Replicate

Salesforce BLIP. Vision-language model for image captioning and visual question answering. Given an image it writes a short natural-language caption, or answers a question about the image when one is supplied. A widely used baseline for automatic captioning.

pharmapsychotic's CLIP Interrogator. Takes an image and produces a Stable-Diffusion-style text prompt by combining BLIP captioning with CLIP to rank likely subjects, artists, mediums and styles. Commonly used to reverse-engineer a prompt from an existing picture.

Monocular depth-estimation model trained on 595k labeled and 62M unlabeled images. Strong zero-shot generalization in indoor and outdoor scenes.

Black Forest Labs' flagship text-to-image model. Faster generation than FLUX.1 Pro at higher prompt adherence, with strong photorealism and reliable spatial composition. Runs as a hosted Replicate model.

FLUX 1.1 Pro in ultra mode. Up to 4 megapixel images with raw mode for photorealism.

FLUX 1.1 Pro in Ultra mode by Black Forest Labs. Generates up to 4 megapixel images with a raw mode for less processed, more natural-looking photography. Best FLUX option when output resolution and fine detail matter.

Black Forest Labs' development model. Fast, high-quality image generation with LoRA support.

Google DeepMind's Imagen 4 text-to-image model, hosted on Replicate. Sharp detail, accurate text rendering, and strong prompt adherence across photographic and illustrated styles. Outputs up to 2K resolution.

Google's state-of-the-art video generation model. Simulates real-world physics with various visual styles.

Google's Veo 3 served via Replicate. Text-to-video with native synchronized audio generation. High-fidelity motion and scene coherence in short clips.

Latest Veo with image-to-video and context-aware audio

Tencent's HunyuanVideo, a 13B open-weights text-to-video diffusion transformer. Produces high-motion, photorealistic clips with smooth temporal consistency and was one of the first open models to rival closed systems on motion quality.

SDXL fine-tune by galleri5 for slick flat icons and pop constructivist graphics with thick edges. Trained on Bing generations, it produces clean single-subject icon art that suits app icons, badges and UI glyphs. Raster output, not true vector.

The highest-quality tier of Ideogram v3. Improved photorealism and prompt adherence over v2 while keeping Ideogram's best-in-class text rendering. Supports style references and inline text layout.

Whisper Large v3 wrapped with Hugging Face Transformers optimizations (batched inference, flash attention) for very high throughput. Transcribes hours of audio in minutes on a single GPU. Maintained by Vaibhav Srivastav. Good when you need bulk transcription fast.

InstantID makes realistic portraits of a real person from a single reference photo without per-user training. Combines a face encoder with an IdentityNet adapter on SDXL to keep identity and pose while following a text prompt, so it is fast and tuning-free.

Kuaishou's Kling v2.1, generating 5 and 10 second videos at 720p or 1080p from text or an image. Known for cinematic camera work and realistic physical motion, available on Replicate via the official KwaiVGI account.

Kuaishou's premium Kling v2.1 Master. Generates 1080p 5s and 10s clips from text or an image with strong dynamics and prompt adherence. The top tier of the Kling 2.1 family.

Cinematic video up to 15s with multi-shot and native audio

Most versatile: multi-reference images, video editing, native audio

The latest Midjourney model. Industry-leading aesthetic quality and prompt adherence for image generation.

MiniMax Hailuo 02 on Replicate. Text-to-video and image-to-video producing 6s or 10s clips at 768p standard or 1080p pro. Known for accurate real-world physics and stable motion.

Meta's music generation model. Generate up to 1 minute of music from text descriptions.

Turns any single selfie into a clean professional headshot using FLUX Kontext image editing. Keeps the person's face while swapping to business attire, a studio background and even lighting. Aimed at LinkedIn-style profile photos.

Recraft's faster, cheaper vector model. Outputs editable SVG paths instead of raster pixels, so logos, icons and flat illustrations scale to any size without blur. Defaults to a vector_illustration style and supports line art and engraving looks. Hosted API only.

