{"slug": "clip", "name": "CLIP", "category": "ai-models", "category_name": "AI models", "description": "", "recommenders": 5, "commenters": 12, "warners": 0, "score": 6.098, "rising": -0.31, "first": "2025-12-03", "latest": "2026-09-24", "why": "CLIP should give the images that actually look similar.", "why_is_quote": true, "has_verdict": false, "sentiment": {"recommend": 5, "neutral": 5, "compare": 2, "warn": 0}, "spark": [0, 0, 2, 0, 0, 0, 1, 1, 0, 0, 1, 0], "listed": true, "loved": [], "caveat": "", "short_quote": "CLIP should give the images that actually look similar.", "kind": "", "icon": false, "domain": "openai.com", "github_repo": null, "official_url": "https://github.com/openai/CLIP", "pricing": "", "aliases": ["CLIP"], "verdict": null, "praise": [{"label": "", "summary": "", "quotes": [{"id": 797248576, "item_id": 49828036, "url": "https://news.ycombinator.com/item?id=49828036", "by": "andy12_", "date": "2026-09-24", "version": null, "stance": "recommend", "quote": "CLIP[1] for actions! Very cool.", "excerpt": "CLIP[1] for actions! Very cool. [1] https://openai.com/index/clip/", "context": "", "use_case": "", "switched_from": "", "story": "Contrastive Language Models"}, {"id": 775269280, "item_id": 48454330, "url": "https://news.ycombinator.com/item?id=48454330", "by": "krackers", "date": "2026-06-08", "version": null, "stance": "recommend", "quote": "See CLIP\nhttps://github.com/openai/CLIP", "excerpt": ">which describes images in natural language, See CLIP https://github.com/openai/CLIP", "context": "", "use_case": "", "switched_from": "", "story": "Playing with Vision Embeddings"}, {"id": 773500017, "item_id": 48343751, "url": "https://news.ycombinator.com/item?id=48343751", "by": "sureglymop", "date": "2026-05-31", "version": null, "stance": "recommend", "quote": "smaller specialised things like whisper, CLIP, SAM, etc.", "excerpt": "What I'm actually excited about in AI are smaller specialised things like whisper, CLIP, SAM, etc. Just stacking them together as little building blocks has been amazing. Great app to hopefully let more people realize that it's not all just huge LLMs that have to run somewhere else!", "context": "", "use_case": "", "switched_from": "", "story": "Ask HN: What Is the State of App Development in 2026?"}, {"id": 744378256, "item_id": 46523641, "url": "https://news.ycombinator.com/item?id=46523641", "by": "smusamashah", "date": "2026-01-06", "version": null, "stance": "recommend", "quote": "CLIP should give the images that actually look similar.", "excerpt": "As suggested in another comment, CLIP should give the images that actually look similar. This is a great collection of images, and using those you will find very similar covers. http://same.energy/search?i=yN5ou Also https://mood.zip/", "context": "", "use_case": "image similarity search", "switched_from": "", "story": null}, {"id": 744333184, "item_id": 46520824, "url": "https://news.ycombinator.com/item?id=46520824", "by": "Samin100", "date": "2026-01-06", "version": null, "stance": "recommend", "quote": "Using a CLIP or Dino v2 model to produce image embeddings would probably improve the similarity search a lot", "excerpt": "This is awesome! Using a CLIP or Dino v2 model to produce image embeddings would probably improve the similarity search a lot - kind of similar to http://same.energy/ .", "context": "", "use_case": "image embeddings", "switched_from": "", "story": null}]}], "caveats": [], "who": [], "tags": [], "use_cases": ["image embeddings", "vision encoder for vlms", "image similarity search", "visual recognition frontend"], "compared": [{"slug": "claude-opus", "name": "Claude Opus", "co": 2, "switched_from": 0, "switched_to": 0}], "makers": [], "threads": [{"url": "https://news.ycombinator.com/item?id=46518106", "title": "item 46518106", "mentions": 2}, {"url": "https://news.ycombinator.com/item?id=49826221", "title": "Contrastive Language Models", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=49458502", "title": "Getting video models to learn better, faster", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=48676795", "title": "Apple to Skip High-End M6 Mac Chips in Favor of AI-Focused M7 Line", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=48413366", "title": "Playing with Vision