June 2025 Summaries

OmniGen2 is now natively supported in ComfyUI, offering a robust multimodal model developed by VectorSpaceLab that integrates text-to-image generation, image editing, and multi-image composition. With its 7 billion parameters, OmniGen2 allows users to create high-quality images from text prompts, make precise edits using natural language commands, seamlessly combine elements from multiple images, and generate clear text content within images, all powered by the Qwen-VL-2.5 for superior visual understanding. To get started, users can update ComfyUI or its desktop version and follow the documentation to download models and workflows, enabling creative processes such as transforming a character into crystal material or altering the time of day in an image while maintaining composition.

FLUX.1 Kontext [dev] is an open-source diffusion transformer model developed by Black Forest Labs, which has been integrated into ComfyUI and is a 12 billion parameter sibling of the FLUX.1 Kontext model. It retains the core features of the FLUX.1 Kontext suite, such as character consistency, local editing, and style reference, allowing users to maintain unique elements across images, make precise modifications, and generate novel scenes with preserved styles guided by text prompts. Although the previous API version offered superior fidelity and speed, FLUX.1 Kontext [dev] operates entirely on local machines, providing flexibility for developers and researchers to experiment. Users can access it by updating ComfyUI and exploring various templates to create outputs like transforming images into 1960s pop art or altering scenes to display different perspectives or lighting conditions, with detailed documentation available for further guidance.

ComfyUI has announced native support for NVIDIA's Cosmos-Predict2 model suite, which is designed for generating physics-aware images, videos, and world states to aid in Physical AI development. The Cosmos-Predict2 models, which can be used commercially under the NVIDIA Open Model License, are known for their high performance and are available at Comfy-Org/Cosmos_Predict2_repackaged. To get started, users should update ComfyUI to the latest version, access the Cosmos templates through the Workflow → Template option, and download the necessary models. The suite includes workflows such as Cosmos-Predict2 Text-to-Image and Video2World (Image-to-Video), and users are encouraged to refer to the documentation for more details and to explore the example outputs.

Following the announcement of a new subgraph feature, the development team has refined the input and output connection system, making it available for early beta testing. Key updates include improved behavior when collapsing subgraphs, enhanced handling of links within subgraphs, and overall stability improvements. Developers are encouraged to test these changes by updating to the master branch of ComfyUI and launching with a test frontend, although these updates are not yet included in stable releases. Feedback from early testers is highly valued, with suggestions and findings welcome on Discord and GitHub to help shape the feature before its stable release.

ComfyUI is undergoing significant improvements with its Nodes v3 initiative, aimed at enhancing the custom node experience by addressing challenges such as stability, dependency conflicts, and dynamic input/output support. Custom nodes, which are central to ComfyUI, previously faced issues due to a lack of separation between internal and public interfaces, causing potential breakage with updates. Nodes v3 proposes creating a public API to ensure backward compatibility and improve stability, while the introduction of isolated Python processes for custom node packs is being explored to resolve dependency issues. Dynamic inputs/outputs and custom widgets are set to receive first-class support through a more explicit, object-oriented schema, making it easier for developers to manage node functionality without memorizing numerous class methods and property names. The initiative also seeks to streamline model management and enable future capabilities like distributed and parallel execution. Communication with custom node developers is facilitated via a private Discord channel, and the project is open to feedback as development progresses. The new API will support backward compatibility and facilitate easier maintenance of custom nodes, incorporating versioned API objects and async calls to support parallel and out-of-process execution.

The upcoming release of subgraphs in ComfyUI is set to transform workflow management by allowing users to consolidate complex processes into custom super-nodes, maintaining the same editing tools as regular workflows without a new learning curve. These subgraphs replace the previously attempted group nodes, providing a more efficient and practical solution for managing intricate workflows. Users can collapse nodes, groups, and reroutes into a single subgraph node, significantly reducing visual complexity and improving readability and manageability. Subgraphs offer ease of editing, where changes are automatically updated across all instances in a workflow, and cloning options allow unique modifications without affecting the original. Additional features include widgets for exposing controls at the subgraph level and the ability to nest subgraphs, enabling the creation of utility libraries and complex abstractions. The public beta will soon be available, and users are encouraged to update their ComfyUI for early access and engage with the community on Discord.

ComfyUI has introduced a built-in help menu to assist users in understanding and effectively using nodes within its interface, making workflow creation more accessible, especially for beginners. These help pages are hand-written, available in multiple languages, and include descriptions of nodes, their inputs, and outputs, even if custom documentation is not provided. The documentation for core nodes is open source, inviting community contributions and improvements, and extension authors can add their own documentation by following a developer guide. This feature aims to reduce the time spent on trial and error by offering easy access to node information, and it is currently available in the latest nightly version of ComfyUI's frontend, with feedback welcomed through GitHub and other official channels.