How to upscale a low resolution rendering before you turn it into a video

The only copy of your best rendering is a 900-pixel JPEG pulled from an old email, and you want it in a vertical video. Blown up as it is, it goes soft and blocky on a phone, and an AI video model animates that softness instead of the building. Upscaling can rescue it, but every upscaler adds pixels that were never rendered, and some add whole details. This guide covers how much resolution AI video actually needs, what each type of upscaler invents, how Photoshop, Topaz Gigapixel and Magnific differ, and when re-rendering from the model is the better answer.

Key takeaways

• Aim for a long side of at least 1,500 pixels before an image goes into an AI video tool; a 2x upscale gets most low-resolution renderings there without inventing much.
• Every upscaler adds pixels that were never rendered; the difference is whether it smooths between them, predicts plausible texture, or generates new detail on purpose.
• Photoshop's Super Resolution doubles width and height, Topaz Gigapixel offers faithful and generative models side by side, and Magnific's creative upscaler has a creativity control that sets how much it hallucinates.
• If you still have the 3D model, re-render at a higher resolution instead; it is the only way to get real detail rather than predicted detail.
• Compare every upscale with the original at 100% zoom on straight lines, repeating patterns, text and faces, because a video model will treat any invented detail as real and move it.

How do you upscale a low resolution rendering?

Find the largest original you have, clean up compression blocks first, then run a faithful AI upscale at 2x, such as Photoshop's Super Resolution, Topaz Gigapixel's High Fidelity or Low Resolution model, or Magnific's Precision upscaler, until the long side reaches 1,500 pixels or more. Check the result against the original at 100% zoom before you use it. Use a creative or generative upscaler only at low settings and only when you accept invented detail. If the 3D model still exists, re-rendering at a higher resolution beats any upscale.

A clean 2x upscale of a decent source is close to invisible; a 4x upscale of a 400-pixel thumbnail is a new picture that happens to look like yours.

How much resolution AI video actually needs

A vertical film is delivered at 1080 × 1920, so a source image smaller than that in its long dimension has to be enlarged somewhere. If you don't, the video tool does it anyway, and the model reads soft edges and JPEG blocks as part of the scene.

Arch2Video accepts images with a long side of at least 512 pixels and gives the sharpest scenes from 1,500 pixels or more, which is a useful target for any image-to-video tool. Two related checks are worth a minute: export settings, which rendering export settings for Instagram and AI video covers in detail, and whether the image is worth the effort at all, since one strong view does more than a rescued weak one (the six roles a project video needs).

What an upscaler actually invents

All upscaling makes up pixels. The question is how, and how much of it you'd stand behind in front of a client.

Interpolation (bicubic and similar resampling in any image editor) averages neighboring pixels. It invents nothing recognizable, but it can't create sharpness, so a 2x interpolated image looks like a soft version of the original.

Faithful AI upscaling predicts what a sharper version of the same pixels would probably look like, learned from pairs of low- and high-resolution images. Adobe describes Super Resolution as a deep neural network trained on "millions of pairs of low-resolution and high-resolution image patches" (Adobe blog). It sharpens edges and texture you already have, though at small sizes the prediction can still guess wrong.

Generative or creative upscaling adds detail on purpose. Magnific's API documentation says its creative upscaler "can introduce or infer new visual details guided by prompts or context," while its Precision upscaler improves clarity "without inventing new elements" (Magnific docs, as of October 2026). Generative output can look stunning and still be wrong.

What goes wrong is specific to buildings:

• Repeating elements drift. Brick coursing, board spacing, railing balusters and mullion grids get regularized or broken, so the count no longer matches the drawings.
• Text turns to glyphs. Signage, house numbers and a plaque on the wall become plausible nonsense letters.
• Materials change. A smooth plaster wall gains a stone texture, or a matte metal panel picks up a grain it never had.
• People and plants mutate. Small entourage figures get new faces, and trees grow leaves of a species nobody specified.
• Edges get halos. Heavy sharpening leaves bright outlines around rooflines against the sky, which flicker once the frame moves.

Photoshop, Topaz Gigapixel and Magnific compared

All three can rescue a low-resolution rendering; they differ in how much control you get over invention.

Photoshop Super Resolution

Super Resolution "doubles the linear resolution of the photo," giving "twice the width and twice the height of the original photo, or four times the total pixel count," and it works on raw files, JPEGs, PNGs and TIFFs, writing the result as a new DNG (Adobe blog). In Adobe's description you right-click the image in Camera Raw, choose Enhance, tick Super Resolution and press Enhance. That post dates from 2021 and Adobe's current help page could not be loaded for this guide as of October 2026, so the dialog may look different in your version of Camera Raw.

