Reference Image Tool Crashes With 1000+ Images: Fixes and Alternatives (2026)

By refern | Last updated: June 2026

Your reference tool crashes or slows to a crawl once your image collection grows past a few hundred items. This is a documented architectural limit in several popular tools, not user error. Allusion's database stops displaying images past roughly 120,000 files; PureRef loads every image uncompressed into RAM; Milanote's canvas lags above 300 to 500 cards per board; Miro freezes on boards with 2,000 or more elements. This page explains why each tool hits its ceiling and what a streaming-index architecture does differently.

Why reference tools break at scale

Reference tools fail at large libraries for one of three reasons.

RAM loading. The tool reads every image into memory when you open a project or board. This works fine for 50 images and becomes unmanageable at 5,000. PureRef is the clearest example of this model.

Database ceiling. The tool stores metadata in a local database that was never designed to grow past a certain size. Once the file on disk exceeds a threshold, queries time out or fail entirely. Allusion hits this ceiling around 120,000 images.

Canvas element limits. Web-based board tools like Milanote and Miro render everything on a single browser canvas. Past a few hundred to a few thousand elements, rendering performance collapses.

The fix is a pipeline that never loads the full library into memory: index images from disk, store metadata in a streaming-capable database, and fetch only what the current view needs.

How each popular tool fails at scale

PureRef: all images loaded into RAM uncompressed

PureRef is an excellent focused reference board for per-project use. For large collections it has a hard architectural constraint: every image on a board is held in memory uncompressed.

The developers confirmed this on their official forum: "PureRef will hold all loaded images in memory uncompressed, so it can get out of hand if you go crazy with your board." Their recommended workaround is to split a large board into multiple smaller ones. A user on the same forum reported 30-second load times for images that "aren't even that large," with significantly higher RAM usage after an update. The developer response was: "We want PureRef to use less ram in the future... I can't promise when we will have that done as it's somewhat complicated."

For artists who use PureRef as a session overlay while painting, this is manageable. For anyone trying to maintain a library of hundreds or thousands of images in a single board, it is not.

Other PureRef scale limitations: no search, no tags, no cross-project library. Every .pur file is a standalone board; there is no "all references I have ever collected" view. If you collect references across years and projects, PureRef has no way to find them.

PureRef is genuinely excellent at what it does. Its always-on-top overlay, transparent-to-mouse click-through, and laser focus are best-in-class for per-project boards. The scale problem is simply outside its design scope.

Allusion: database ceiling at roughly 120,000 images, plus a severe RAM leak

Allusion is a free, open-source desktop reference library manager. It indexes images from watched folders on disk, hierarchical tags, and basic search. For small to medium libraries it works well.

Two documented failures affect larger libraries.

RAM leak during thumbnail generation. GitHub issue #640 (November 2024) documents 14.4 GB of RAM consumed while generating thumbnails for just 358 images on Windows 11. Memory is never reclaimed without a restart. No fix has been shipped by the maintainers.

Database size ceiling. GitHub issue #604 documents the app stopping its display of all images once the database file exceeds roughly 81 to 82 MB, which corresponds to approximately 120,000 or more images. The user had to restore from backups repeatedly, but the problem recurred because the database kept growing. No fix has been shipped.

GitHub issue #423 also documents crashes at 300,000 files (5 GB or more RAM, crash on restart), 40,000 files (crash during file-check), and 90,000 files (crash during startup file-check). The startup file-check scan is the identified bottleneck.

An additional concern: Allusion's last official release was v1.0.0-rc.10 on February 6, 2023. A GitHub issue filed April 2025 is titled "Project no longer maintained - Try these forks instead." There are 83 open issues with no maintainer responses. These RAM and database bugs are unlikely to be fixed in the official project.

Allusion's strongest advantages are that it is completely free, open source (GPL-3.0), and cross-platform. For small to medium libraries and users who only need basic tag-based search, it works. For libraries above 50,000 images, the documented failures become critical.

Milanote: performance degrades at 300 to 500 cards per board

Milanote is a cloud-based visual board tool aimed at creative project planning. It is polished, easy to use, and good for collaboration and client presentations. It is not an image library manager.

Boards with 300 or more images, or 500 or more total cards, report lag, slow interactions, and occasional crashes. Image-heavy boards hit this threshold earlier than text-heavy ones. This is a browser-canvas rendering ceiling that affects all web-based board tools.

Milanote also has no folder hierarchy for a library of thousands of images, no bulk tagging, no duplicate detection, no color search, and no visual similarity search. It is built for organizing creative projects, not for maintaining a growing reference archive.