Recraft's text-to-image model that topped the Hugging Face text-to-image arena at release. Strong long-text rendering, brand-style consistency, and precise control over image dimensions and color palettes.

Recraft's raster-to-vector converter. Takes a PNG or JPG and traces it into a clean SVG with precise vector paths, aimed at logos, icons and graphics that need to scale. Image-to-SVG counterpart to Recraft's text-to-SVG models.

Top-ranked for motion quality and visual fidelity

Runway's Gen-4 Turbo on Replicate. Fast image-to-video generation producing 5s and 10s clips at 720p with strong character and scene consistency across shots.

Meta Segment Anything 2. Promptable segmentation across images and video with temporal memory. Zero-shot, point/box/mask prompts, fast on a single H100.

Stability AI's SDXL 1.0 with the optional refiner. The 3.5B base plus 6.6B ensemble UNet that became the default open image model before FLUX. Good for fine-tuning and LoRAs, broad community support.

fofr's sticker generator that outputs graphics with transparent backgrounds, so the result drops straight into chat apps or print sheets. Runs an SDXL-based pipeline at high speed (default 17 steps) and returns die-cut style art without manual background removal.

OpenAI's Whisper running on Replicate. General-purpose speech recognition trained on 680k hours of multilingual audio. Transcribes and translates 99 languages, robust to accents and background noise, and outputs plain text, segments, or word-level timestamps.

Background removal model from 851-Labs that outputs a clean cutout with a transparent alpha channel. One of the most-run background removers on Replicate, handles people, products and objects on busy backgrounds.

Product advertising photo generator. You upload a cut-out product shot and a prompt describing the scene; it places the product on a new generated background with matching lighting and shadows, so a plain packshot becomes an ecommerce or ad-ready hero image without a photo studio.

Plug-and-play motion module that animates personalized Stable Diffusion models without further training. 16-frame clips at 512x512.

ByteDance distillation of AnimateDiff. 4-step sampling for over 10x faster inference at comparable quality to multi-step base model.

Latent-diffusion model for general-purpose text-to-audio. Generates speech, music, and sound effects with a unified prior.

fal.ai's fully open-source 6.8B flow-based text-to-image model. Up to 1536x1536 resolution.

Suno's text-to-audio model. Generates realistic speech, music, and sound effects.

BiRefNet high-resolution dichotomous image segmentation for background removal. Bilateral reference network that produces sharp matting on fine detail like hair, fur and thin structures, often cleaner than older U2Net or rembg models.

BRIA AI's commercial background removal model trained on fully licensed data. Produces accurate cutouts for e-commerce and design, with attention to clean edges around products and people.

BRIA's first commercial-safe background-removal model. Trained on fully-licensed data, suitable for production e-commerce and design pipelines.

BRIA's professional background-removal model trained on fully-licensed data. Commercial-safe.

Microsoft Research pipeline by Ziyu Wan et al. that restores scanned old photos, removing scratches, dust and fading and optionally enhancing faces in one pass.

ByteDance's Seedance 1 Pro on Replicate. Text-to-video and image-to-video producing 5s or 10s clips at 480p or 1080p. Strong motion quality and prompt following from the Seedance family.

catacolabs Cartoonify turns a photo into a flat cartoon illustration. Takes a single image and returns a stylized cartoon version with clean shapes and bold outlines. Straightforward one-input model for avatars and profile pictures.

Content-Consistent Super-Resolution model. Reduces hallucination compared to typical diffusion-based upscalers while keeping perceptual quality high.

Champ controllable human image animation. Uses 3D parametric guidance (SMPL) for realistic full-body motion transfer from a single reference image.

Resemble AI's open Chatterbox TTS. Zero-shot voice cloning from a short audio prompt with an exaggeration control for emotion intensity, plus CFG weight to balance pacing and fidelity.

High-resolution image upscaler with creative detail re-imagination via SD-based hallucination. Strong for photography and product shots.

