portrait, the static background will show a seam artefact at the crop boundary.

Common Questions

Which models support region masking for cinemagraphs?

As of 26 September 2026, three commercial models offer region-mask controls: Pika 2.0 with pixel-level paint masks, Kling 1.6 with Motion Brush and bounding box, and Runway Gen-3 Alpha with a soft bounding box. Luma Dream Machine does not support region masking; its motion control is limited to whole-frame camera moves. No open-weights model ships with native region masks, though Stable Video Diffusion, Wan 2.1, and LTX-Video can be composited manually in ComfyUI with frame-by-frame editing.

How do I prevent mask boundary artefacts in a cinemagraph?

Paint the mask at least 20 pixels inside the moving element's edge, never to the exact boundary. In Pika 2.0, use a tight brush and keep motion strength below 60. In Kling 1.6, use the bounding box as a 30-pixel buffer. If the artefact persists, reduce motion strength or shorten the clip to 4 seconds. For Runway Gen-3 Alpha, the soft bounding box cannot be tightened, so accept softer edges or switch to Pika 2.0.

Can I use a cinemagraph for commercial work?

Pika 2.0 permits commercial use on paid plans with metadata-only watermarking. Kling 1.6 permits commercial use on paid plans. Runway Gen-3 Alpha permits commercial use on paid plans but shows a visible watermark on the free tier. Luma Dream Machine does not permit commercial use under standard terms, you need an enterprise agreement. Stable Video Diffusion and LTX-Video prohibit commercial use under their standard licences; Wan 2.1 permits it under Apache 2.0.

What source images work best for cinemagraphs?

Images with a single element that can plausibly move in isolation: flowing water, steam, hair, smoke, flags, curtains, rotating fans, swaying grass, falling leaves. Avoid faces, reflective surfaces, and thin vertical lines against high-contrast backgrounds. The source image must be at the model's native resolution, 1920×1080 px for Pika 2.0 and Kling 1.6, and in a 16:9 landscape or 9:16 portrait aspect ratio.

How do I handle flicker in the static background?

Preprocess the source image to reduce local contrast, flat, evenly lit backgrounds flicker less. If flicker persists, post-process the clip in FFmpeg: extract frames with <code>ffmpeg -i clip.mp4 frames/frame_%04d.png</code>, apply a temporal median filter to the background region only, then re-encode at 24 fps with H.264. Do not use frame interpolation; it introduces motion where you want stillness.