When to Upscale a Source Image Before Feeding It to an Image-to-Video Model

Upscaling a source image before image-to-video generation solves one problem and creates another. The problem it solves: feeding an image below the model's native resolution forces the model's internal pre-processing to upscale, introducing softening and hallucinated detail that the temporal attention layers then animate as flicker. The problem it creates: upscaling externally with an AI upscaler before generation can introduce synthetic texture patterns that the video model interprets as real detail and exaggerates into warping. The decision depends on the model's native resolution and the source image's original quality. For models with high native resolution like Kling 1.6, upscaling a low-res source is often necessary. For models with lower effective resolution like Stable Video Diffusion, external upscaling may add artefacts the model cannot resolve. This page gives the decision framework, not a blanket recommendation.

Risk one: the model's internal upscaling. When you feed a source image smaller than the model's native input resolution, the model must upscale it before it can generate motion. This internal upscaling, usually a fast, low-quality method like bicubic resampling, softens the image. The temporal attention layers then animate that softened detail as flicker, producing a jittery, unstable output. The softer the input, the more the model hallucinates detail to fill the gaps, and those hallucinated details move on their own across frames.

Risk two: external AI upscalers introduce false texture. Tools like Real-ESRGAN or Topaz Gigapixel AI use GAN-based architectures that hallucinate detail, they invent texture that looks plausible in a still frame but does not exist in the original. When you feed that invented texture to a video generation model, the model treats it as real and animates it. The result: warping, morphing, and a 12-18% higher rate of hallucinated motion compared to using a non-AI upscaler like Lanczos resampling, according to a SIGGRAPH Asia 2024 study.

Make your decision in three steps. First, identify the model's native generation resolution. Second, measure the effective resolution of your source image, not its pixel dimensions, but the level of real detail it contains. A 4K JPEG with heavy JPEG blocking and compression artefacts has lower effective resolution than a clean 720p PNG. Third, apply the upscaling rule: upscale to match the model's native generation resolution on the shortest side, not to an arbitrary high resolution. Exceeding native generation resolution triggers internal downscaling, negating the upscale.

Aspect-Ratio Distortion Trap

If your source image's aspect ratio does not match the model's training aspect ratios, pad with black bars or use outpainting rather than stretching. Stretching causes aspect-ratio distortion that the model animates as unnatural compression or expansion. Crop the output after generation.

Upscaling Method Matters

Nearest-neighbour upscaling before generation increases aliasing artefacts in the output motion. AI upscalers that hallucinate detail introduce false textures. The research recommends bicubic or Lanczos resampling when you do not trust the AI-upscaled detail. If you must use an AI upscaler, choose one with a non-commercial licence like Real-ESRGAN (BSD 3-Clause) or SwinIR (MIT) over proprietary tools.

Model	Minimum Input (px)	Native Generation Resolution (px)	Upscaling Recommendation
Runway Gen-3 Alpha	768×768	1280×768 (landscape), 768×1280 (portrait)	Upscale to at least native generation resolution before upload
Kling 1.6	384×384	960×540, 960×960, 540×960	Match source resolution to target output resolution before upload
Luma Dream Machine	576×576	1360×752 (16:9), 1080×1080 (1:1), 752×1360 (9:16)	Upscale to at least 1080p on shortest side before upload
Pika 2.0	512×512	1088×640 (16:9), 1088×1088 (1:1), 640×1088 (9:16)	Upscale to at least 1080p on longest side before upload
Veo 3	256×256	1280×720 (16:9), 1280×1280 (1:1), 720×1280 (9:16)	Upscale to at least 720p on shortest side before upload
Sora	480×480	1920×1080 (16:9), 1080×1920 (9:16), 1080×1080 (1:1)	Upscale to at least 1080p on shortest side before upload
Stable Video Diffusion (SVD)	576×1024	576×1024	Resize source to exactly 576×1024 px before inference
Wan 2.1	480×480	832×480 (16:9), 480×832 (9:16), 480×480 (1:1)	Upscale to at least 480p on shortest side before upload
LTX-Video	256×256	768×512 (3:2), 512×768 (2:3), 512×512 (1:1)	Upscale to at least 512p on shortest side before upload
Seedance 1.0	512×512	1280×720 (16:9), 720×1280 (9:16), 1024×1024 (1:1)	Upscale to at least 720p on shortest side before upload
Hailuo I2V-01	480×480	1280×720 (16:9), 720×1280 (9:16), 1080×1080 (1:1)	Upscale to at least 720p on shortest side before upload

Each model has a claimed resolution and a real effective resolution. The table above gives the claimed native generation resolution, but the effective resolution, the level of detail the model can actually resolve, is typically one tier lower due to latent-space compression. A model that outputs 1080p files often has effective detail closer to 720p upscaled. This gap matters because upscaling a source to the claimed resolution but not the effective resolution wastes the extra pixels.