Meta's 13B Code Llama tuned for instruction following. A faster mid-size option for code generation and completion, supporting infilling for inserting code at a cursor position. Served on Replicate per call.

Meta's 34B Code Llama tuned for instruction following. A balance of size and quality for code generation, completion, and explanation, with strong coverage of Python, JavaScript, and other common languages. Runs on Replicate per call.

Meta's largest Code Llama, a 70B Llama-2 derivative specialized for programming and tuned to follow instructions in chat form. Handles code generation, completion, and explanation across common languages. Served on Replicate as a per-call endpoint.

Meta's smallest Code Llama at 7B parameters, tuned for instruction following. The cheapest and fastest member of the family for quick code generation, completion, and infilling. Served on Replicate per call.

Robust face-restoration model using a transformer-based codebook prior. Handles severe degradation, occlusion, and old-photo restoration with adjustable fidelity-quality tradeoff.

CogVideoX-5B from Tsinghua/Zhipu AI, an open 5B-parameter text-to-video diffusion transformer. Generates 6-second 720p clips with coherent motion and is widely used in research for its open weights and reproducibility.

Zhipu/Tsinghua's 5B open text-to-video model. 720x480 @ 8fps, 6s clips, image-to-video variant available.

Tsinghua CogVLM2 19B with Llama-3 8B base plus 11B vision expert. Strong document understanding and visual reasoning, 8k context.

fofr's model generates the same character in many poses and angles from one reference image. Useful for building an avatar set or character sheet where the face and design stay consistent across outputs. Can produce a grid or individual images.

ControlNet conditioned on Canny edge maps. Preserves composition and outlines while restyling with Stable Diffusion 1.5 or SDXL backbones.

ControlNet conditioned on depth maps. Preserves the 3D scene layout while letting the prompt change style, lighting and content.

DDColor by Xiaoyang Kang et al. colorizes black-and-white photos using dual decoders that jointly learn pixel colors and semantic color queries, giving vivid and natural results on old images.

Quantized GGUF build of DeepSeek's 33B code model, trained on roughly 2T tokens that are about 87 percent code. Designed for repository-level completion and project-aware generation thanks to a 16k context window. Runs on Replicate as a per-call endpoint.

DeepSeek-VL 7B chat model. Vision-language model with hybrid vision encoder and strong real-world visual question answering performance.

Naver CLOVA Donut OCR-free document-understanding transformer. End-to-end JSON extraction from forms, receipts and invoices without explicit OCR.

Rednote Hilab Dots OCR. End-to-end document parsing model with layout, text and reading-order prediction in one transformer.

Generative Gaussian-splatting model for fast image-to-3D synthesis. Produces textured meshes in two minutes via differentiable rasterization.

Tencent DynamiCrafter. Animates still images into short videos preserving texture and structure, with strong open-domain coverage.

JaidedAI EasyOCR. Simple Python OCR wrapper supporting 80+ languages with deep-learning text detection and recognition.

Ant Group EchoMimic. Lifelike audio-driven portrait animation with editable landmark conditioning for fine-grained motion control.

Try-on pipeline aimed at ecommerce listings. You give it a photo containing clothing on a body pose plus a separate face image; it composes a person wearing that clothing with the supplied face, controllable by prompt, CFG, and output size. Useful for generating on-model product shots from a flat garment image.

Enhanced Super-Resolution GAN, the original 2018 architecture. Produces sharp 4x upscales with strong perceptual quality on natural images.

Open-source flow-matching TTS with strong zero-shot voice cloning. Code MIT, weights CC-BY-NC.

F5-TTS, a flow-matching TTS that clones a voice from a reference clip plus its transcript and reads new text in that voice. Fast non-autoregressive synthesis with optional silence removal.

fofr's face stylizer converts a face photo into 3D render, emoji, pixel art, video-game character, claymation or toy styles. Uses InstantID plus style LoRAs on SDXL to keep the likeness while applying a chosen art style. Popular for fun avatars.

fofr's model turns a face photo into a die-cut sticker with a white border and transparent background. Uses InstantID to hold the likeness and outputs a clean PNG suitable for chat stickers or print. Simple single-image input.