Milanote's free tier allows only 100 total items (notes, images, and links combined), which is exhausted quickly in active creative work. The Individual plan is $9.99 per month billed annually (as of 2026). Over three years that is $360 versus a one-time purchase for a dedicated library tool.

Milanote is genuinely good at what it was designed for: polished collaborative boards with real-time co-editing, 100 or more profession-specific templates, and sharing with clients. The scale ceiling is outside its intended use case.

Miro: freezes above 2,000 elements, not designed for image libraries

Miro is an enterprise collaborative whiteboard built for team workshops, sprint planning, and design-thinking sessions. It is not a reference image manager.

A community thread with 56 replies documents 5 to 6 second response delays for basic interactions, with the issues correlating with boards containing 2,000 or more elements. Users quoted: "It takes up to 5-6 seconds for the board to respond to an interaction" and "I'm close to going back to other apps because the lag in creating, editing and moving objects is just too much."

Beyond performance, Miro has no image library features: no folder hierarchy for images, no bulk tagging by image, no metadata, no duplicate detection, no color search. Images pasted onto boards are not indexed or searchable by content.

Miro's per-seat subscription pricing ($8 to $20 per seat per month, annual billing, as of 2026) is also designed for teams, not individual artists.

Where Miro genuinely wins is real-time multi-user collaboration, 5,000 or more templates, and 160 or more integrations with enterprise tools like Jira, Slack, and Figma. For team workshops and corporate workflows it is well-suited. For an artist's personal reference library, it is the wrong tool.

What a streaming-index architecture does differently

The core difference is that a streaming pipeline never loads the full library into memory.

refern is a desktop reference manager for artists that combines Eagle-style organization with a PureRef-style infinite canvas and an Obsidian-style relationship graph. It costs $30 one time, runs on Windows, macOS, and Linux, and does not copy your files.

How refern handles large libraries

refern indexes images from a workspace folder on disk. The files stay exactly where they are; refern stores an index (SQLite with FTS5 full-text search) and thumbnails alongside the originals.

The indexing pipeline is streaming by design. Images are processed in a bounded worker pool using a rayon parallel bridge with a capped channel. Memory stays bounded regardless of library size because the pipeline never loads all pending images into RAM at once. A user with 27,000 images confirmed smooth performance.

If the process is interrupted (crash, forced quit, power loss), the pipeline resumes from where it stopped. The pipeline state is written to disk before work begins at each phase.

The database itself stays fast because it uses SQLite in WAL mode with FTS5 indexing and a layout-descriptor architecture: the grid fetches lightweight descriptors (~50 bytes per item) for layout, then fetches full entity data only for the roughly 50 items visible in the viewport. A 500,000-image library does not require 500,000 full rows loaded into the UI at once.

Thumbnails are stored as WebP files in a hashed two-level directory structure. The thumbnail pipeline is separate from the UI thread; thumbnail generation never blocks browsing.

For libraries above 10,000 images, this architecture behaves differently from tools that were not built with scale in mind.

Comparison table

How to migrate from a tool that is hitting its ceiling

From Allusion. refern includes an Eagle importer, not a direct Allusion importer. If your images are in a folder that Allusion watches, you can open that same folder as a refern workspace. refern will index the images in place, reading embedded EXIF/IPTC/XMP metadata on import. Your files are not moved or copied.

From PureRef. PureRef embeds images inside a proprietary .pur binary. To migrate, export your images from PureRef first (File > Export Images), save them to a folder, then open that folder as a refern workspace. refern will index and thumbnail them. For future work, you can use refern's canvas as a direct replacement for per-project PureRef boards, with the addition of a persistent searchable library behind it.

From Milanote or Miro. Both tools store images on their cloud servers. Download a ZIP of your board images using the board's export function, save them to a local folder, and open that folder in refern.

In all cases: refern never copies your original files. The workspace is a normal folder; refern stores the index and thumbnails inside it alongside your files. There is no vendor lock-in.

See also: refern vs PureRef comparison and refern vs Eagle comparison.

Honest limitations of refern at scale

refern is actively developed and launched in June 2026. A few caveats for very large libraries:

Cloud sync is not yet available (planned for Phase 2). If you work across multiple machines today, your workspace folder can live on a synced drive (Dropbox, iCloud, OneDrive) and refern will open it on each machine.

AVIF files are not supported in the current version. The backend has no AVIF decoder. AVIF files are skipped during indexing.

3D model preview is not yet shipped (planned).

Auto-tagging with a local model is not yet shipped (planned).

The streaming pipeline is designed to scale to millions of files. If your library is above 500,000 images, behavior at that scale has not been publicly validated by a user report.