Coding agents can write code that renders a film or a game, but they can't hear the result. They can't tell a cheap key click from a good one, or notice a hit that lands 80 ms late. Earshot turns audio into text, numbers and images an agent can reason about.

It uses local models on CPU plus model-free measurements such as loudness and spectrum. It can describe a file, check how well it fits a description, find the best moment in a long recording, and tell a real recording from code-made sound. It can also produce a second-by-second listening transcript of a film, an annotated image, and a scorecard against seven professional launch films.

It also lets the agent make sound. forge writes sounds as synthesis recipes and improves them using a critic's feedback. fit turns a recording into editable code. A bank of 252 code-made effects renders by id and can be timed to the picture. The README is open about limits: the search can overfit earshot's own models, and its blind tests used one listener.

Read more on interesting repos · GitHub

Desktop app that enlarges low-resolution images using AI, on Linux, macOS and Windows

Small or pixelated images look rough when you enlarge them. Upscayl takes such an image and uses an AI model to guess what the missing detail should look like, then produces a larger version.

It is an Electron app built in TypeScript. The upscaling itself is done by a separate open-source backend, upscayl-ncnn, which runs Real-ESRGAN models through Vulkan. That means the work happens on your own graphics card, not on a server. The README says you need a Vulkan-compatible GPU, and many integrated GPUs will not work.

The README is clear about limits: it can enhance low-resolution and pixelated images, but it cannot fix blur or bad focus. It ships with a few models and lets you add more from a separate custom-models repository or convert your own. The README compares itself to Topaz Gigapixel only through its topic tags, so check the before-and-after examples it links to judge quality.

Read more on interesting repos · GitHub

A local web app for making and editing images with Stable Diffusion and Flux models

Invoke is an image-generation app you install on your own computer. It starts a local web server and a React interface in your browser, where you write prompts, pick a model and make images. It is free to use under the Apache-2.0 license, which allows commercial use.

Beyond plain text-to-image, it has a Unified Canvas for painting, inpainting (filling in part of an image) and outpainting (extending an image past its edges). It also has a node-based workflow editor, where you connect steps into a custom generation pipeline and share it with others.

The README lists many models: Stable Diffusion 1.5, 2.0, SDXL and 3.5, the Flux family, Qwen Image, Z-Image and more. A few, such as Nano Banana, GPT Image and Wan, work through an API only. It also includes a model manager, a gallery that keeps the prompts and settings behind each image, and upscaling tools.

Read more on interesting repos · GitHub

ACE-Step 1.5 generates music from text. You describe a song, or give lyrics and style tags, and it produces audio from 10 seconds up to 10 minutes long. It runs locally, so you don't need a paid music service. The project also offers a free hosted version at acemusic.ai for people without a GPU.

It works in two stages. A language model plans the song first, expanding a short prompt into a blueprint with metadata, lyrics and a caption. A diffusion transformer (DiT) then turns that plan into audio. The README says quality lands between Suno v4.5 and Suno v5 on its own metrics. Those are the authors' claims, so listen to the output yourself.

Beyond plain generation, it edits audio. It can make covers, repaint sections of a track, split stems, and add accompaniment to a vocal. You can also train a LoRA on a few of your own songs to capture your style. Model sizes run from 2B to 4B (XL), so you can pick one that fits your GPU memory.

Read more on interesting repos · GitHub

Self-hosted tool that cuts long videos into 9:16 shorts with captions and dubbing

Making short vertical clips from a long recording is slow work. You find the good moments, reframe the picture for a phone, add captions and post it. OpenShorts does those steps in one dashboard. You upload a video, and it transcribes it, picks 3 to 15 moments, crops each one to 9:16 and burns in subtitles.

Google Gemini reads the transcript and scene boundaries to choose moments. faster-whisper makes word-level subtitles, and MediaPipe and YOLOv8 track faces for the crop. There are layouts for one speaker, two speakers stacked with captions between them, and screen shares. If you don't want a Google key, you can point it at a local model such as Ollama through an OpenAI-compatible URL.

The README compares it to hosted clip tools like Opus Clip, Vizard and Klap. The main difference is that you can self-host it under the MIT license with no watermark or upload limit. It also bundles an AI-actor video maker and a YouTube toolkit. The same software is offered hosted at openshorts.app, with GPU speed, from $12 a month.

Read more on interesting repos · GitHub

Desktop app for cloning voices, dubbing video and dictating, run on your own machine

VoiceStudio puts several speech tasks in one app: text to speech, voice cloning, voice design, video dubbing, dictation, transcription and audiobook creation. The project says it covers 646 languages. The default engine is k2-fsa/OmniVoice, and you can switch to other engines listed in the feature catalog.

