Runway Gen-4 Image-to-Video Generation: Director-Mode Camera Controls and the Credit System

You have a still image and you need a moving shot from it. The model you used before only gave you a motion-strength slider that made your subject wobble and the background warp. The Runway Gen-4 image-to-video camera controls solve exactly that: they give you pan, tilt, zoom, roll, dolly, and crane axes, replacing the single slider with a director-mode panel that treats the virtual camera as a controllable instrument. This page documents every axis, what the credit system costs you in practice, and how this model differs from Gen-2.

The defining feature of Runway Gen-4 for image-to-video generation, introduced on 2025-05-01 per Runway's blog and documentation at docs.runwayml.com, is the Camera Control panel. It exposes six discrete axes: pan (camera rotates left or right on its vertical axis), tilt (camera rotates up or down), zoom (focal length change, not camera movement), roll (camera rotates around its forward axis, like a Dutch angle), dolly (camera tracks forward or backward on a linear path), and crane (camera moves vertically up or down). Each axis has its own strength slider, independent of the Motion Brush system that animates objects within the frame. This is the core answer to what camera-motion controls Runway Gen-4 exposes for image-to-video generation, and it is the single most important difference from Gen-2.

Gen-2, documented in Runway's help pages as of 2024, had only a motion-strength slider. That single slider controlled everything: how much the subject moved, how much the background moved, and how fast. You could not tell the model to dolly past a stationary person while zooming in. Gen-4's director mode separates the virtual camera's movement from the subject's movement. The Motion Brush, first added in Gen-3 Alpha and retained in Gen-4, lets you paint motion vectors onto specific regions of the source image. The Camera Control panel moves the viewpoint itself. The confusion with Gen-2 is resolved by noting that Gen-4 has two independent control systems, subject animation and camera movement, where Gen-2 had one.

Pan rotates the camera left or right on its vertical axis. It works best when the source image has horizontal detail, a landscape, a line of people, a car on a road, because the model has texture to slide across the frame. A pan on a plain wall produces a smear, not a camera move. Zoom changes the focal length, pulling the scene toward or away from the viewer. It does not change perspective; the camera stays in place. Roll rotates the image around the camera's forward axis, producing a Dutch-angle effect. Gen-4 handles roll more cleanly than Gen-2, which introduced horizon tilt changes mid-clip.

Dolly and Crane: The Linear Axes

Dolly moves the camera forward or backward on a straight path. It is the most physically demanding axis for the model because the perspective of the image must expand or contract as the camera approaches or retreats. Subjects drift sideways under dolly when the source image lacks depth cues. A close-up of a face with a flat background will cause the subject to slide relative to the frame. Crane moves the camera vertically. It pairs naturally with dolly for a camera-arc shot, but using both axes at high strength simultaneously increases the risk of motion collapse, where the model generates a nearly static clip because it cannot resolve the combined movement.

Motion Brush vs. Camera Control: When to Use Which

Use the Motion Brush to make a flower sway, a car's wheels spin, or a character's hair move. Use Camera Control to dolly toward a building, pan across a crowd, or zoom into a product shot. They work together: you can set a camera pan at medium strength and then paint a region for the subject to walk against the pan direction. Source-image fidelity at the clip end depends on keeping camera-motion strength moderate (under 60 percent on the slider for each axis). The model's attention mechanism loses track of the source image when the virtual camera moves too far from the original framing. At the 10-second clip length, subject drift becomes visible on 40 percent of generations even with zero camera movement, per independent testing reported in the dossier's claimed-vs-real analysis.

Gen-3 Alpha introduced the Camera Control panel and Motion Brush. Gen-4, released on 2025-05-01 per Runway's product announcements, does not add new axes. The change is in motion coherence and artefact reduction. Runway's documentation date-checked 2026-09-26 on docs.runwayml.com states that Gen-4 improves temporal consistency on camera moves by 35 percent relative to Gen-3 Alpha, though no benchmark scores or evaluation methodology are published to support the claim. No standalone model card for Gen-4 exists on runwayml.com or docs.runwayml.com as of 2026-09-26.

The practical difference is that Gen-4 shows less background sliding during dolly and crane moves. Background sliding, where the background moves independently of the foreground subject as if on a separate layer, was the most common failure mode in Gen-3 Alpha during camera movements. Gen-4 reduces but does not eliminate it. If the source image has a clear depth boundary (a person in front of a distant horizon), the model treats them as separate layers; if the image is a single flat texture, background sliding persists.

Limb Mutation in Gen-4

Limb mutation, arms, legs, or fingers warping, multiplying, or detaching during motion, is also less frequent in Gen-4 but still present when the source image contains people in non-standard poses or when the camera movement passes through the subject. The model lacks a skeletal model of the subject; it predicts pixel motion from patterns learned during training. That is why a hand held up in a source image can turn into three hands over the clip duration.

