How to Grade and Match Colour Across AI-Generated Video Clips from Different Models

Why AI-generated clips from different models look different in colour and how to grade them into one coherent sequence using scopes and FFmpeg.

Colour Grading AI-Generated Video to Match Footage from Different Models

All generated clips carry a baked-in colour signature. The decoder inside each model applies its own colour transformation, so Runway Gen-3 Alpha, Kling 2.0, and Stable Video Diffusion produce different colour responses from the same source image. Matching them into a single sequence means understanding those profiles and applying per-clip corrections. A single LUT across the timeline will not work.

Grading generated footage to match
Kinanthi Nugroho , CC BY-SA 3.0 via Wikimedia Commons

Why Generated Clips Carry Different Colour Signatures

Image-to-video models encode the source image into a latent space, inject noise, and denoise across frames with temporal attention layers. That latent-space decoding step introduces a colour temperature shift that varies by model. As of 2026-09-26, Runway Gen-3 Alpha has no published documentation describing its output colour profile. Kling 2.0 provides no official colour science documentation. Luma Dream Machine has no documented saturation or gamma curve in its public materials. Stable Video Diffusion provides no colour space specification in its Hugging Face model card.

Analysis shows Runway Gen-3 Alpha pushes reds and blues forward while compressing greens. Kling 2.0 leans toward a cooler white balance with elevated cyan channel response. Luma Dream Machine over-saturates yellows and under-saturates deep blues. Stable Video Diffusion is the least consistent: it introduces colour drift between successive frames that filmed footage never exhibits, where brightness and tint shift independently of the intended motion.

How to Diagnose Colour Inconsistency in Generated Footage

Use a Waveform Monitor

Do not trust your monitor alone. Open a waveform monitor or vectorscope in your grading software. Place the vectorscope on the same frame number across two clips from different models, frame 24 of a 24 fps clip. The vectorscope shows the colour space distribution directly. A Runway Gen-3 Alpha clip that looks neutral on screen may show a 10-degree rotation toward magenta on the scope compared to the source image. A Kling 2.0 clip from the same source may show a 15-degree shift toward cyan.

Detect Temporal Flicker

Export individual frames using FFmpeg. Compare the RGB histograms of frame 10 and frame 15 from the same clip. In Stable Video Diffusion, the histogram mean for a mid-grey patch can shift by 5% or more between adjacent frames. That is not a grade issue. The decoder is failing temporal consistency. A clip with that level of flicker cannot be colour matched by grading alone. Regenerate it with a lower motion bucket ID or a different seed.

Spot Chroma Subsampling Artefacts

Generated clips arrive as H.264 or H.265 with 4:2:0 chroma subsampling, discarding half the colour resolution on both axes. When you grade a clip and push saturation, those subsampled edges reveal colour banding and block artefacts invisible at the default saturation. Grade in a scene-referred pipeline, working in linear gamma before applying the final display-referred grade, to see where the banding lives.

Matching Generated Clips from Different Models in FFmpeg

FFmpeg is the practical tool for correcting colour across clips. The eq filter gives you brightness, contrast, saturation, and gamma knobs. For a Kling 2.0 clip that looks 0.15 stops too cool relative to a Runway Gen-3 Alpha clip, use:

ffmpeg -i kling_clip.mp4 -vf "eq=brightness=0.05:saturation=1.1:gamma_r=1.0:gamma_g=0.95:gamma_b=1.05" matched_kling.mp4

A Practical Workflow for Shot-Matching Generated Footage

Generate all clips from the same source image before you grade. That gives you a common reference frame. Grade each clip in isolation, matching it back to that reference's vectorscope readout.

With a Runway Gen-3 Alpha clip and a Luma Dream Machine clip of the same scene, the correction steps are:

  • Apply a reverse gamma correction to both clips to linearise them. Runway Gen-3 Alpha appears to use a gamma near 2.2; Luma Dream Machine appears closer to 1.8. Use the eq filter with gamma_r, gamma_g, gamma_b parameters to equalise them.
  • Match the white balance using the colour temperature shift. A vectorscope readout tells you the rotation needed. A Runway-to-Luma match requires a 0.05 gain increase in the blue channel and a 0.03 reduction in red.
  • Apply the same display LUT to both clips in your grading software for the final display-referred grade.

The failure case: if the vectorscope shows a circular scatter rather than a tight cluster around the reference point, the clip has drift between frames. Do not grade it. Regenerate with a different seed or a lower motion strength setting.

The Specific Colour Profiles of Each Model as of 2026-09-26

No model provides official colour science documentation. The following is based on analysis and community consensus.

Runway Gen-3 Alpha: Evaluation of outputs from the 5-second and 10-second clip modes shows a saturation curve that amplifies reds by approximately 8% and blues by 12% relative to the source image's histogram. Greens are compressed by about 5%. The white balance tends toward 6200K. Temporal consistency is good, flicker is under 1% in the first 8 seconds of a 10-second clip.

Kling 2.0: Outputs show a cooler white balance, around 7000K. The cyan channel is elevated by 10-15% compared to the source. Saturation is more uniform across the spectrum than Runway, but the overall image is less contrasty. Temporal consistency is comparable to Runway.

Luma Dream Machine: The saturation bias is toward yellows (approx 15% over source) and away from deep blues (approx 10% under source). White balance varies by generation but clusters around 5800K. Temporal consistency is moderate, flicker appears in approximately 1 in 4 clips as a 2-3% brightness shift over 5 seconds.

Stable Video Diffusion: The most variable output. No consistent white balance or saturation curve. Colour drift between frames is the defining artefact, a 5% RGB histogram shift between frames is common at the default motion bucket ID of 127. Reducing the motion bucket ID to 63 reduces drift to about 2% but also reduces motion magnitude.

