Upscaling: Magnific and Topaz

Upscaling with Magnific and Topaz Video AI - when to upscale, how creative hallucination works, and the step-by-step workflow for images and video.

TL;DR: AI generators output at 720p or 1024x1024 because that is what fits in a GPU budget. Magnific and Topaz Video AI are the two tools that close the gap between raw generation and publish-ready quality - Magnific for still images and short video clips (with an option to hallucinate richer detail, not just stretch pixels), and Topaz Video AI for full-length footage where temporal consistency and model selection matter most.

Why you need to upscale at all

Most AI image generators output at 1024x1024 pixels. Most AI video generators cap at 720p or 1080p. Those resolutions work fine on a phone screen, but the moment you crop, zoom, print, or publish to a platform that sharpens thumbnails aggressively, the lack of pixel density shows.

Standard resizing (bicubic, Lanczos) just interpolates between existing pixels. The edges get smoother, but no new detail appears. The image looks stretched, not sharp. AI upscaling solves a different problem: it infers what the missing detail should look like based on everything it knows about the content type, then synthesizes it.

Common triggers for upscaling:

• Publishing AI stills to a feed or marketplace that displays at 2x retina resolution
• Printing or using an image in a thumbnail larger than 1080px wide
• Exporting a generated video to YouTube, Instagram Reels, or a TV/broadcast context
• Repurposing archival or compressed footage that looks soft next to native 4K content
• Delivering client work where the brief specifies 4K or print-ready resolution

Magnific: upscaling with hallucinated detail

Magnific (formerly Freepik, rebranded in April 2026) does something conventional upscalers cannot: it invents plausible detail. Skin pores, fabric weave, brick texture, leaf veins - these do not exist in the source file. Magnific's generative engine reads the image context and synthesizes what should be there if the image had been captured at higher resolution. The company calls this "generative hallucination."

The tool offers two modes:

• Creative mode - the AI reimagines and adds detail, guided by sliders and an optional text prompt. Scale up to 16x. Best for AI-generated content, illustrations, or any source where inventing detail is acceptable.
• Precision mode - faithful enlargement with no creative reinterpretation. Best for real photography or professional footage where accuracy matters. Important caveat: it faithfully upscales artifacts too, so your source needs to be clean.

The three sliders that matter

In Creative mode, three sliders (-10 to +10) control the result:

• Creativity - how aggressively the AI invents new detail. At 0 it behaves like a standard upscaler. At higher values it adds wrinkles, stitching, grain, and texture that was never in the source. For realistic content, stay in the 2-5 range. For artistic work, higher values create legitimate creative output.
• Resemblance - how closely the output stays true to your original. Higher resemblance preserves faces, brand elements, and identifiable features. Lower values give the AI more freedom to drift. When upscaling portraits or characters, keep this high (7-9) to avoid identity shift.
• HDR - boosts dynamic range contrast. Useful for adding punch to flat AI-generated images. Use sparingly on already-contrasty sources.

The text prompt is your secret weapon

Creative mode accepts a text prompt alongside the image. This is more useful than it sounds. If you are upscaling a portrait, prompting "highly detailed skin texture, natural lighting, photography" tells the model what kind of micro-detail to synthesize. For architecture: "sharp stone texture, mortar detail, natural shadow". The prompt does not redraw the image - it shapes what the hallucinated detail looks like.

Magnific also ships 13 content presets (Portrait, Art, 3D Renders, Photography, Nature, and more) that pre-tune the enhancement strategy for each content type. Start with the matching preset before touching sliders.

Magnific video upscaling

Magnific's video upscaler applies the same Creative/Precision split to video clips. The key technical difference from image upscaling is temporal awareness: the model analyzes motion, texture continuity, and subject consistency across frames before generating new detail. Processing each frame in isolation causes flickering (temporal incoherence). The video mode locks the hallucination seed across the sequence so that invented textures stay consistent frame-to-frame.

Video controls differ slightly from image controls:

• Strength (0-100%) - overall enhancement intensity
• FPS Boost - doubles frame rate for smoother playback
• Sharpen - edge definition
• Smart Grain - preserves or adds photographic grain

Magnific video works best on clips under 30 seconds. For feature-length or long-form content, Topaz Video AI is the right tool.

Magnific workflow (images)

1. Open magnific.com/ai/image-upscaler
2. Upload your image (drag-and-drop or URL)
3. Choose Creative or Precision mode
4. Pick a content preset (Portrait, Photography, etc.)
5. Set scale factor: 2x or 4x for Creative; up to 16x for Precision
6. Adjust Creativity, Resemblance, HDR sliders
7. Add a text prompt describing what the image contains
8. Generate - compare with the original using the built-in viewer
9. Download or use in a follow-on generation

Magnific pricing

Magnific runs on a credit model. Plans (annual pricing) as of 2026: Essential at the entry tier, Premium with more credits, Premium+ which adds a commercial license and music rights, and Pro for high-volume production. Credits do not reset monthly - they are valid for one year. There is a 30-day refund guarantee. No free tier exists.

Topaz Video AI: the right model for the right footage

Topaz Video AI is a local desktop application (with optional cloud rendering) built specifically for video enhancement. It runs on your GPU. The key design choice is model specialization: rather than one general AI, it ships a library of purpose-trained models, and choosing the right one for your footage type is where most of the quality difference comes from.