Topaz Gigapixel

Topaz Gigapixel runs as a standalone app or plugin on Mac and Windows and lists separate models: Standard and Standard Max for general work, High Fidelity for "true-to-input details," Low Resolution for degraded sources, Art & CG for computer-generated imagery, Text & Shapes, Recover for restoration, and Redefine plus the Wonder models for creative and generative upscaling (Topaz Labs, as of October 2026). For a rendering, start with High Fidelity or Art & CG, and try Low Resolution on a heavily compressed thumbnail. It is sold as a subscription.

Magnific

Magnific's creative upscaler offers 2x to 16x, a "3d_renders" optimization, and three sliders that run from -10 to 10: creativity, which will "increase or decrease AI's creativity," resemblance, which adjusts "the level of resemblance to the original image," and HDR for definition (Magnific API docs, as of October 2026). For built or designed work, lower creativity and raise resemblance; if the result needs to be faithful, Magnific's Precision upscaler is the mode built for it. Magnific's own homepage could not be loaded for this guide, so check current plans there.

Re-rendering vs upscaling: which to choose

If the model exists, re-render. A new render at 2,000 pixels or more gives real detail, the right mullion count and a clean sky, and it costs render time rather than accuracy.

Upscale when the model is gone or was never yours:

• An old project whose files were archived in a format you no longer open (see turning old project photos into a video for the archive side).
• A rendering a visualization studio delivered at screen size.
• An image generated in Midjourney or ChatGPT at its native size; the Midjourney-to-video guide covers its own upscale button.
• A client's copy that went through a messaging app, after you've asked for the original and come up empty.

Ask for the original first: a file sent by email or shared drive is usually larger than a copy saved from a chat app or a social post.

A step-by-step upscale workflow that keeps the design intact

1. Collect the largest version. Check the visualizer's delivery folder, the original email, the PDF presentation it came from (re-export the page at high resolution) and the website's media library.
2. Remove compression first. If you see 8 × 8 pixel blocks in the sky or banding in gradients, run a JPEG artifact removal or light denoise before enlarging, or the upscaler will sharpen the blocks.
3. Upscale 2x with a faithful model. Check the result, and only then decide whether you need another step. Two moderate passes are easier to judge than one big jump.
4. Compare at 100%. Put the original (scaled up with plain interpolation) and the upscale in two layers and toggle between them. Look at rooflines, window grids, railings, text, faces and plants.
5. Paint back invented detail. If a creative pass changed a facade, mask it back to the faithful version.
6. Export in sRGB. Save a high-quality JPG or a PNG, keep each file under your video tool's limit (30 MB per image for Arch2Video), and name it so you know it was upscaled.

Then review the finished video the same way you checked the stills: the AI video quality checklist covers geometry, text and materials in motion.

When not to upscale

Some images aren't worth saving. A tiny thumbnail is better replaced by a fresh render, a new photograph of the built work, or a different image that fills the same role. Floor plans and drawings with fine linework and dimensions are a special case: an AI upscaler can smear or rewrite numbers, so export the drawing again from the CAD file or PDF at a higher resolution rather than upscaling a screenshot. And if a project will be presented to a client, a planning board or a publication, don't let an upscaler add detail that isn't in the design; it reads as a promise.

Frequently asked questions

Will an upscaled rendering look fake in an AI video?

Usually not, if you stay at 2x with a faithful model and check the result. The fakeness comes from generative passes at high settings, which add texture the video model then animates as if it were real: stone that crawls, foliage that boils, edges that shimmer. A modest, clean upscale gives the model sharp edges to hold on to.

How many pixels should a rendering be for an AI video?

For a vertical 1080 × 1920 film, aim for a long side of 1,500 pixels or more; 2,000 or more leaves room if the tool recomposes the frame. Arch2Video accepts a long side from 512 pixels but gives its sharpest scenes at 1,500 and above. Below about 750 pixels, even a good upscale is guessing at details, so look for a larger original first.

Is it OK to upscale a rendering made by a visualization studio?

Technically yes; contractually, check first. Visualization agreements often license specific uses, and a modified file used in a video may be a new use. Ask the studio for a higher-resolution export, which often makes upscaling unnecessary, and see architectural photo licensing for video for the questions to ask. This is general information, not legal advice.

Should I upscale before or after I crop the image to vertical?

Upscale the whole image first, then crop if you need to. Upscalers use surrounding context, so working on the full frame gives better edges, and you keep the option of a different crop later. For an AI video tool that recomposes images itself, such as Arch2Video, you usually don't need to crop at all; just keep the important part of the building away from the very edges of the frame.