Google FILM frame interpolation. Synthesizes high-quality intermediate frames between near-duplicate inputs, designed for large motion gaps.

Microsoft Florence-2 Large. Unified prompt-based vision foundation model for captioning, detection, segmentation and OCR with a single 770M-param backbone.

Microsoft Florence-2 unified vision model with referring expression segmentation. Text-prompted region and mask generation in one model.

PuLID identity customization running on FLUX.1-dev. Inserts a face from one reference photo into prompt-driven scenes using contrastive alignment, giving higher likeness and detail than SDXL-era ID adapters. Good for realistic avatars and character portraits.

The fastest Flux model. Generate images in under 2 seconds. Great for prototyping.

The open-weight 12B rectified-flow transformer from Black Forest Labs. Close to FLUX Pro quality with a guidance-distilled checkpoint released under a non-commercial license. The most widely fine-tuned base in the FLUX family.

The fastest FLUX model from Black Forest Labs, distilled to produce images in 1 to 4 steps. Apache 2.0 licensed for commercial use. Built for high-volume generation and real-time previews.

Black Forest Labs' fastest open-weights image model. Apache-2.0 licensed, ~1-4 step inference.

FLUX structural control via Canny edge maps. Preserve composition while restyling.

Open-weight edge-guided FLUX model from Black Forest Labs. Extracts Canny edges from a control image and regenerates it from your prompt while holding the original composition and outlines, so you can restyle a scene without changing its structure.

FLUX structural control via depth maps. Keep 3D scene layout while changing style/content.

Open-weight depth-guided FLUX model from Black Forest Labs. Derives a depth map from the control image and regenerates from your prompt while preserving 3D spatial layout, useful for re-texturing rooms, products, or scenes without moving objects.

Black Forest Labs' inpainting/outpainting model for FLUX. Fill masked regions with prompt-guided content.

Black Forest Labs' open-weight inpainting and outpainting model, guidance-distilled from FLUX.1 Fill [pro]. You supply an image plus a mask and a prompt; it fills the masked region or extends the canvas with prompt-guided content that matches lighting and texture.

Open-weight version of FLUX.1 Kontext by Black Forest Labs. Instruction-based editing: pass an input image and a plain text edit ('change the jacket to red', 'remove the person on the left') and it applies the change while keeping the rest of the scene and identity consistent.

FLUX image-variation adapter. Generate variations and remixes from a reference image.

FLUX.1-dev inpainting wrapper that fills masked parts of an image from a prompt. Useful when you want FLUX-quality fills with a simple image plus mask plus prompt interface and adjustable mask strength.

Tencent ARC face-restoration GAN. Reconstructs realistic facial detail in low-quality or compressed photos using a pretrained StyleGAN2 prior.

Global-Local Path Networks depth-estimation model. Combines hierarchical transformer encoder with selective feature fusion for sharp boundaries.

Faster cheaper Veo 3 with audio

Faster Veo 3.1 with image-to-video and audio

StepFun GOT-OCR 2.0. Unified end-to-end OCR-2.0 model handling text, formulas, charts, sheet music and geometric shapes in one architecture.

IBM Granite 20B Code Instruct. Larger Granite code model balancing quality and inference cost for enterprise CI/CD code-review automation.

IBM Granite 8B Code Instruct. Trained on permissively-licensed code, strong on multi-language code completion and instruction-following.

xAI video with native audio and lip-sync, up to 15s

Grounding DINO plus SAM. Open-vocabulary text-prompted detection and segmentation in one pipeline for fully-automatic mask generation.

Minimax model for realistic human motion and VFX

Tencent's Hunyuan3D 2.0 image-to-3D pipeline. Two-stage shape and texture generation producing high-resolution textured meshes.

Refreshed Hunyuan3D 2.1 with improved texture fidelity and PBR-material support. Image-to-3D with textured GLB output.

Tencent's 13B open-weights video diffusion transformer. SOTA among open video models at release.