As of 2026, Topaz Video AI can upscale to 4K and 8K, reduce noise and compression artifacts, interpolate frames (up to 120 fps), stabilize shaky footage, and convert SDR to HDR. It integrates as a plugin into Premiere Pro and DaVinci Resolve.

Model selection: the most important decision

For progressive (modern) video, the core models are:

• Proteus - the default for most live-action. Versatile, with manual parameter control. Start here if you are unsure. Good for low-to-medium quality footage where you want a customizable result.
• Proteus Natural - same as Proteus but respects the source more aggressively, avoiding over-enhancement. Use when the original footage is decent and you just need resolution.
• Iris - face specialist. Recovers detail in degraded face footage. Use on talking-head content, interview footage, or any clip where faces are the subject. Keep sharpening moderate - faces punish over-sharpening faster than any other content type.
• Nyx / Nyx XL - denoising specialist. Runs at 1x to remove noise while preserving resolution, or at 2x to denoise and upscale simultaneously. Use for high-ISO footage, compressed social media downloads, or anything with visible grain/banding.
• Rhea / Rhea XL - excellent with organic textures. Good choice for 4x upscaling of nature, wildlife, and documentary footage.
• Gaia / Gaia 2 - conservative precision for CG, animation, and high-quality source material. Frame-by-frame processing; no modifiable settings in Gaia 2.
• Artemis - balanced enhancement across six quality-level variants. Use when Proteus feels too aggressive and Gaia feels too soft.

For interlaced video (broadcast archives, old DV tapes), use the Dione model family (Dione DV, Dione TV, Dione Dehalo) before anything else. Deinterlacing is a prerequisite - sending interlaced frames to Proteus will produce artifacts.

A practical two-pass workflow for noisy footage

One of the most effective Topaz workflows is a two-pass approach for difficult source material:

Pass 1: Denoise
- Import noisy clip
- Apply Nyx at 1x (output at original resolution, just cleaner)
- Export as lossless ProRes or DNxHD

Pass 2: Upscale
- Import the clean file from Pass 1
- Apply Proteus or Rhea at 2x or 4x
- Export final delivery file

Running noise reduction and upscaling in a single pass forces the model to do two hard things at once. Splitting them gives each model a clean input and produces noticeably better results, especially on compressed source footage from social platforms.

Topaz Video AI workflow (standalone app)

1. Import your video file
2. Select Video Type (Progressive, Interlaced, Interlaced Progressive)
3. Pick an AI model from the panel (see guide above)
4. Set output resolution - 2x, 4x, or a specific target (e.g. 3840x2160)
5. Click Preview - set a 5-10 second window to test before full export
6. Review the side-by-side comparison
7. Configure export: codec (H.265 for 4K+), container (MP4), bitrate
8. Export - locally on GPU or via Cloud Export for long files

Topaz pricing

Topaz Labs retired the perpetual license in October 2025. As of 2026, Topaz Video runs on a subscription: $299/year for Personal (limited commercial use for organizations under $1M revenue), $699/year for Pro (full commercial rights). A Studio Bundle at $399/year bundles additional Topaz apps.

Magnific vs. Topaz: choosing the right tool

• Still images - Magnific. It is purpose-built for this and the creativity slider gives you control over how much the AI invents versus preserves.
• Short AI-generated video clips (under 30 seconds) - either tool works. Magnific's Creative mode adds hallucinated texture; Topaz is more conservative and temporally stable. Try both on a short preview.
• Long-form video, archival footage, or delivery to broadcast - Topaz Video AI. Local processing, deep model specialization, and plugin support for NLEs make it the professional choice.
• Batch / API workflows - Magnific has an API with credit-based access. Topaz is a desktop app with no batch API, though its CLI mode enables scripting for local pipelines.
• Budget - both require paid subscriptions. Factor in that Magnific charges per upscale via credits; heavy use on large images can exhaust a plan's monthly allocation quickly. Topaz is a flat annual fee with local compute (your GPU).

Key takeaways

• AI upscaling synthesizes new detail - it does not just stretch existing pixels. This is what separates Magnific and Topaz from standard resize operations.
• Magnific's Creativity slider is the defining control: 0 is a faithful upscaler, higher values let the AI hallucinate texture that was never in the source.
• Always use the Resemblance slider to anchor faces and brand-critical elements when running high Creativity values.
• Text prompts in Magnific's Creative mode steer what kind of micro-detail gets hallucinated. Write one that describes your image's content.
• Topaz model selection matters more than any settings dial. Proteus for general live-action, Iris for faces, Nyx for noise, Gaia for CG.
• Two-pass workflows (denoise first, upscale second) consistently outperform single-pass on compressed or noisy source footage.
• Preview before full export in Topaz. A 10-second preview costs seconds; a bad full export costs hours.
• Precision mode in Magnific faithfully upscales defects too. Clean your source before running Precision - it does not fix problems, it magnifies them.

Try this next: once your images and video are upscaled and publish-ready, learn how to structure them into a polished final output. Video post-production basics covers the cut-to-export pipeline that takes enhanced footage all the way to a finished deliverable.