Removing Compression Artefacts and Noise from Source Images for Image-to-Video Models
How JPEG compression artefacts and noise in a source image become amplified into video flicker during image-to-video generation, and how to clean them beforehand.
Cleaning Compression Artefacts Image-to-Video Source Images
Compression artefacts in your source image become visible flicker and frame-to-frame lighting shifts in your generated video. The mechanism is straightforward: image-to-video models encode the source image into a latent space, then decode across frames through temporal attention layers. JPEG blocking and chroma subsampling create subtle patterns that the latent encoder treats as genuine image features. The temporal attention then propagates these patterns inconsistently across frames, producing flicker and banding that were invisible in the still. You must clean compression artefacts from your image-to-video source images before generation, not after. This page explains why and how.
Why A JPEG's Invisible Flaws Become A Video's Visible Flicker
The latent diffusion architecture that powers models like Stable Video Diffusion and LTX-Video does not 'see' an image the way you do. It compresses the source into a lower-dimensional latent space, adds noise, and then denoises across a sequence of frames. A JPEG saved at quality 85 carries 8x8-pixel blocking artefacts in shadow areas and chroma subsampling in gradients. In a still image viewed at 100% zoom, these are barely perceptible. In a video, the latent encoder treats the blocking edge as a hard boundary, and the temporal attention layer shifts that boundary by one or two pixels between frames. The result is a flickering block that cycles on and off at 24 fps. The same mechanism amplifies chroma subsampling into colour banding that slides across the frame. This is not a model bug. It is a direct consequence of encoding compression damage as if it were real detail.
JPEG Artefact Animation Flicker In AI Video: The Mechanism
Runway Gen-4's model-card notes sensitivity to high-frequency patterns in input images. The same issue affects every latent diffusion model that uses a VAE encoder. When a 24x24 pixel JPEG block exists in a dark area of the source, the VAE sees a sharp luminance step. The temporal attention layer, working across 25 frames, tries to animate that step smoothly. Because the step is an encoding artefact rather than a real edge, the interpolation is inconsistent. One frame places the step at pixel 120, the next at pixel 122, then back to 120. The eye perceives this as a pulsing or crawling artefact. In gradients, the problem is worse: chroma subsampling at 4:2:0 halves colour resolution horizontally and vertically, creating a checkerboard of colour blocks that the model animates independently. The result is banding that shifts colour temperature every second.
Noise Reduction Before Generative Video: Your Cleanup Workflow
Inspect every source image at 100% zoom. Look for 8x8 pixel blocks in shadow areas, stair-step edges on diagonals, and colour stepping in smooth gradients. These are the artefacts that will fail. Apply a deblocking filter before any generation step. FFmpeg is the standard tool: the pp=deblock filter with strength 15-20 removes JPEG blocking without softening real edges. Follow with a denoising pass using nlmeans at strength 5 to suppress chroma noise. Save the result in a lossless format: PNG or TIFF. If your generation tool accepts JPEG input only, save at quality 99 or higher. Do not re-save a JPEG that has already been compressed once. Every re-encode compounds the artefact pattern.
| Model | Provider | Documented Input Sensitivity | Recommended Source Format |
|---|---|---|---|
| Stable Video Diffusion 1.1 | Stability AI | Sensitivity to JPEG blocking in dark areas; noise amplification at motion_bucket_id above 127 | PNG, 1024x576, no compression |
| LTX-Video 2.0 | Lightricks | Fast inference propagates chroma subsampling artefacts; 4:4:4 chroma recommended | PNG or TIFF, 1280x720 minimum |
| Wan 2.1 | Alibaba | Artefact propagation documented in shadow gradients; model card recommends deblocking before input | PNG, 1280x720 |
| Runway Gen-4 | Runway AI | High-frequency pattern sensitivity; camera control amplifies blocking in panning shots | PNG or JPEG quality 99, 1080p |
| Kling 1.6 | Kuaishou | Noise in source reduces start-and-end frame correspondence; clean input improves temporal consistency | JPEG quality 98 or higher, 1080p |
Compression Blocking Generative Video Failure: What Happens Without Cleaning
Generate a clip from a JPEG saved at quality 70. The output will show flicker in every shadow area, banding across the sky or any gradient, and a general loss of source-image fidelity at clip end. The model spends its limited temporal attention budget trying to animate compression artefacts rather than the actual subject. Motion coherence collapses because the artefact patterns compete with real edges. Subject drift accelerates. In extreme cases, the model generates a frame where the artefact block is treated as an object boundary, producing a visible seam that persists across the clip. This is the failure mode that causes newcomers to assume the model is broken. The model is working exactly as trained. The input is the problem.
