June 2026 Summaries

Comfy MCP has been introduced in public beta, offering a seamless integration between the Comfy Ecosystem and various agents like Claude, Codex, Hermes, and Cursor to enhance creative workflows across image, video, 3D, and audio models. This platform allows users to automate and update best-practice workflows, enabling tasks such as product placement, character design, and script-to-shot ideation at scale, all through natural language commands without the need for technical expertise in node graphs or GPUs. Comfy MCP is designed for reproducibility and teamwork, allowing agents to build, edit, and execute workflows while sharing them with ease, making it suitable for long-term projects. Users are encouraged to provide feedback during its beta phase, and additional resources, including a CLI and a community-driven Comfy Skill repository, are available to expand the user experience.

Seedance 2.0 Mini and 4K is now integrated into ComfyUI, providing users with a more cost-effective option for video creation through ByteDance's Seedance 2.0 video nodes, including Text to Video, First-Last-Frame to Video, and Reference-to-Video functionalities. Seedance 2.0 Mini allows for the generation of structured multi-shot sequences from text prompts at resolutions of 480p and 720p, maintaining character consistency and visual style for storytelling and branded content, though it lacks the 4K capability of the full Seedance 2.0. Meanwhile, the full Seedance 2.0 offers native 4K output with 10-bit color, ensuring fine detail and smooth transitions, ideal for high-resolution projects that require detailed and gradation-rich visuals. Users can start using Seedance 2.0 Mini by updating ComfyUI, selecting a Seedance 2.0 Mini template, and setting the desired resolution to begin generating videos.

HappyHorse 1.1, now integrated into ComfyUI as a Partner Node, offers advanced audio-native video generation capabilities suitable for real-world production scenarios such as episodic series, commercials, and game cutscenes. This version emphasizes synchronized audio, producing dialogue, sound effects, and background music in a single render pass, while addressing five critical production features: dynamic motion, consistent character rendering, reliable prompt adherence, stable text rendering, and authentic cinematic framing. The model supports three distinct nodes—Text-to-Video (T2V), Image-to-Video (I2V), and Reference-to-Video (R2V)—each tailored for specific tasks like building scenes from scratch, animating static frames, and orchestrating multi-character stages. Enhancements include smoother motion, improved multi-image reference-to-video capabilities, flexible character-scene combinations, upgraded instruction following, and refined audio synchronization, with outputs available in various aspect ratios and resolutions, ensuring seamless integration into creative workflows.

Krea 2 Open-Source Models, available in ComfyUI, feature two checkpoints—Krea 2 RAW and Krea 2 Turbo—offering diverse aesthetic outputs and allowing local training and generation on personal hardware. Krea 2 RAW serves as a base model for fine-tuning and research, producing diverse outputs without post-training, while Krea 2 Turbo, an 8-step distilled checkpoint, is optimized for fast, high-quality text-to-image generation with consistent outputs. The models are designed to work synergistically, with LoRA training on RAW enhancing performance when applied to Turbo. Krea 2 OSS supports a broad stylistic range, is trained on varied natural-language prompts, and employs a 12B dense DiT architecture with multi-layer feature aggregation. Users can leverage these models by updating ComfyUI to version 0.26.0 and following specific setup instructions to enhance their workflows.

Xindi Zhang's documentary animation "The Song of Drifters," which explores themes of belonging and memory through a personal narrative inspired by an ancient Chinese poem, has garnered critical acclaim, winning the 2025 Student Academy Award and being shortlisted for the Oscars. Zhang, a Chinese animation director and visual artist, utilized ComfyUI, a node-based AI tool, to enhance her creative process, integrating her own illustrations and live-action footage into the animation via advanced AI techniques like style transfer and custom LoRAs. Her innovative approach challenges stereotypes about AI-generated art, emphasizing the importance of artists actively participating in each stage of their creative workflow rather than relying on automated processes. This approach not only expanded her artistic expression but also led to career opportunities, including a position as an AI Creative at Amazon AI Studio. Zhang’s experience underscores the growing demand for professionals who can merge traditional artistry with AI tools, and she advises aspiring artists to prioritize storytelling and purposeful use of AI to maintain authenticity in their work.