Embeddings", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=48337409", "title": "Ask HN: What Is the State of App Development in 2026?", "mentions": 1}], "sentiment_by_month": {"2025-11": {}, "2025-12": {"neutral": 2}, "2026-01": {"recommend": 2, "neutral": 1}, "2026-02": {"compare": 1}, "2026-03": {}, "2026-04": {"neutral": 1}, "2026-05": {"recommend": 1}, "2026-06": {"recommend": 1, "neutral": 1}, "2026-07": {}, "2026-08": {"compare": 1}, "2026-09": {"recommend": 1}, "2026-10": {}}, "versions": [], "version_unnamed": 12, "family_ref": null, "runs_on": null, "lines": [], "quotes": [{"id": 797248576, "item_id": 49828036, "url": "https://news.ycombinator.com/item?id=49828036", "by": "andy12_", "date": "2026-09-24", "version": null, "stance": "recommend", "quote": "CLIP[1] for actions! Very cool.", "excerpt": "CLIP[1] for actions! Very cool. [1] https://openai.com/index/clip/", "context": "", "use_case": "", "switched_from": "", "story": "Contrastive Language Models"}, {"id": 791349680, "item_id": 49459355, "url": "https://news.ycombinator.com/item?id=49459355", "by": "schopra909", "date": "2026-08-27", "version": null, "stance": "compare", "quote": "the hidden states of multimodal LLMs act as better conditioning mechanisms", "excerpt": "Aah, for this we're just trying to filter not generate. When it comes to conditioning, you'll still need a model that understands text since the primary control is text. In the original Stable Diffusion, CLIP doubled as part of this LAION Aesthetic Filter as well as the text control for the generative model. But in the last couple of years, the field has drifted away from CLIP towards LLM hidden states (e.g. Qwen, Mi", "context": "", "use_case": "", "switched_from": "", "story": "Getting video models to learn better, faster"}, {"id": 778950849, "item_id": 48684428, "url": "https://news.ycombinator.com/item?id=48684428", "by": "Mashimo", "date": "2026-06-26", "version": null, "stance": "neutral", "quote": "the CLIP model for the context search", "excerpt": "Not for faces, but the CLIP model for the context search https://docs.immich.app/features/searching/ That needs to be in (v)ram for searches.", "context": "", "use_case": "", "switched_from": "", "story": "Apple to Skip High-End M6 Mac Chips in Favor of AI-Focused M7 Line"}, {"id": 775269280, "item_id": 48454330, "url": "https://news.ycombinator.com/item?id=48454330", "by": "krackers", "date": "2026-06-08", "version": null, "stance": "recommend", "quote": "See CLIP\nhttps://github.com/openai/CLIP", "excerpt": ">which describes images in natural language, See CLIP https://github.com/openai/CLIP", "context": "", "use_case": "", "switched_from": "", "story": "Playing with Vision Embeddings"}, {"id": 773500017, "item_id": 48343751, "url": "https://news.ycombinator.com/item?id=48343751", "by": "sureglymop", "date": "2026-05-31", "version": null, "stance": "recommend", "quote": "smaller specialised things like whisper, CLIP, SAM, etc.", "excerpt": "What I'm actually excited about in AI are smaller specialised things like whisper, CLIP, SAM, etc. Just stacking them together as little building blocks has been amazing. Great app to hopefully let more people realize that it's not all just huge LLMs that have to run somewhere else!", "context": "", "use_case": "", "switched_from": "", "story": "Ask HN: What Is the State of App Development in 2026?"}, {"id": 765679874, "item_id": 47854992, "url": "https://news.ycombinator.com/item?id=47854992", "by": "jperryjperry", "date": "2026-04-21", "version": null, "stance": "neutral", "quote": "generated CLIP embeddings for all 191,922 of them", "excerpt": "I fetched every Met Museum CC0 artwork with a published image, generated CLIP embeddings for all 191,922 of them, and searched for visually near-identical pairs separated by 2,000+ years. I used Cursor (Opus 4.7, max effort) to help build it, including spinning up a ~500 vCPU Burla cluster for the image downloads and embedding jobs. I also had it study popular museum websites for design inspiration while building the", "context": "", "use_case": "image embeddings", "switched_from": "", "story": "Embedded 191,922 Met artworks to find hidden twins across 4k years"}, {"id": 752324081, "item_id": 47020255, "url": "https://news.ycombinator.com/item?id=47020255", "by": "teendifferent", "date": "2026-02-14", "version": null, "stance": "compare", "quote": "CLIP