Image-to-Video Source Image Cleanup Workflow: Step By Step
Start with the original camera file if you have it. A RAW file carries no compression artefacts. Convert to 16-bit TIFF at the model's native resolution. If you only have a JPEG, open it in an editor that can save without re-compression. Apply FFmpeg deblocking filter with strength 15. Check for banding: if gradients show visible steps, add a dithering pass. Save as PNG. For models that require JPEG input, set quality to 99 and disable chroma subsampling if your encoder allows it. Run a test generation on a 2-second clip before committing to a full-length output. Inspect the test at 100% zoom. If flicker or banding persists, increase deblocking strength to 20 and re-run.
Frame-To-Frame Lighting Shift: The Final Symptom Of Uncleaned Input
The most visible failure from uncleaned source images is not flicker in texture but a global lighting shift between frames. The latent encoder interprets compression banding in a gradient as a lighting condition. As the temporal attention layer processes subsequent frames, it varies the banding pattern, which the decoder reads as the light changing. This produces a clip where the scene appears to brighten and darken in a 2-3 second cycle. The fix is identical: clean the source. A clean gradient with no banding gives the temporal attention nothing to vary, and the lighting remains stable. Models like LTX-Video, with their fast inference pipelines, are especially sensitive because they use fewer denoising steps, leaving less room for the model to correct artefact-driven errors.
Temporal Consistency Depends Entirely On Source Purity
Every frame-to-frame lighting shift, every crawling block, every colour banding artefact in your generated video traces back to a compression artefact in the source image. The temporal attention layers in latent diffusion models have no way to distinguish a real edge from an encoding edge. They animate both equally. The models that are most sensitive, Stable Video Diffusion at low motion_bucket_id values, LTX-Video during fast inference, Wan 2.1 at its maximum clip length, are the ones where cleaning matters most. The solo filmmaker who needs b-roll without a crew, the social-media creator animating a product shot, and the open-source tinkerer running models on a consumer GPU all face the same constraint: garbage in, flicker out. The educator researching failure modes and the legal professional clearing generated content for broadcast both need this workflow in their assessment. Someone who wants to generate from a text prompt only, with no source image, should skip this entirely and work with a text-to-video tool. Someone shopping for an all-in-one video editor with generative features built in is on the wrong page. The single thing that most often goes wrong is assuming the still looks clean enough because it looks clean on screen. It does not. Zoom to 100%, find the blocking, and fix it first.
Common Questions
Can I clean the source image after generation instead of before?
No. The artefact pattern is already embedded in every frame. Post-processing filters like temporal denoising can reduce visible flicker but cannot recover the detail lost when the latent encoder treated JPEG blocks as real edges. The source must be clean before inference.
What is the minimum FFmpeg deblocking filter strength I should use?
Start at strength 15 for JPEGs saved at quality 90 or higher. Use strength 20 for quality 80 or lower. Strength above 25 begins to soften real edges, which reduces motion coherence in the output. Test one clip before processing a batch.
Does chroma subsampling matter if I am generating in black and white?
Yes. Chroma subsampling affects colour resolution, but the blocking pattern it creates is luminance-visible in smooth areas. A black-and-white gradient from a 4:2:0 JPEG still shows the 2x2 chroma block grid, which the model animates as texture.
Which models are least sensitive to input compression artefacts?
Models with higher native resolution and more denoising steps, such as Runway Gen-4 at 1080p, show less artefact amplification than models that run at 576x1024 or use aggressive latent compression. No model is immune. The Wan 2.1 model card explicitly recommends deblocking before input.
Can I use an online image compressor to save space before generation?
No. Any re-compression adds artefacts. Store source images in PNG format on disk. If the generation tool requires JPEG upload, re-save at quality 99 with maximum chroma resolution. Do not let a platform's upload workflow re-compress your file.
What about noise from high ISO camera files instead of JPEG compression?
High ISO noise behaves similarly. The latent encoder treats luminance noise as texture and animates it as flicker. Apply a light denoising filter before generation. Leave a small amount of natural grain; heavy denoising creates a plastic look that the model amplifies into unnatural motion.
How do I test whether my source image is clean enough?
Generate a 2-second clip at minimum resolution with no motion prompt. Play it back at 100% zoom. If you see any flicker, banding, or crawling blocks in flat areas, the source needs more cleaning. A clean source produces a static-looking clip with no visible temporal change in flat regions.