Most work runs on your own computer. It uses NVIDIA CUDA on Windows and Linux, Apple Silicon acceleration on Macs, and falls back to the CPU on machines without a GPU, which is slower. Remote workers and services are optional, and usage analytics only turn on with your consent.

It positions itself as a local, open-source alternative to ElevenLabs. It also offers a local API and an MCP server so coding agents and other programs can call it. The code is AGPL-3.0, and each model has its own license, so check them before commercial use.

Read more on interesting repos · GitHub

This repo holds the code behind a music video set to the song "I'm Upping My P(doom)". Claude Opus 5.5 wrote everything in it, working in Claude Code. The only direction was to use the Clawd character design and give each lyric interesting visuals and transitions.

Each frame is painted in a web page (studio.html) using p5.js and p5.brush for brushstroke effects. The video is split into nine chapters, one file each, plus shared code for characters, props, lyrics and the timeline. A Node script, render.mjs, opens the page in headless Chrome, saves frames, then uses ffmpeg to join them with the song.

The repo also keeps the working documents: a storyboard Opus wrote, and an animation guide it wrote to brief parallel subagents. A first-generation version lives in legacy/. The README links to a separate, more general base for Claude animations by the same author.

Read more on interesting repos · GitHub

Running a vision-language model on a live camera feed is slow, and it is hard to know which tweak actually helps. Glance Speedlab is a research harness for that. It grabs the newest camera frame only when the model is ready, sends one request with several questions, and reports capture, transport, image-prefix, scoring and full-loop latency separately.

It sits on top of Glance, a local layer that turns open vision-language models into typed yes/no decisions. Speedlab keeps a PyTorch MPS reference path next to an optional MLX 8-bit scorer, so you can switch between them in the same UI and compare.

Its main value is the written record. Each change starts as a preregistered experiment with a hypothesis and a quality guardrail, and the README tables what was kept and rejected. Results cover one Apple M5 laptop, Qwen3-VL-2B and 4B, and a small fixed test suite. It is a research prototype, not a production monitor.

Read more on interesting repos · GitHub

Editing video with an agent is hard because the agent cannot see or hear the footage, and most editors are not built to be driven by code. Diffusion Studio stores each project as a folder of JSX files. The agent edits the code, and the editor canvas redraws.

It works in both directions. If you drag a shape on the canvas or trim a clip on the timeline, the change is written back to the code on the element that owns it, so the agent always sees the current version. The README likens it to an IDE that renders a video canvas instead of text.

The desktop app ships a `diffusion` command-line tool, which is also exposed as MCP tools. They let an agent probe media, grab frames, view waveforms, transcribe speech, ask a multimodal model about a clip, and check a composition for mistakes like black-frame gaps. It works with Claude Code, Codex, Cursor, Copilot, Gemini CLI, OpenCode or Pi.

Read more on interesting repos · GitHub

Getting an image model to give good results twice is hard. Most people write loose prose prompts and get uneven output. This project collects community examples and rewrites them as structured prompts, so you can see what was asked and what came out.

Each case splits a prompt into parts such as subject, lighting, materials, layout and text. The author calls this Prompt as Code. Parts like these are easier to reuse in scripts, templates, batch jobs and agents than one long paragraph. Cases are sorted into groups like UI, posters, products, photography, illustration and characters.

There is also a website for browsing large previews, copying prompts and filtering by style, plus a section comparing GPT Image 2.5 output on shared prompts. The README says that section does not add live generation, and that original generation conditions for some cases are unverified.

Read more on interesting repos · GitHub

TLDR publishes 14 newsletters, and reading them all takes time. TLDR Radio turns each one into an audio episode you can play while cooking or driving. You pick a date, press Broadcast, and a few minutes later the episodes are in your browser, one per newsletter, with each story as its own chapter.

It runs as two Docker containers: a FastAPI app with a plain JavaScript frontend, and Kokoro-82M speech synthesis on the CPU. The pipeline is fixed from start to finish. It fetches the archive page, parses it by page structure, fills in a script from templates and a pronunciation dictionary, and sends that script to the speech model. No language model summarizes or rewrites anything.

The README contrasts it with listen-to-your-newsletter tools that need an account, a subscription and access to your inbox. This one needs only Docker and about 8 GB of disk. It is meant for personal use, and the README asks you not to publish the audio.

Read more on interesting repos · GitHub

Some of the best image models are mostly used through web pages. Imagine gives you a command, imagine -p "your prompt", that calls them directly through their APIs and saves the result as a PNG in your current folder. It supports Gemini, Vertex AI and OpenAI.