dominates on average", "excerpt": "OP here. I wanted to test whether the vision encoder's pre-training strategy matters when you stitch it into an LLM. So I froze three encoders (CLIP, I-JEPA, supervised ViT), stitched each into Qwen2.5 with a small trainable projector + LoRA (~3M params), and compared. Key findings: - CLIP dominates on average (language-aligned embeddings make the projector's job trivial). - But I-JEPA \u2014 which has never seen text dur", "context": "", "use_case": "Vision encoder for VLMs", "switched_from": "", "story": "Stitching Vision Encoders into LLMs: Clip vs. I-JEPA vs. ViT Comparison"}, {"id": 746918256, "item_id": 46682391, "url": "https://news.ycombinator.com/item?id=46682391", "by": "mattdesl", "date": "2026-01-19", "version": null, "stance": "neutral", "quote": "I'm currently exploring similar things with CLIP to guide fitness", "excerpt": "Nice work, the outputs look ethereal and quite beautiful. For some related work using genetic algorithms and evolution strategies, see[1]. Sketch synthesis is an area I'm pretty interested in lately; I'm currently exploring similar things with CLIP to guide fitness, natural evolution strategy to optimize the rendered results, and using an implicit neural representation to represent pen plotter paths (rather than a se", "context": "", "use_case": "", "switched_from": "", "story": "Iterative image reconstruction using random cubic b\u00e9zier strokes"}, {"id": 744378256, "item_id": 46523641, "url": "https://news.ycombinator.com/item?id=46523641", "by": "smusamashah", "date": "2026-01-06", "version": null, "stance": "recommend", "quote": "CLIP should give the images that actually look similar.", "excerpt": "As suggested in another comment, CLIP should give the images that actually look similar. This is a great collection of images, and using those you will find very similar covers. http://same.energy/search?i=yN5ou Also https://mood.zip/", "context": "", "use_case": "image similarity search", "switched_from": "", "story": null}, {"id": 744333184, "item_id": 46520824, "url": "https://news.ycombinator.com/item?id=46520824", "by": "Samin100", "date": "2026-01-06", "version": null, "stance": "recommend", "quote": "Using a CLIP or Dino v2 model to produce image embeddings would probably improve the similarity search a lot", "excerpt": "This is awesome! Using a CLIP or Dino v2 model to produce image embeddings would probably improve the similarity search a lot - kind of similar to http://same.energy/ .", "context": "", "use_case": "image embeddings", "switched_from": "", "story": null}, {"id": 743070709, "item_id": 46441919, "url": "https://news.ycombinator.com/item?id=46441919", "by": "solid_fuel", "date": "2025-12-30", "version": null, "stance": "neutral", "quote": "using the existing CLIP model as a visual frontend and then training a small classifier on the output of CLIP", "excerpt": "I've actually done a fair bit of ML work in Elixir, in practice I found: 1) It's generally harder to interface with existing libraries and models (example: whisperX [0] is a library that combines generic whisper speech recognition models with some additional tools like discrete-time-warping to create a transcription with more accurate time stamp alignment - something that was very helpful when generating subtitles. B", "context": "", "use_case": "visual recognition frontend", "switched_from": "", "story": "Go Away Python"}, {"id": 738236083, "item_id": 46139755, "url": "https://news.ycombinator.com/item?id=46139755", "by": "laidoffamazon", "date": "2025-12-03", "version": null, "stance": "neutral", "quote": "normalized for CLIP embedding on GPU", "excerpt": "Sure. I had a lot of help from Claude Opus 4.5, but it was roughly: - Using pyscenedetect to split each video on a per scene level - Using the decord library https://github.com/dmlc/decord to pull frames from each scene at a particular sample rate (specific rate I don't have handy right now, but it was 1-2 per scene) - Aggregating frames in batches of around 256 frames to be normalized for CLIP embedding on GPU (had ", "context": "", "use_case": "", "switched_from": "", "story": "Qwen3-VL can scan two-hour videos and pinpoint nearly every detail"}]}