Runway uses a credit-based pricing system for Gen-4 image-to-video generation, documented in its Terms of Service effective 2025-11-12 and on its pricing page at app.runwayml.com. Users purchase credit packs or subscribe to a monthly plan that allocates a fixed number of credits per billing cycle. One generation consumes one or more credits depending on output resolution and clip length. 720p 5-second clips consume fewer credits than 1080p 10-second clips. The number of credits per generation is not published in terms-of-service text; it appears only on the logged-in pricing page, which falls under the Perishable category and is not quoted here.

The commercial-use terms are straightforward. Runway's Terms of Service state that users retain all ownership of outputs under the section titled 'User Content'. Runway does not use user inputs or outputs to train models without explicit opt-in, per the same section, effective 2025-11-12. Commercial use is permitted subject to the Acceptable Use Policy effective 2025-11-12. Runway embeds C2PA Content Credentials metadata in Gen-4 outputs, announced 2025-05-01. This is an invisible metadata watermark conforming to the C2PA specification version 2.1. There is no visible watermark on Gen-4 output files. For platform compliance, C2PA metadata is read by Adobe Content Credentials and by YouTube's synthetic-content labelling system, but not by TikTok's or Instagram's detection systems as of 2026-09-26, per platform documentation cited in the dossier. That means a clip generated in Gen-4 may be flagged by TikTok's video-analysis algorithm even though its metadata watermark is intact, leaving the uploader exposed to a platform labelling requirement they cannot satisfy with the existing C2PA data.

Credit Consumption and Throttling

The dossier's claimed-vs-real analysis flags the gap between 'unlimited generations' on subscription tiers and actual fair-use limits. Runway's Terms of Service do not guarantee a specific number of generations per credit allocation; degraded service or throttling may apply to high-volume users. The documentation advises that API queue times can multiply total turnaround by 2-10 times during peak usage, and this wait time is excluded from API-documented inference speed figures. If you need predictable generation throughput, the API with a reserved-tier plan is the only reliable path, and even then the 25 MB maximum file size limit for API requests applies to the source image, not the output.

The most common failure mode with camera-control use is motion collapse, where the model generates a nearly static clip despite a prompt requesting significant camera movement. This is caused by the model's bias toward low-motion outputs, which have fewer visible artefacts during training. It occurs most often when multiple axis sliders are set above 70 percent simultaneously. Reduce strength on two of the axes and keep the primary movement on one.

Subject Drift

Subject drift is the second most common failure. The person or object in the source image gradually transforms into something else over the clip duration. It is caused by the model's temporal attention losing the conditioning signal from the source image. Counter-intuitively, zero camera movement does not prevent subject drift; it is a function of clip length, not of camera motion. At 10 seconds, approximately 40 percent of generations show visible subject drift even with no camera control active. The only way to avoid it is to generate shorter clips, 5 seconds, and concatenate them in post-production, though that introduces seam artefacts at the join because Gen-4 does not support start-and-end-frame conditioning.

Background Sliding and Limb Mutation

Background sliding occurs when the background moves independently of the foreground subject during camera moves. It is worst on dolly and crane moves with images that have no depth separation. A source image of a person standing against a plain wall will cause the wall to slide left while the person stays centred. The cause is the model failing to infer a consistent 3D scene structure from the single image. There is no fix; the workaround is to use source images with natural depth boundaries.

Limb mutation, arms, legs, or fingers warping, multiplying, or detaching, occurs when the camera movement passes near or through a subject's limbs. Gen-4 reduces the frequency of limb mutation relative to Gen-3 Alpha, but the cause remains the same: the model has no skeletal or physical model of the subject. It predicts pixel motion from learned patterns. A hand held up in the source image is statistically predicted to become multiple hands over time because the model's training data includes many images of moving hands.

Content Moderation and Rejection

Runway's Acceptable Use Policy effective 2025-11-12 defines prohibited content. Content moderation strictness on Gen-4 is high. Benign source images are sometimes rejected by the safety classifier. The dossier reports content-moderation false positives where skin tones, certain object shapes, or patterns that resemble prohibited content trigger rejection. If your source image is rejected, choose a different image. There is no appeal process documented in the Acceptable Use Policy. If you plan to animate a real person, obtain their explicit consent before uploading their image; the platform's terms do not waive this legal obligation.

Generate a 5-second clip with one camera-control axis at medium strength before you try multiple axes or the full 10-second length. Test the source image with zero camera movement first to check for subject drift and limb mutation. If the source image produces a clean 5-second static clip, add one axis. Pan is the safest start, at 40 percent strength. Review the output for background sliding and subject drift before adding a second axis. Most users who get frustrated with Gen-4's camera controls try to use all six axes at high strength on a complex source image in a 10-second generation. That combination produces motion collapse or severe drift every time. The tool is capable of smooth camera moves, but only within the structural limits of latent diffusion: the source image must be chosen for the move, not the move applied to any image.