What Do the Subject and Purpose Tags Mean on OpenModelDB?

Subject and purpose tags on OpenModelDB are the two filters that answer different questions about a model. Subject tags describe what the model was trained on — anime, cartoon, realistic footage, pixel art, and so on. Purpose tags describe what the model is meant to do to that footage — denoise, deblur, dehalo, upscale, restore, and so on. You pick a subject to match your source material, then a purpose to match the defect or goal you have in mind, and the database narrows from 671 models to the ones actually relevant to your job.

Subject tags: what the model expects as input

Subject tags classify a model by the kind of content it was trained to handle. The site lists these subject categories:

• AI Generated — output from image or video generators
• Anime — Japanese-style animation
• Cartoon — Western-style animation
• CGI — computer-generated imagery
• Faces — close-up or prominent human faces
• Game Screenshots — captured gameplay frames
• Game Textures — in-game texture assets
• Manga — black-and-white comic pages
• Pixel Art — low-resolution sprite-style graphics
• Realistic — live-action photographic footage
• Text — text-heavy frames or scans
• Video Frame — general video frames

The subject tag is a statement about the training data, not a hard restriction. A model tagged Cartoon was trained on cel animation, so it will behave best on cel animation and may produce artifacts on live-action footage. The Sonic example on the site makes this concrete: 90s Sonic 2x (Small) and 90s Sonic 2x (Large) are both tagged Cartoon, Restoration, and Video Frame, and the description says they were designed specifically for Adventures of Sonic the Hedgehog and Sonic SatAm — two specific 1993–1994 shows. That is a very narrow subject match, and using it on unrelated cartoons is a gamble.

Purpose tags: what the model is meant to fix or produce

Purpose tags describe the transformation the model performs. The site lists:

• Anti-aliasing — smoothing jagged edges
• Colorization — adding color to grayscale
• Compression Removal — reducing compression artifacts
• DDS — handling DDS texture formats
• Debanding — removing color banding
• Deblur — sharpening blurred detail
• Dedither — removing dithering patterns
• Dehalftone — removing halftone print patterns
• Dehalo — removing ringing/halo artifacts
• Denoise — reducing noise
• General Upscaler — broad upscaling without a narrow specialty
• Inpainting — filling in missing or damaged regions
• JPEG — repairing JPEG artifacts
• Pretrained — a base model intended for further training
• Research — experimental or academic models
• Restoration — general repair of degraded footage
• Texture Generation — creating texture detail

A single model can carry several purpose tags. StarSample V2.0 Lite NS is tagged Anime and Restoration, and its description explains that V2.0 improved its handling of compression and reduced "AI smudging" — that is a purpose claim about compression removal and artifact suppression, not just a subject claim.

How the two tag types work together

The tags are designed to be combined, not read in isolation. A practical selection flow looks like this:

1. Identify your subject. Is your source anime, live-action, a game screenshot, pixel art, or something else? Filter by that subject first.
2. Identify your purpose. Is the problem noise, blur, compression, halos, banding, or is the goal simply to upscale clean footage? Filter by that purpose second.
3. Read the model description. Tags are broad; descriptions are specific. The Sonic models are tagged Cartoon + Restoration, but the description reveals they target two named shows and even recommends QTGMC deinterlacing for maximum quality — details the tags alone cannot convey.
4. Check the scale factor. Model names often carry a multiplier such as 2x or 1x. Adore is described as a "2x model for Realtime 1080p anime upscaler," while StarSample V2.0 Lite NS is listed at 1x. Match the scale factor to your target resolution.

Why the distinction matters when choosing

Subject and purpose tags fail in different ways if you ignore them:

• Wrong subject → the model applies detail patterns from the wrong domain. A pixel-art model run on photographic footage can invent blocky edges that were never there.
• Wrong purpose → the model does work you did not ask for. A denoise model run on already-clean footage can soften real detail; a general upscaler run on heavily compressed video may amplify the compression instead of removing it.

The site's stated aim is to let you "find and compare models" better than existing sources, and the tag system is the mechanism for that comparison. Two models can share a subject but differ in purpose — for example, two Anime models where one is a General Upscaler and the other is Restoration — and the purpose tag is what tells you which one fits a degraded source versus a clean one.

A quick reference

Use the subject tag to match the model to your material and the purpose tag to match it to your defect. When the two overlap with several candidates, the model description — not the tags — is where you confirm the specific shows, resolutions, and scale factors a model was built for.