Setup is one config file at ~/.config/imagine/config.yaml, with no environment variables needed. You can add a provider with imagine providers add. For OpenAI you can use an API key, or sign in with a ChatGPT Plus, Pro or Team plan. The README says the subscription route uses an unpublished endpoint, so it may stop working. The API key is the stable fallback.

The same command generates or edits: add -i reference.png and it edits that image instead. You can also pass a YAML or JSON file to -p to run many jobs in parallel, each with its own prompt, provider and size. The README names its predecessor, banana-cli, which only supported Google's models. Imagine is built so new providers can be added as a directory under providers/.

Read more on interesting repos · GitHub

RoomGPT takes a photo of a room and returns a redesigned version of it. It is the early, open-source version of the paid RoomGPT.io product, with login, payments and extra features removed.

You upload a photo, and a Next.js API route sends it to ControlNet, a model that generates new images guided by the structure of an input image. The model is hosted on Replicate, so you need a Replicate API key but no GPU. Bytescale stores the images.

Because the app is minimal, it works well as a reference for building an image-generation product on hosted models. Rate limiting is optional and uses an Upstash Redis database.

Read more on interesting repos · GitHub

Making a video usually means juggling many tools: research, script, voice, images, clips, music, captions and editing. OpenMontage puts that work behind your AI coding assistant. You describe the video in plain language, and the agent follows production pipelines and guide files stored in the repo.

It works with Claude Code, Cursor, Copilot, Windsurf and Codex. The agent calls tools for image and video generation, text-to-speech, music, FFmpeg and Remotion composition. The README says provider choices are scored and logged, and you approve the creative decisions along the way.

Unlike tools that only animate a few still images, it can also build a video from free stock footage and open archives. It can also start from a reference video (YouTube, Short, Reel, TikTok or a local clip) and propose similar concepts with cost estimates.

Read more on interesting repos · GitHub

Making video by hand in an editor is slow to repeat and hard for an AI agent to drive. HyperFrames takes a different route: you describe a video as an HTML page with timing attributes, and it renders that page to an MP4. The output is deterministic, so the same input gives the same video each time.

Animations have to be seekable, meaning the renderer can jump to any moment and get the right frame. The README lists GSAP, Lottie, Three.js, Anime.js, CSS and WAAPI as supported runtimes. The repo topics point to Puppeteer and FFmpeg for capture and encoding, though the README excerpt doesn't say so directly.

Its main twist is a set of 21 skills that teach coding agents such as Claude Code, Codex, Cursor and Gemini CLI the loop of planning, writing HTML, linting, previewing and rendering. A router skill picks a workflow for requests like product launch videos, explainers, PR summaries, captions or slide decks. There is also a one-way port from Remotion, a React-based video framework.

Read more on interesting repos · GitHub

Making a short video by hand means writing a script, finding clips, recording a voiceover, timing subtitles and editing it all together. MoneyPrinterTurbo chains those steps into one automated workflow. You give it a topic or keyword, and it produces a finished video.

A language model writes or rewrites the script and picks search keywords. Footage comes from stock libraries such as Pexels, Pixabay and Coverr, from your own files, or from AI video and image services. A text-to-speech service reads the script, subtitles are generated, music is added, and the pieces are combined. Every stage can be replaced with your own content.

It works with many model providers, including OpenAI, Claude, Gemini, DeepSeek and local Ollama. You can use it through a web UI, an API, a command line or an agent. The README does not compare it with other tools by name.

Read more on interesting repos · GitHub

Supertonic turns text into speech on your own machine. It uses ONNX Runtime, so there are no cloud calls and no API keys. The model has 99 million parameters, which is small for a TTS system, so downloads are modest and it starts quickly. It needs no GPU.

You download the model files (ONNX models, config and preset voice styles) from Hugging Face, then run an example in your language of choice. The repo has examples for Python, Node.js, browser (WebGPU), Java, C++, C#, Go, Swift, iOS, Rust and Flutter. Output is 44.1kHz 16-bit WAV audio. It supports 31 languages, and you can pass lang="na" when you don't know the input language.

The README compares it with larger open TTS models such as VoxCPM2, OmniVoice and Qwen3-TTS, and reports competitive error rates on the Minimax-MLS-test benchmark. Those numbers are the authors' own. The repo is archived: the authors say there will be no updates, fixes or support, and the hosted demos and Voice Builder service are not part of it.

Read more on interesting repos · GitHub

Running diffusion models usually means writing scripts or using a fixed web form. ComfyUI replaces that with a canvas where each step (load a model, encode a prompt, sample, upscale, save) is a node, and you connect nodes with wires to build a workflow.

Because the workflow is a graph, you can reuse parts, swap models, and re-run only the pieces that changed. The README says it manages VRAM and RAM carefully, offloads models, and supports quantized ones, so large open models can run on modest GPUs; it claims as little as 4GB VRAM with 8GB RAM. That is the project's own claim, so expect speed to vary.