Comfy Go is a native mobile app developed using the Comfy Cloud API, showcasing the potential of rapid app development with minimal resources. This app provides four generation flows—text-to-image, image-to-image, text-to-video, and image-to-video—allowing users to generate content directly on their phones without needing a laptop or browser. Built by a single developer in just one week, the app emphasizes accessibility both in terms of user experience and development, as it leverages a streamlined public API consisting of only three essential calls: submit a workflow, stream events, and receive outputs. The app's creation was facilitated by the ComfySwiftSDK, a lightweight Swift layer that simplifies interaction with the API, making it feasible for developers to create robust applications quickly. This project is part of the Comfy Vibes initiative, which encourages developers to explore innovative uses of the Comfy API, highlighting its simplicity and effectiveness in building functional, native applications.

At Comfy, a system was developed to enhance code review processes by leveraging AI models from four different labs, each offering unique perspectives and reducing blind spots that a single model might miss. This system fans out a pull request (PR) diff to four AI models from OpenAI, Anthropic, Google, and Moonshot, conducting two review passes per model, and consolidates the results through a single judge model. This approach aims to catch nuanced bugs such as concurrency issues and API contract drifts, which might be overlooked by human reviewers due to fatigue or by models sharing the same training priors. Operating as a $200/month GitHub Action, this system runs in continuous integration (CI) and is designed to avoid being manipulated by malicious PRs. It has successfully identified significant bugs in about 110 PRs so far, and the architecture is open-sourced to encourage further development and feedback from the engineering community.

Ideogram 4.0 is a newly released 9.3 billion parameter open-weights text-to-image model that is natively supported in ComfyUI from day one, providing tools for generating images from structured JSON prompts. This model allows users to exercise precise control over aspects of image creation such as color palettes, bounding-box layouts, and typed text elements, which significantly enhances the specificity and groundedness of the results. Unlike using flat text prompts, the JSON format enables detailed customization, with Ideogram 4.0 being particularly adept at handling scenes with exhaustively described elements. The model includes an inherent safety filter that cannot be modified by users, ensuring certain content restrictions are maintained during image generation. Users are encouraged to update to the latest version of ComfyUI, download the necessary workflow, and explore the model's capabilities using either natural language or JSON prompts, with comprehensive resources available for setup and experimentation.

TripoSplat, an open-source model from Tripo, is now natively supported in ComfyUI, allowing users to transform a single image into a 3D Gaussian asset suitable for modern 3D pipelines. This model excels in rendering stylized subjects like characters and props by utilizing adaptive density control to apply detail selectively, optimizing file size and rendering cost. TripoSplat provides flexibility in detail levels for different applications, such as using fewer Gaussians for background elements or more for prominent assets, and fits well into workflows involving 3D previews, AR/VR, and interactive content creation. Users can access the model's weights and inference code for local customization and community workflows, emphasizing its open-source nature under the MIT license.

In May, ComfyUI integrated 11 new models across various domains, including video, image, 3D, audio, and multimodal, enhancing its capabilities significantly. Notable integrations include Krea 2 for image and style transfer, VOID by Netflix for video object removal, and Tripo 3.1 for 3D generation, offering comprehensive text-to-model and image-to-model functionalities. Luma UNI-1 introduces advanced image editing with a decoder-only autoregressive transformer, while Claude from Anthropic brings multimodal understanding to workflows. OpenRouter provides access to over 20 language models, and Google DeepMind's Gemma 4 offers a versatile multimodal model. Additional integrations such as HidDream-O1-Image for reasoning-guided image generation, Stable Audio 3 for audio and sound effects, BiRefNet for high-resolution background removal, and MoGe by Microsoft for 3D geometry and depth enhance the platform's diverse toolset. ComfyHub's growing library of over 500 workflows indicates its expanding community and resource base.