High-Resolution AI Stills: Getting Real 2K and 4K

Canva MCP resolution boundaries and prompt prompt limits

Canva's MCP generate-image tool returned 1088×1456 (about 1.6 MP) at 3:4. It exposes only prompt, aspect ratio and image references; writing '2K' or '2048px' into the prompt changed nothing.

Automated API tooling often enforces rigid output limitations that diverge significantly from capabilities advertised in consumer web interfaces. When querying Canva through its Model Context Protocol endpoint, our automated testing suites recorded a fixed image output dimension of 1088×1456 at a standard 3:4 portrait aspect ratio. This yields approximately 1.6 megapixels of spatial resolution—sufficient for mobile interface cards but completely inadequate for high-resolution print editorials or digital billboard campaigns. Crucially, the MCP schema exposes only basic prompt, aspect ratio, and reference image fields. Attempting to force higher output density by injecting resolution strings like '2K' or '2048px' into the prompt payload produced zero effect on the output raster.

By our verdict, Canva's own Upscale fails on ordinary photos, so we route those to Magnific Classic 2k or to our video finishing chain instead.

Evaluating post-processing utilities requires objective visual assessment. In our comparative trials, applying Canva's native upscaler to photographic still frames introduced smudging and plastic artifacts across fine details. For mission-critical commercial deliverables requiring substantial scaling, our studio routes source frames through Magnific Classic 2k or our dedicated offline video finishing pipelines.

Dream Lab native output and Boost resolution capabilities

The Dream Lab web UI locks its model to Canva Craft-1 2K with no picker, and returns 1760×2336 at 3:4 for each generation.

Accessing the Dream Lab web interface directly provides access to an entirely different generation tier. Within the browser workspace, the generation model is locked to Canva Craft-1 2K without providing model selection dropdowns. Submitting a generation run at a 3:4 aspect ratio consistently delivers a native 1760×2336 raster file. This baseline format provides significantly greater micro-detail than the standard MCP output, resolving fine skin pores and fabric weaves cleanly.

Boost resolution, one click per image, creates a new 3520×4672 file (about 16 MP, about 1.2 MB JPG) at 3:4. Boost makes a new image; the original stays.

To achieve true high-resolution production output, Dream Lab includes a specialized Boost resolution feature. Triggered via a single button click on any selected generation card, this utility synthesizes a new 3520×4672 raster file at 3:4 proportions. This operation yields approximately 16 megapixels of visual information compressed into an efficient 1.2 MB JPEG file. Importantly, Boost does not overwrite the source render; it compiles an independent asset alongside the original frame, allowing direct side-by-side technical evaluation.

A 4-image batch came back in about 25–30 seconds.

Thumbnail inspection traps and local pixel measurement

The grid thumbnails are 414×549 resize URLs, so we read size only from the downloaded file with PIL. Expected: 1760×2336 or 3520×4672.

Evaluating generative output fidelity directly within browser inspection consoles frequently leads to false conclusions. In Dream Lab, the dynamic grid interface populates candidate frames using resized thumbnail URLs rendered at 414×549 pixels. Inspecting DOM element properties or capturing browser viewport screenshots will severely misrepresent the actual underlying generative resolution. In our testing protocols, automated verification scripts download the raw binary asset and inspect image dimensions programmatically using PIL, confirming true raster dimensions of 1760×2336 for native renders or 3520×4672 for Boosted assets.

In video upscaling tests, a broken 2x output passed every dimension check. The green-cyan speckle in eye catchlights only showed up when we compared the same frame before and after.

Dimension metrics alone do not guarantee structural integrity. During our video upscaling experiments, corrupted upscaler outputs successfully matched target 2x pixel counts while embedding minute green-cyan color noise in reflective eye catchlights. Rigorous quality assurance requires comparing identical cropped regions against the source frame before approving commercial delivery.

Macro texture fidelity and surface artifact evaluation

On a 1:1 crop of an oxblood leather bag test, grain, stitching, edge paint and brushed-brass hardware resolved as real detail after Boost, not as a soft upscale.

Verifying whether Boost resolution performs genuine high-frequency synthesis or merely applies a bicubic interpolation filter requires examining 1:1 pixel crops. In our benchmark test evaluating an oxblood leather luxury handbag, zooming into 100% scale revealed crisp, authentic material rendering. Natural leather grain patterns, individual thread twists in the structural stitching, sealed edge paint seams, and micro-scratches across brushed-brass metallic hardware resolved as tangible geometric detail rather than blurred or hallucinated noise.

The one artefact we saw: faint wormy texture on flat leather at 100% zoom. Invisible at phone and web size.

Despite impressive textural clarity, diffusion upscalers exhibit subtle characteristic artifacts across smooth uniform surfaces. At 100% magnification on untextured leather panels, our reviewers detected a faint, wormy latent pattern characteristic of high-step denoising filters. However, this microscopic artifact remained entirely invisible when viewed at standard web or mobile viewport resolutions, making it fully acceptable for commercial digital campaigns.

Production unknowns and commercial graphic boundaries

Unknown: the credit cost per generation and per Boost (no counter in the UI or MCP), whether Boost can chain past 4.7K, and whether persona identity holds in Dream Lab. We have not tested any of these.

Maintaining strict studio transparency requires distinguishing verified empirical facts from unmeasured platform parameters. Currently, Canva provides no public credit consumption counters within the Dream Lab UI or MCP documentation, leaving the exact credit expenditure per generation or Boost pass unquantified. Furthermore, our team has not tested whether Boost can be chained sequentially past 4.7K, nor have we verified whether facial identity locks hold reliably across synthetic talent reference sets within Dream Lab.

Brand wordmarks through Dream Lab are untested. We assume the model will smudge them and graft the real letters in afterwards.

Typography and trademark rendering represent another untested operational boundary. Until thorough character and wordmark benchmarks are completed, our pipeline assumes Dream Lab will distort fine lettering, requiring real vector wordmarks to be grafted manually in post-production.