Imagga alternatives for tagging catalogs and media libraries

If you use Imagga to tag or categorize a product catalog or media library, the realistic shortlist is shorter than most comparison lists suggest. Google Cloud Vision and Amazon Rekognition both return general labels at per-image prices. A self-hosted open model such as SigLIP 2 can sort images into your own categories if you're willing to run it. Two names you'll see recommended need care: Clarifai's public API no longer appears to be running, and Microsoft is retiring the Azure Image Analysis API that returns tags.

Keep Imagga in the comparison too. Its tagging API can return tags in 46 languages, and its categorization API includes public categorizers plus custom ones built for your taxonomy. If either matters to you, the alternatives may add work rather than remove it.

This comparison is based on each provider's documentation and pricing pages, checked on September 24, 2026. It is not an accuracy test we ran.

Match the replacement to the tagging job

What you would be replacing in Imagga

Imagga's API documentation describes two tagging endpoints. The v2 tagging API returns a flat list of tags with confidence scores, a threshold parameter and optional output in 46 languages, English included. Imagga says most of those translations are generated automatically and may be less precise for niche or abstract concepts, so the translation check in the table applies to Imagga too. The newer v3 structured tagging returns grouped tags (objects, scene, mood, colors and more) without confidence scores, aimed at automation. Categorization is separate: you pick a categorizer, such as general_v3 with about 4,000 categories or personal_photos with 13, and each image receives categories with scores.

Custom categorizers are built with Imagga's team. Its custom training page describes supplying non-overlapping categories and sample images, after which Imagga trains the model and serves it through the categorization API, in its cloud or on your premises.

Imagga's pricing, as of September 24, 2026, lists a free plan with 100 API requests a month (it asks for a "Powered by Imagga" credit), Indie at $79 a month for 70,000 requests, Pro at $349 a month for 300,000 requests, and custom Enterprise pricing above 1,000,000 requests. Plans differ in which APIs they include. The documentation also notes that one request can count as several processed items on some endpoints, so check the usage report before comparing per-image prices.

Clarifai and Azure: two familiar names to handle carefully

Clarifai still appears in many alternatives lists. In May 2026, Nebius announced that Clarifai's core team was joining Nebius, and that its license to Clarifai technology excludes Clarifai's legacy computer vision models. When checked on September 24, 2026, Clarifai's API hostname did not resolve and its website did not respond. One competing vendor's migration guide reports that Clarifai's services ended on July 17, 2026; no official Clarifai notice confirming that date could be retrieved, so treat it as unconfirmed. Either way, don't plan new work on it. If you relied on its food or general models, test the candidates below on your own images.

Microsoft's migration guide says the Azure Image Analysis API, versions 3.2 and 4.0, will be retired on September 25, 2028, including container deployments, after which calls fail. Both versions list tagging among their features. Microsoft points image-analysis users toward generative models, its Content Understanding service (whose image support is labelled preview) or embedding models such as SigLIP. Azure remains an option for Microsoft-centered teams, but not through the tagging API older comparisons describe.

Google Cloud Vision: broad English labels with per-feature billing

Label detection returns a description, a confidence score, a topicality score and a Knowledge Graph identifier for each label. It returns 10 labels by default unless you set maxResults, and labels are English only. For a catalog in several languages, that means a translation step. The Knowledge Graph identifier can help you merge labels that mean the same thing.

Cloud Vision pricing, as of September 24, 2026, bills each feature applied to an image as a unit. Label detection is free for the first 1,000 units a month, $1.50 per 1,000 units up to 5,000,000, and $1.00 per 1,000 after that. SafeSearch is free when requested with label detection. The free allowance suits evaluation. Beyond it, a "free Google Vision alternative" in practice means hosting a model yourself.

For custom categories, Google's option is AutoML image classification, a separate training product with its own costs. Its training guide now sits in Google's Gemini Enterprise Agent Platform documentation, and older Vertex AI links redirect there.

Amazon Rekognition: a label hierarchy that helps with navigation

Rekognition's label detection returns each label with its ancestors, a category such as "Food and Beverage", and aliases like "Cell Phone" for "Mobile Phone". You can include or exclude labels and categories in the request. For a catalog, that hierarchy can map onto browse filters with less mapping code than a flat tag list. Labels are returned in English. Dominant colors and image quality come from its Image Properties option, which is priced separately.

AWS's pricing page places DetectLabels in its Group 2 image APIs, and its worked example uses $0.001 per image for the first million. Check your region's rate. Custom Labels bills for training and for every hour a model is running, whether or not it processes images, so batch work needs a start-and-stop routine.

Self-hosted zero-shot models: no license fee, not free to run

SigLIP 2, released by Google under Apache 2.0, supports zero-shot image classification: you pass an image and a list of candidate labels, and it scores each label. Your taxonomy becomes the label list, so adding a category means editing a list rather than retraining a model.

That flexibility moves work to your team. You host the model, batch requests, monitor accuracy and decide when to update it. It returns scores only for the labels you supply, not a general tag list, and it includes no moderation. For a media library with steady volume, that can cost less than per-image fees. For a small catalog, a managed API is probably less work.

Run a like-for-like tagging test

Take a few hundred images that represent your catalog, including difficult ones. Suppose a recipe site needs "grilled salmon", "vegan" and "dessert" as browse categories. Run the same images through Imagga and each candidate, then map every provider's output to those categories before scoring anything. General labels such as "food" or "dish" may be accurate and still useless for navigation.

For each candidate, record:

• Correct and incorrect assignments per category at the threshold you would use in production.
• Categories the provider never returns, which you would need custom training or zero-shot labels to cover.
• Output languages, and the cost and quality of any translation step.
• The total monthly cost at your volume, counting every feature you request per image.
• How model updates are signalled. Rekognition, for example, returns the label model version with each response.

Switch when a candidate covers your categories as well or better at a cost and maintenance load you accept. If Imagga's categorizers and language support already fit, the better move may be tuning thresholds or commissioning a custom categorizer, not a migration.