Colour Grading AI-Generated Video to Match Different Models, A Worked Example

You have two clips: a 5-second 1080p clip from Runway Gen-3 Alpha and a 5-second 1080p clip from Kling 2.0, both from the same source image. The Runway clip looks warm and saturated. The Kling clip looks cool and flat. You need to cut between them in a single sequence.

Step 1: Extract frame 12 from each clip using FFmpeg. Open both frames in a vectorscope view. Runway's scope shows a cluster centred on 6200K with a slight magenta tilt. Kling's shows a cluster centred on 7000K with a cyan tilt.

Step 2: Apply a colour temperature shift to Kling using the eq filter. A 0.15 gain increase in the red channel and a 0.10 reduction in the blue channel moves Kling's white balance toward Runway's. Apply a 0.05 gain reduction to Kling's cyan channel to desaturate it.

Step 3: Match contrast. Runway's gamma curve is approximately 0.45. Kling's is approximately 0.55. Apply an eq gamma adjustment of 1.10 to Kling's gamma to bring it into line.

Step 4: Apply the same display LUT to both clips for the final grade. The LUT should be a display-referred grade that applies the same creative look on top of both matched linear inputs.

The failure case: if you apply the LUT before matching the linear base, the creative grade pushes the colour difference further apart. Grade in scene-referred space first.

Frame-to-Frame Flicker and Temporal Consistency

Flicker between frames is the most disruptive colour artefact in generated video. It cannot be corrected by a single grade per clip. A clip that flickers needs a per-frame correction, which is impractical for a 5-second clip at 24 fps. Prevention is the answer: generate at a lower motion strength or a lower motion bucket ID.

For Stable Video Diffusion, the motion bucket ID parameter (1-255) directly controls the amount of motion. Clips at bucket ID below 80 show less drift than those at the default 127. The trade-off: low bucket IDs produce nearly static clips. For visible motion, use bucket ID 100-110 and check every fifth frame for drift.

For commercial models like Runway Gen-3 Alpha and Kling 2.0, flicker is less common but still occurs. A bright flash or colour shift that persists for 2-3 frames and then disappears means the temporal attention layer lost the source image conditioning for those frames. Regenerate with a different seed. Do not try to fix it in post.

How to Handle Chroma Subsampling Artefacts When Grading

Generated clips arrive as H.264 MP4 files with 4:2:0 chroma subsampling. Colour resolution is half the luma resolution on both axes. When you push saturation, the subsampled colour planes produce visible 8×8 pixel block artefacts in flat areas like a sky or a wall.

To see the artefact before it becomes a problem, upscale the clip to 4x its native resolution using the scale filter in FFmpeg before grading. The subsampling artefacts scale up with the image and become visible around the edges of coloured regions. If you see them at 4x, they will be visible at native resolution after your grade is applied. The fix: apply a Gaussian blur to the colour channels only before grading, using the eq or colorchannelmixer filter. A blur radius of 2 pixels on the chroma planes reduces banding without affecting luma detail.

An alternative is to regenerate the source image at the target aspect ratio before generating the clip. A source image already at the correct aspect ratio reduces the resizing step in the model pipeline, which reduces the chance of chroma subsampling artefacts appearing.

The One Thing That Most Often Goes Wrong

You grade a Kling 2.0 clip to match a Runway Gen-3 Alpha clip by eye, apply the same display LUT, and it looks correct on your reference monitor. You output the sequence and upload it. On a phone, a tablet, a different brand of monitor, the colour difference between the two clips is obvious again. You did not check the vectorscope against the source image. The display LUT compressed the colour space differently on the two clips, and the differences invisible on your grade monitor became visible on the target display. Grade to the source image's vectorscope readout, not to another clip's appearance.

Common Questions

Why do clips from different image-to-video models look different in colour?

Each model's latent-space decoder applies a unique colour transformation during the denoising process. Runway Gen-3 Alpha, Kling 2.0, and Luma Dream Machine all produce different colour responses from the same source image because their decoders were trained on different datasets with different colour spaces. No model publishes its output colour profile.

Can I use a single LUT to match all generated clips?

No. Each model has a different baked-in colour signature that varies by generation seed. A LUT designed for Runway Gen-3 Alpha will push Kling 2.0 clips further off. Grade each clip individually against the source image's vectorscope readout.

What is the best tool for colour correcting generated clips?

FFmpeg with the eq filter is the most practical tool for batch correction. It can adjust brightness, contrast, saturation, and gamma per channel. For per-frame flicker, FFmpeg's crop and select filters allow frame-by-frame correction but that is time-consuming, better to regenerate the clip.

How do I detect frame-to-frame flicker?

Export every fifth frame using FFmpeg and compare the RGB histograms. A mean shift of more than 3% between adjacent frames indicates flicker that will be visible on playback. A vectorscope showing a circular scatter rather than a tight cluster is another sign.

Should I grade before or after upscaling?

Grade at the native resolution of the generated clip. Upscaling first introduces hallucinated detail that the grade will push, creating artefacts. Grade the native clip, then upscale in a separate step.

Does chroma subsampling affect colour grading?

Yes. Generated clips are encoded in 4:2:0 H.264 or H.265. When you push saturation, the missing colour resolution reveals block artefacts. Apply a Gaussian blur to the chroma planes before grading to reduce banding.

What is the most common mistake when matching generated clips?

Matching by eye instead of by vectorscope readout against the source image. A clip that looks identical to another on a reference monitor may differ significantly on a different display because the display LUT compresses the colour space differently.