Lvmin Zhang's IC-Light packaged by zsxkib. Relights a product or portrait from a text prompt or a chosen light direction while keeping the subject's shape and detail intact, so a flat product photo can be given studio, window, or dramatic side lighting without re-shooting.

Hugging Face Idefics3 8B. Llama-3 based open-source vision-language model with strong document QA and chart-understanding performance.

Ideogram's text-to-image model known for accurate in-image text and typography. Handles posters, logos, and signage where other models garble lettering. Supports magic prompt expansion and multiple aspect ratios.

Ideogram's fast v3 model, the fastest and cheapest tier of the v3 family. Known for accurate in-image text rendering and reliable typography, which most diffusion models still get wrong. Hosted API only.

IDM-VTON virtual try-on from the CVPR 2024 paper. You give it a photo of a person and a garment image; it dresses the person in that garment while preserving pose, body shape, and the garment's pattern and text. Good for showing a clothing product on a model for an ecommerce listing.

Image-to-3D mesh generator from sparse-view diffusion. Produces textured meshes in under one minute on a single A100.

Berkeley InstructPix2Pix. Edits an image from natural-language instructions in a single forward pass. Trained on GPT-3 plus Stable Diffusion synthetic pairs.

Tencent's face-identity conditioning adapter for SD/SDXL. Face embedding + CLIP for ID-consistent generation.

DeepSeek's unified multimodal model. Decouples vision encoding for both understanding and generation tasks.

Kuaishou's Kling v1.6 Pro on Replicate. Generates 5s and 10s clips in 1080p from text or an image, with cinematic motion and physics realism. The widely used pro tier of the 1.6 generation.

Open-weights 82M-parameter TTS. Punches above its size class on naturalness benchmarks at a fraction of the inference cost of larger models.

Kuaishou's bilingual (CN/EN) latent diffusion text-to-image model with strong text rendering.

Kuaishou LivePortrait. Efficient portrait animation driven by reference videos with stitching, retargeting and motion-control parameters.

Meta Llama 3.2 11B Vision served via Ollama on Replicate. Open-weights multimodal model for image captioning, document and chart reading, and visual question answering.

Meta Llama 3.2 90B Vision. Largest open-weights Llama vision model. Strong visual reasoning, chart, OCR and document understanding.

LLaVA 1.6 (LLaVA-NeXT) with a Vicuna-13B language backbone. Open vision-language chat model that describes images, answers questions, reads charts and reasons about scenes. Version 1.6 adds higher input resolution and better OCR and reasoning than LLaVA 1.5.

LLaVA v1.6 on a Nous-Hermes-2 34B base, served on Replicate. Open-source vision-language assistant for image question answering, description and visual reasoning at higher resolution.

SDXL fine-tune by mejiabrayan aimed at logo generation. Produces simple, centered mark and wordmark style logos from a text prompt. Useful for quick brand concepts and mockups. Raster PNG output, not vector.

Lotus generative depth model. Treats depth as a generation task using a diffusion model, producing higher-fidelity depth on textured surfaces.

Lightricks' 2B DiT video model. Realtime generation on consumer GPUs (~6s @ H100, 24fps).

Fast affordable video with I2V support

Luma Labs' Ray-2 at 720p on Replicate. Text and image-to-video producing 5s and 9s clips with fast, coherent motion and strong camera control. Successor to Dream Machine.

ByteDance MagicAnimate. Temporally consistent human-image animation driven by a DensePose motion sequence with strong identity preservation.

UIUC Magicoder S CL 7B. CodeLlama-7B fine-tuned with OSS-Instruct synthetic data. Strong HumanEval Plus and MBPP Plus performance per parameter.

MAGNeT is Meta's masked, non-autoregressive audio generator. Instead of predicting tokens left to right it fills masked audio tokens in parallel over a few decoding steps, so generation is faster than autoregressive MusicGen at similar quality. This Replicate packaging exposes the text-to-music and text-to-sound variants.