Wan 2.1 Resolution Tiers and LTX-Video Speed-Optimised Architecture

Wan 2.1 offers multiple parameter-size options (1.3B and 14B). The 1.3B model has lower effective resolution than the 14B version. LTX-Video uses a speed-optimised architecture that trades resolution for inference speed; its effective resolution is closer to 480p despite outputting 768×512. For both models, upscaling beyond the effective resolution adds noise without benefit.

Stable Video Diffusion Input Size: The Zero-Upscaling Case

Stable Video Diffusion is the exception. Its minimum input resolution (576×1024) matches its native generation resolution exactly. There is no internal upscaling step. The SVD documentation explicitly advises resizing the source to exactly 576×1024 px before inference. Any upscaling beyond that triggers internal downscaling, negating the effort. This is the cleanest pipeline for testing whether pre-generation upscaling helps or hurts: with SVD, it almost always hurts.

Skip upscaling in three scenarios. First, when the source image is already at or above the model's effective resolution. A 1080p source for Kling 1.6 (native resolution 960×540) needs no upscaling and will be downscaled internally. Second, when the source image contains compression artefacts or JPEG blocking. Upscaling those artefacts makes them larger and the video model animates the block boundaries as flicker. Third, when using a model with a wide gap between claimed and effective resolution, like LTX-Video or Wan 2.1's 1.3B parameter variant. The extra pixels from upscaling are lost in the latent-space compression pipeline.

Published Evidence for the Decision

A CVPR 2024 workshop paper found that upscaling a source from 512p to 720p before SVD inference reduced the flicker artefact rate by 34%. A separate arXiv 2024 study found that upscaling from 512p to 1080p before Runway Gen-2 inference reduced temporal inconsistency by 22%. But the SIGGRAPH Asia 2024 poster showed that AI upscalers with GAN artefacts increased hallucinated motion by 12-18% compared to Lanczos resampling. The lesson: measured upscaling with a non-GAN method helps; AI upscaling with hallucinated detail hurts.

Build your pre-processing pipeline in this order: check the source image's effective resolution, compare it to the model's effective resolution (not the claimed resolution), choose an upscaling method (bicubic or Lanczos over GAN-based AI upscalers), upscale to the model's native generation resolution on the shortest side, and pad for aspect-ratio mismatch. Do not exceed the native generation resolution.

Runway Gen-3 Alpha Source Guidelines

Runway's documentation (dated June 2025) recommends upscaling to at least 1280×768 for landscape or 768×1280 for portrait before upload. The minimum input is 768×768. The documentation cites Topaz Gigapixel AI and Adobe Super Resolution as upscaling tools, both proprietary. For a free alternative, use Real-ESRGAN with the general model, not the anime model, unless your source is anime.

Kling 1.6 High Native Resolution Advantage

Kling 1.6 has a high native resolution (960×540, 960×960, or 540×960) and a low minimum input (384×384). This 2.5× gap means upscaling is almost always necessary for low-res sources. Kling's documentation (July 2025) recommends matching source resolution to target output resolution. The cited upscaling tools are Real-ESRGAN and Topaz Gigapixel AI.

Stable Video Diffusion Input Size: The Tight Constraint

SVD is the most restrictive model for input resolution. Its native generation resolution is exactly 576×1024. The documentation (November 2023) advises resizing to exactly those dimensions. Real-ESRGAN is the cited upscaling tool, with separate 4× anime and general models. If your source is smaller than 576×1024, resize with bicubic, not nearest-neighbour.

The single most common error is using an AI upscaler that introduces synthetic texture, feeding that hallucinated detail to the video model, and watching the output warp and morph as the temporal attention layers animate the false texture. The published studies are clear: GAN-based upscalers increase hallucinated motion by 12-18% over Lanczos resampling. The fix is to use bicubic or Lanczos for the final upscale, not an AI upscaler, unless you are certain the source image has no GAN artefacts and you need the extra resolution for a model like Kling 1.6 with a high native resolution.

The second most common error is upscaling to an arbitrary high resolution instead of to the model's native generation resolution. This triggers internal downscaling, wasting the upscale and adding an extra compression step. Always check the model's documentation for the exact native resolution and upscale to that value on the shortest side.