Vision & Utilities

Whisper large-v3 optimized for speed (38M+ runs) — transcribes roughly 150 minutes of audio in under 100 seconds using batched inference. Chunk-level or word-level timestamps.

Meta's Muse Voice Transcribe speech-to-text over the Meta developer API - 25 languages with code-switching (Korean included), speaker diarization, endpointing, keyword and language biasing. Billed per second of audio (0.11625 credits/s, about 7 credits per minute), up to 10 minutes per request.

Meta's Muse Voice Transcribe over a WebSocket - stream 16-bit PCM (24 kHz or 16 kHz mono) and receive partial and final transcripts while you speak, with endpointing and diarization. Billed per second of audio Meta processed (0.11625 credits/s, 35 credits per 5-minute block), sessions up to 20 minutes.

OpenAI's Whisper large-v3 speech recognition (144M+ runs) — the standard for transcription. Automatic language detection across ~100 languages, English translation, and plain text / SRT / VTT output formats.

Whisper transcription with speaker diarization (8M+ runs) — returns who said what, with per-segment speaker labels and timestamps. The go-to for meetings and interviews.

Salesforce BLIP (173M+ runs) — image captioning, visual question answering, and image-text matching in one model. The classic choice for bulk captioning at 1 credit per image.

CLIP ViT-L/14 embeddings for text AND images (163M+ runs) — puts both in the same vector space for cross-modal search, image dedup, and zero-shot classification. 1 credit per run.

Microsoft's Florence-2 all-in-one vision model — captioning, object detection, phrase grounding, OCR, and segmentation in a single API. Pick a task, optionally add text input, done.

Text-prompted object detection (39M+ runs) — describe what to find in natural language ('red car, person wearing a hat') and get bounding boxes with confidence scores plus an annotated image.

Tencent's flagship 3D generation — create high-polygon textured 3D models from a text prompt OR an image, with optional PBR (physically based rendering) materials.

Small but capable vision-language model (14M+ runs) — ask free-form questions about any image and get detailed answers. Efficient VQA for tagging, moderation prep, and rich alt-text.

Multilingual text embeddings (74M+ runs) — 1024-dimension vectors across 100 languages including Korean. Pairs perfectly with Core.Today customer databases' vector search (knn_vector).

The standard NSFW image classifier (127M+ runs) — returns 'normal' or 'nsfw' for any image. An essential, ultra-cheap moderation gate for UGC platforms.

Simple, massively-used OCR (91M+ runs) — extracts text from an image with a single input and returns plain text. Great default for receipts, screenshots, and scanned documents at just 1 credit.

Microsoft's TRELLIS image-to-3D (836K+ runs — the most-used 3D model on Replicate). Turns one or more images into a textured GLB 3D asset, with turntable render videos and optional Gaussian PLY.

Add a text watermark to any video — simple, fast brand protection for generated or user content at 2 credits per video.

Video understanding and captioning — ask free-form questions about a video and get detailed answers about actions, scenes, and content. Great for video search indexing and moderation prep.

ByteDance's open-source lipsync — re-syncs a video's mouth movements to any audio track using latent diffusion. State-of-the-art open lipsync quality for dubbing and localization.

Add AI-generated sound to any video (5M+ runs) — synthesizes synchronized audio (ambience, effects, foley) from the video content and an optional text prompt. The perfect finisher for silent AI-generated clips.

Video upscaling with Real-ESRGAN — enhance videos to FHD, 2K, or 4K frame by frame. The go-to open-source video upscaler for old footage and AI-generated clips.

Talking-head video from a single photo and an audio track — animates the face with natural head motion and eye blinks, with optional GFPGAN face enhancement.

Veed Lipsync v2 via Fal.AI. Replaces a source video's mouth movements to articulate a new audio track — takes any source video + a new audio track and produces a lip-synced output video.

FFmpeg-powered video utilities (20M+ runs) — convert to mp4/gif, extract audio as mp3, or dump zipped frames, all with one task parameter. The Swiss-army knife for media pipelines.