Meta MAGNeT non-autoregressive music generator. Up to 7x faster than MusicGen with comparable quality via masked generative transformers.

Detail-hallucinating upscaler in the Magnific style. Adds plausible high-frequency texture using a Stable Diffusion refiner conditioned on the low-res input.

ETH Zurich Marigold. Diffusion-based monocular depth-estimation model fine-tuned from Stable Diffusion with strong fine-detail recovery.

Marker PDF-to-Markdown conversion pipeline. Combines layout, OCR and equation models to produce clean Markdown with preserved tables and formulas.

Meta Mask2Former universal image-segmentation transformer. Single architecture for panoptic, instance and semantic segmentation tasks.

Intel MiDaS v3.1 relative depth-estimation model. Robust zero-shot single-image depth across diverse domains and resolutions.

OpenBMB MiniCPM-V 2.6. 8B vision-language model with strong single-image, multi-image and video understanding plus OCR capabilities.

MiniMax's video generation model. Fast, high-quality video output with text-to-video capabilities.

Genmo's Mochi 1, an open text-to-video model with high-fidelity motion built on a 10B Asymmetric Diffusion Transformer. Released under Apache 2.0, it was the largest open video model at launch and is strong on smooth, physically plausible movement.

Genmo's 10B open-weights text-to-video model. AsymmDiT architecture, 5.4s @ 480p.

Allen AI Molmo 7B-D on Replicate. Open vision-language model trained on the PixMo data, notable for pointing at and locating objects in images, not just describing them.

Moondream2 small vision-language model on Replicate. About 1.9B params, designed to run on edge devices, handles captioning, visual QA and short OCR-style reads at very low cost.

Tencent MuseTalk real-time lip-sync model. Audio-driven mouth-region editing in latent space at 30+ fps on a single GPU.

Meta's 3.3B-parameter MusicGen Large. Text-conditioned music generation with single-stage autoregressive transformer, supports melody conditioning.

Allen AI olmOCR. Open-source 7B vision-language model fine-tuned for high-fidelity document parsing including math, code and tables.

OOTDiffusion virtual try-on. Takes a clear photo of a model and an upper-body garment and renders the garment onto the person using an outfitting-fusion diffusion approach that keeps the garment's texture and the model's pose. A lightweight alternative to IDM-VTON for clothing previews.

CMU OpenPose multi-person 2D pose estimator. Real-time keypoint detection for body, hand, face and foot using Part Affinity Fields.

MyShell OpenVoice v2. Multilingual zero-shot voice cloning with accurate tone-color reproduction and style/emotion control.

Baidu PaddleOCR v3 PP-OCR pipeline. Lightweight detector plus recognizer optimized for production use with 80+ language support.

Hugging Face Parler-TTS Mini. Lightweight TTS conditioned on a natural-language style description for fine-grained control over voice characteristics.

Phind CodeLlama 34B v2. Highly tuned CodeLlama variant focused on retrieval-augmented developer assistant workflows.

Tencent ARC PhotoMaker. Identity-preserving stylized photo generation from a stacked-ID embedding. Realistic re-styling of a subject in seconds.

Physics-accurate video generation up to 1080p

Playground AI's diffusion model tuned for aesthetics. SDXL-based architecture trained on the EDM formulation, rated by users as more visually pleasing than SDXL in their study. Strong on vivid color and contrast.

OpenAI Point-E text-to-point-cloud system. Fast 3D point-cloud generation from text, optionally lifted to a mesh via marching cubes.

Alibaba Qwen's instruction-driven image editor. Extends Qwen-Image's text-rendering ability to editing, so it handles both semantic edits (swap objects, change style) and precise text edits inside the image while preserving the original layout and unedited regions.

Alibaba Qwen2-VL 7B served on Replicate. Open-weights vision-language model that chats about images and video, with dynamic resolution and strong OCR and document QA for its size.

AI-Upscaler that increases image resolution up to 4x while preserving texture and detail. Trained on synthetic and real data to reduce common ESRGAN artifacts.

