img2threejs/img2threejs
Agent skill for Claude Code and Codex that rebuilds the object in a reference image as procedural Three.js TypeScript code through a gated, vision-reviewed, token-efficient pipeline.
Awesome AI Agents › Image Processing & Analysis Agents
LYL1015/JarvisEvo
JarvisEvo is the released implementation of a CVPR 2026 paper on a photo retouching agent that improves itself through joint optimisation of an editor and an evaluator. The agent is built on a multimodal large language model, published as an 8 billion parameter checkpoint of roughly 17 gigabytes, that reads a source photograph together with a natural language editing request and produces a sequence of concrete retouching operations. Those operations are applied through Adobe Lightroom rather than by generating pixels directly, so the result is a set of interpretable, adjustable edits in a real editing tool instead of an opaque image synthesis. The repository releases an agent-to-Lightroom protocol, a client and server design that lets training and inference run across several machines and GPUs while talking to distributed Lightroom instances, which is what makes reinforcement style training against a real editor practical. Training follows several stages: supervised fine tuning code is released, while the synergistic editor policy optimisation and reinforcement fine tuning stages are listed as still to come. The evaluator half of the system scores candidate edits and supplies the reward signal that lets the editor keep improving without a human rating every attempt. Alongside the model the authors release ArtEdit-Bench, an evaluation dataset of about one gigabyte for measuring instruction-following photo editing quality, and provide inference code, checkpoints and a project page with visual comparisons. The work follows the earlier JarvisArt retouching agent from the same group. It is aimed at researchers working on multimodal agents, tool use over professional software, and reward modelling for creative tasks, and at practitioners interested in agents that operate existing editing applications.
https://github.com/LYL1015/JarvisEvo
Also in Image Processing & Analysis Agents
apple/ml-ferret
Ferret is an end-to-end multimodal large language model developed by Apple that excels in fine-grained referring and grounding tasks with open vocabulary, supported by a large-scale dataset and evaluation benchmark for research purposes.
THUDM/CogVLM
CogVLM and CogAgent are state-of-the-art open-source visual language models designed for advanced image understanding, multi-turn dialogue, and GUI agent capabilities, achieving top performance on multiple cross-modal benchmarks.
11cafe/jaaz
Jaaz is a local and free AI design agent that enables users to design, edit, and generate images, posters, and storyboards with advanced AI-powered tools and a creative canvas for fast iterations.
SamurAIGPT/Generative-Media-Skills
Collection of agent skills and workflow recipes giving Claude Code, Cursor, Gemini CLI and OpenCode access to image, video and audio generation models through muapi-cli and an MCP server.
QIN2DIM/hcaptcha-challenger
hCaptcha Challenger is an open-source project that uses multimodal large language models and advanced machine learning techniques to automate solving hCaptcha challenges without relying on third-party services or scripts.
roboflow/inference
Roboflow Inference is a platform that turns any computer or edge device into a command center for deploying and managing computer vision models and workflows, enabling advanced AI-powered visual applications.
mbzuai-oryx/groundingLMM
GLaMM is a groundbreaking multimodal AI model that generates natural language responses integrated with object segmentation masks, enabling advanced visual grounding and conversational tasks.