May 2026 Summaries

Krea 2 Image is a foundational image model introduced as a Partner Node within ComfyUI, enabling users to have precise control over the creative direction, style, and mood of image generation. It includes two model variants, Krea 2 Large and Krea 2 Medium, each offering distinct characteristics suited for different artistic needs such as photorealism and expressive styles. The model allows users to manipulate various parameters, including creativity levels and moodboard conditioning, to produce diverse and nuanced outputs. A standout feature is its ability to perform style transfer by incorporating reference images, allowing for seamless application of existing styles to new creations. For those seeking to maintain visual consistency across projects, moodboards offer a refined method to guide the model's output. Users can access Krea 2 Image through the ComfyUI platform or via Comfy Cloud, making it a versatile tool for artists and studios aiming for specific aesthetic outcomes.

Stable Audio 3.0, Stability AI's latest suite of music models, is now integrated into ComfyUI, offering enhanced capabilities for artistic experimentation. These models are trained on fully licensed music data and support variable-length generation, enabling users to create both sound effects and extended musical tracks within existing workflows. The models come in small and medium sizes, with small models capable of running on CPUs for quick sound effects and short music loops up to two minutes, while medium models, requiring a GPU, can generate longer and more structurally complex tracks up to six minutes. Stable Audio 3.0 is licensed for commercial use and is designed to be user-friendly, supporting a range of audio applications from music production to sound design for multimedia projects.

Three new open-source models have been integrated into ComfyUI, enhancing its capabilities in text, image, and video modification. Netflix's VOID is a video inpainting model designed to remove objects and their physical interactions, such as shadows and reflections, from videos through a unique quadmask approach that considers cause and effect. BiRefNet, originating from CAAI AIR 2024, excels in high-resolution background and object segmentation, producing detailed masks for complex images and supporting tasks like salient and camouflaged object detection. Google's Gemma 4 is a multimodal reasoning model capable of processing various inputs including text, image, audio, and video, with a configurable thinking mode for step-by-step reasoning. Each model is available for download through ComfyUI, with workflows provided for easy implementation, offering users the ability to experiment and create with these advanced tools.

Tripo 3.1 is now supported in ComfyUI via Partner Nodes, enabling the generation of high-detail 3D assets with improved geometry density and surface quality that are suitable for close-up production use. This tool is designed to assist creators in transitioning quickly from concept to usable models, making it ideal for applications such as game production, marketing visuals, product rendering, and 3D printing. Key features of Tripo 3.1 include high-density geometry for sharper edges and cleaner silhouettes, PBR-ready material behavior for reliable lighting response, and hero asset quality that remains stable in close-up shots. It also allows for cross-scenario reuse, making models versatile across different workflows. Typical use cases for Tripo 3.1 include game hero assets, marketing visuals, product visualization, and 3D printing preparation, where geometric fidelity and material quality are prioritized.

Luma's Uni-1 is an advanced autoregressive image model now accessible via ComfyUI Partner Nodes, distinguishing itself from traditional diffusion models by using a decoder-only autoregressive transformer that treats text and images as a unified interleaved sequence. This innovative approach allows the model to reason through prompts before generating images, effectively decomposing instructions, resolving constraints, and planning composition akin to a frontier large language model, resulting in superior performance on visual reasoning benchmarks, particularly with infographics and text. Uni-1 offers features such as photorealism with material accuracy, readable text rendering, reference-guided generation, and identity preservation, as well as multi-turn refinement and multi-panel output with temporal consistency. The model supports nine aspect ratios, ranging from ultra-wide panoramic banners to ultra-tall portrait formats, enhancing its flexibility for various creative applications. Users can access Uni-1 by updating ComfyUI or using Comfy Cloud, allowing them to experiment with this cutting-edge technology by simply dropping in prompts and connecting outputs.

In April, ComfyUI experienced a dynamic month marked by the launch of numerous new models, features, and enhancements to its Comfy Hub. The company introduced a range of innovative models such as Seedance 2.0 for video creation, Happy Horse for multi-talented video tasks, and Ace Step 1.5 XL for commercial-grade music generation. Notably, the platform now supports frame interpolation through open-source models RIFE & FILM, and image and video segmentation via Meta's SAM 3 models. The introduction of Parallel Job Execution via API significantly enhances workflow efficiency, allowing users to run multiple workflows simultaneously. Furthermore, the Comfy Hub has expanded, hosting nearly 500 workflows, underscoring its growing utility for users seeking ready-made solutions.