Real-ESRGAN variant fine-tuned for anime, manga, and illustrated artwork. 4x upscaling with cartoon-aware artifact suppression.

State-of-the-art image generation optimized for design and branding. SVG vector output support.

Recraft's v3 variant that outputs vector SVG instead of raster pixels. Generates clean, editable logos, icons and illustrations that scale without quality loss, which is unusual among image models. Hosted API only.

Recraft V4 SVG turns a text prompt into production-ready SVG vector art with clean geometry and structured, editable layers. Newer generation than V3 with improved design quality on logos, icons and flat illustration. Returns true vector paths, not a traced bitmap.

Open-source background-removal tool wrapping U2Net. Produces alpha mattes for photos, products and people with no manual masking.

Lucataco's remove-bg, a rembg-based background removal model that returns the foreground subject on a transparent background. A popular, low-cost option for quick product and portrait cutouts.

Object removal and cleanup using LaMa inpainting. Paint a mask over an unwanted object, logo or person and the model fills the area with plausible background, erasing it from the photo.

Replit's 3B code-completion model, trained on a permissively licensed code subset of the Stack across 20 programming languages. Built for low-latency autocomplete rather than chat. Served on Replicate per call.

Real-Time Intermediate Flow Estimation. Doubles or quadruples FPS of an existing video via learned optical-flow-based frame interpolation.

Stable-Diffusion-based real-time music generator. Operates on spectrogram images then resynthesizes audio, enables seamless transitions and looping.

Retrieval-based Voice Conversion. Converts a source recording into a target speaker's voice, preserving pitch, prosody and rhythm.

Stylized audio-driven talking-head generator. Synthesizes 3D motion coefficients from audio to animate a single portrait image with natural head movements.

SDXL fine-tune by fofr trained on Apple emoji art. Generates rounded, glossy emoji and icon style graphics from a text prompt, useful for custom reactions, app glyphs and playful icon sets. Raster output.

SDXL inpainting built on the Hugging Face Diffusers inpaint pipeline. Replace or remove masked regions of an image with prompt-conditioned content at SDXL resolution. A cheap, well-understood baseline for object removal and local edits.

Meta's SeamlessM4T multimodal translation model. Takes speech or text input and produces transcription or translation across about 100 languages, including speech-to-text and speech-to-speech. One model covers ASR plus cross-lingual translation without chaining separate systems.

Meta SeamlessM4T v2 Large speech mode. Speech-to-speech, speech-to-text, and text-to-speech translation across 100+ languages in a single unified model.

Meta SeamlessM4T v2 Large. Universal multilingual translation across 100+ languages with text-to-text mode for documents and chat.

Budget ByteDance video, fast and cheap

ByteDance video with T2V and I2V, up to 1080p

NVIDIA SegFormer-B5 semantic segmentation. Hierarchical transformer encoder with lightweight MLP decoder, strong ADE20k and Cityscapes results.

OpenAI Shap-E text/image to 3D. Generates implicit neural representations renderable as textured meshes or NeRFs.

Spark efficient TTS with disentangled control over speaker, content and style. Strong cross-lingual zero-shot performance.

Stability AI's Stable Audio Open generates short audio from text prompts, tuned for sound effects, drum loops, instrument riffs and production elements rather than full songs. Open weights, latent diffusion over a 44.1kHz audio autoencoder, with a configurable seconds_total up to about 47 seconds.

Stability AI's 8B MMDiT-based flagship. Open weights at 1MP with improved typography and prompt adherence over SDXL. The largest model in the SD 3.5 release line.

Distilled, 4-step version of SD 3.5 Large from Stability AI. Keeps most of the large model's quality and text rendering at a fraction of the inference time. Open weights under the Stability Community License.

Stability AI's SDXL model via Replicate. High-quality image generation with extensive customization.

BigCode StarCoder2 15B code-generation flagship. Trained on 4T tokens of Stack v2 data with grouped-query attention and 16k context.