It supports many open models natively, covering Stable Diffusion, SDXL, Flux, Wan, LTX-Video and others. Optional paid partner nodes reach closed models, and an --offline flag turns those off. Workflows save as JSON, can be exposed as simple apps through App Mode, and can be called through a local API. It runs on Windows, Linux and macOS, with a desktop app, a manual install, and a paid cloud option.

Read more on interesting repos · GitHub

Making a short narrated video normally means writing a script, finding visuals, producing a voice and editing it all together. Pixelle-Video does those steps from a web page. You type a topic, or paste your own script, and it runs the whole chain.

The pipeline has four stages: write the script, plan the images, process each frame, then compose the video. A language model writes the narration, an image or video model makes the visuals, a text-to-speech engine reads it aloud, and HTML templates lay it out in portrait, landscape or square.

Each stage can be swapped. You can use ComfyUI workflows (local or on RunningHub) or call model APIs directly, such as DashScope, OpenAI, Seedream, Seedance and Kling. The README says it was inspired by MoneyPrinterTurbo, NarratoAI and MoneyPrinterPlus.

Read more on interesting repos · GitHub

Local voice studio for cloning voices, generating speech and dictating into any app

Voicebox puts both halves of voice work in one app. Output means turning text into speech, including in a cloned voice. Input means speaking and getting text back. The README presents it as a free, open-source alternative to ElevenLabs for output and WisprFlow for dictation.

You give it a short audio sample, and it clones the voice using one of seven bundled TTS engines, such as Qwen3-TTS, Chatterbox, Kokoro and HumeAI TADA. Each engine has different strengths in language coverage, speed and size, and you can switch between them per generation. Dictation uses a global hotkey and OpenAI's Whisper. An optional bundled local LLM can tidy up the transcript before it is pasted.

It also has a REST API and a built-in MCP server. An agent such as Claude Code, Cursor or Cline can call a single tool, `voicebox_speak`, and talk to you in a voice you chose. The app is built with Tauri rather than Electron. It runs on Apple Silicon (MLX), Windows (CUDA), AMD ROCm, Intel Arc and Docker. Linux binaries are not yet available, so Linux users have to build from source.

Read more on interesting repos · GitHub

VoxCPM2 turns text into speech. Many TTS systems first convert audio into discrete tokens and then predict those tokens. This one skips that step. It generates continuous speech representations directly, using a diffusion autoregressive design built on the MiniCPM-4 language model backbone.

The model has 2 billion parameters and was trained on over 2 million hours of speech. You give it text in any of 30 languages, with no language tag. You can put a short description in parentheses before the text, such as a young woman with a gentle voice, and it invents a matching voice. Or you can give it a reference clip and it copies that voice, with an optional transcript for a closer match.

It accepts 16kHz reference audio and outputs 48kHz audio, with no separate upsampler. The README reports a real-time factor of about 0.3 on an RTX 4090, and about 0.13 with the Nano-vLLM or vLLM-Omni serving engines. Weights and code are under Apache-2.0.

Read more on interesting repos · GitHub

Whisper turns spoken audio into text. It is a single Transformer model trained on a large, varied set of audio. One model handles several jobs: multilingual speech recognition, speech translation, language identification and voice activity detection.

It splits audio into 30-second windows and predicts text token by token for each window. Special tokens tell the decoder which task to do, such as transcribe or translate. This lets one model replace several stages of a traditional speech-processing pipeline.

You pick a model size to fit your hardware. Tiny needs about 1 GB of VRAM and runs fast. Large needs about 10 GB and is more accurate. Turbo is an optimized large-v3 that is faster with a small loss in accuracy. Code and weights are under the MIT license.

Read more on interesting repos · GitHub

agent-media is a plugin for AI agents that makes short vertical videos in the style of user-generated content (UGC). You give it a script and say who should speak: a text description, a photo, or a saved character. It returns a finished MP4 with the person speaking your lines, lip-synced.

It works through a hosted API. Your agent calls one tool, make_ugc, which starts a job. You poll the job until it succeeds and read the video URL from the result. Short scripts become one clip. Long scripts are split into several takes and joined, and the README says they are never trimmed. A b-roll URL gives a narrated overlay. Captions are off unless you ask for them.

You can reach it as a Claude Code plugin, through the skills installer, as an MCP server, over plain REST, or with a CLI. A second skill, make_podcast, puts two saved characters in one room and cuts between them as they talk. The repo is a mirror of a private monorepo, so it holds skills and docs, not the generation code. Each run costs credits bought on the agent-media site.

Read more on interesting repos · GitHub