StarVector 8B is a multimodal model that generates SVG code directly from an input image. Rather than tracing pixels, it predicts the SVG markup token by token, which can produce compact, semantically structured paths for icons and simple graphics. Research model from the StarVector project.

Picsart StreamingT2V. Generates long, consistent videos by chaining short autoregressive clips with motion and appearance memory.

Style-based TTS using diffusion and adversarial training. Human-level naturalness in zero-shot voice synthesis from a 3-5s reference clip.

Suno's text-prompted generative audio model. Speech, music, ambient sound and effects with non-verbal cues like laughter or sighs.

SUPIR by Fanghua Yu et al. is a large diffusion-based restoration model that recovers photorealistic detail from heavily degraded images and can be steered with a text prompt describing the scene.

SUPIR (Scaling-Up Image Restoration) photo-real restoration model. Combines SDXL prior with language-guided controls for severely degraded inputs.

Transformer-based image super-resolution using Swin-V2 attention. Handles classical, lightweight, real-world, and compressed-input variants with 2x/4x upscaling.

SwinIR transformer-based super-resolution and denoising applied per-frame to video. Handles classic, real-world and lightweight upscaling.

Tencent ToonCrafter generative cartoon interpolation model. Synthesizes smooth in-between frames between two cartoon keyframes.

Multi-voice expressive TTS. Slow but high-quality with strong prosody and natural intonation. Trained for long-form narration use cases.

Microsoft TRELLIS image-to-3D model. Generates textured 3D assets in GLB or Gaussian-splat format from a single reference image.

Stability AI and Tripo single-image 3D reconstruction model. Generates 3D meshes from a single image in roughly half a second.

AI music generation with studio-quality output. Generate full songs with vocals, instruments, and production.

Tencent V-Express. Audio-driven portrait animation with progressive training, weak-condition learning, and expressive lip sync.

PNG/JPG to SVG vectorizer built on VTracer, the open-source raster-to-vector engine. Traces a bitmap into layered color regions and clean paths with controls for color count, area threshold and path simplification. Fast, deterministic alternative to model-based vectorizers.

Tencent VideoCrafter latent video diffusion. Text-to-video and image-to-video generation up to 2s at 1024x576 with strong motion fidelity.

Alibaba's Wan 2.1 open-weights video diffusion model. 14B MoE-based, supports T2V and I2V.

Image-to-video variant of Alibaba's Wan 2.1 14B at 720p, accelerated by WaveSpeedAI. Animates a still input image into a short clip driven by a text prompt, keeping the source composition while adding motion.

Accelerated inference for Alibaba's Wan 2.1 14B text-to-video at 720p, hosted by WaveSpeedAI on Replicate. Open suite of video foundation models with high-resolution output and faster generation.

Ultra-cheap I2V. Upload image and animate it.

Ultra-cheap T2V for pennies

Lip-sync model that re-syncs a target video's lip movement to an arbitrary audio track. Robust to identity and language with a lip-sync discriminator loss.

Whisper Large v3 Turbo combined with pyannote 4.0 for speaker diarization, returning who-said-what segments with timestamps. Built by Thomas Mol. Returns a clean JSON of speaker-labeled segments, handy for meeting notes, interviews, and podcasts.

WhisperX (Large v3) with forced alignment for accurate word-level timestamps plus optional speaker diarization. Uses VAD to cut long files into segments and a wav2vec2 aligner to pin each word to its exact time. Useful for subtitles and per-speaker transcripts.

WizardLM WizardCoder 33B v1.1. Evol-Instruct fine-tune of DeepSeek-Coder-33B with strong code-generation benchmark performance.

Coqui's XTTS v2 multilingual TTS with voice cloning from 6 seconds of reference audio. Supports 17 languages and emotion transfer.

01.AI Yi-VL 34B vision-language model. Bilingual (CN/EN) image understanding, strong CMMMU and MMMU performance among open-weights VLMs.

Intel ZoeDepth metric depth-estimation model. Combines relative-depth pretraining with metric fine-tuning for absolute distance in real units.