March 2026 Summaries

Topaz Starlight Precise 2.5, now available as a Partner Node in ComfyUI, offers enhanced video upscaling capabilities with sharper results and fewer artifacts compared to its predecessor, SLP-2, while maintaining the same cost. This update significantly improves fine detail preservation and reduces hallucination effects, particularly noticeable in workflows upscaling from 720p to 4K, making the output appear more native rather than stretched. Users can easily upgrade by selecting Starlight Precise 2.5 from the model dropdown in their existing Topaz upscaling node without any changes to their workflow, ensuring immediate benefits in video quality without additional expense.

ComfyUI has introduced a new Dynamic VRAM memory optimization system to address the recent rise in hardware RAM prices and enhance the efficiency of running diffusion models on memory-constrained devices. Available for Nvidia hardware on Windows and Linux, this update reduces system RAM usage, eliminates out-of-memory errors, and speeds up loading times by fundamentally changing how model weights are handled. Dynamic VRAM employs a custom PyTorch VRAM allocator that dynamically offloads model weights only when necessary, allowing for higher VRAM utilization and preventing reliance on slower page files. This advancement simplifies memory management by utilizing a Virtual Base Address Register (VBAR) and fault() API for just-in-time allocation of tensors, optimizing memory usage and preventing crashes. The system allows for smooth operation even with models exceeding physical RAM capacity, maintaining high performance without manual VRAM management. Future developments aim to expand hardware support and further reduce RAM footprint while enhancing disk loading speeds.

Grok Imagine has introduced two new features for its video model, available with Partner Nodes, which enhance creative video production using AI. The Grok Reference-to-Video node can generate consistent videos from up to seven reference images, while the Grok Video Extend node allows users to extend a 15-second input video by an additional 10 seconds, ensuring coherent motion and audio. These features offer impressive speed, with the Grok Video Extend generating 10-second additions in about 60 seconds and Reference-to-Video producing 720p videos in roughly 100 seconds. Users can access these features by updating ComfyUI or visiting ComfyCloud, where they can find the nodes in the Node Library or Templates. These updates reflect a creative blend of traditional video production with generative tools and are part of the ongoing news and updates from the Comfy team.

Doodles, an entertainment brand known for its pastel artwork by Canadian illustrator Scott Martin, is launching Doodles AI, a generative platform powered by PRISM 1.0, which is a generative image model trained on Doodles' artwork. This platform allows users not only to generate art in the Doodles style but also to contribute to its evolution through reinforcement learning, effectively turning consumers into collaborators. Engineered by the AI studio SYSTMS, which uses open-source tools like ComfyUI, Doodles AI exemplifies the potential of open-source workflows in producing commercially viable, brand-specific products, distinguishing itself from generic platforms. The initiative marks a shift towards community-driven creativity, allowing for co-creation and fostering a new model of intellectual property sharing. With an ambitious roadmap including 2D and 3D output generation, video, real-time augmented reality, and gaming applications, Doodles AI aims to align with the rapid growth of the AI industry and integrate emerging technologies into its "coded DNA." This approach not only challenges traditional notions of IP protection but also underscores the strategic value of openness in creative industries.

The "Complete Style Transfer Handbook" explores the advanced techniques and models available for AI-driven style transfer in image generation workflows, offering creators multiple reliable methods for applying specific visual styles to their content. It highlights the evolution from basic tools like IP-Adapter to sophisticated models, including custom LoRAs (Low-Rank Adaptations) that allow precise style application without retraining entire models. The handbook discusses the importance of maintaining a consistent visual identity and the limitations of text-based prompting, advocating for style transfer techniques that enable accurate and repeatable results. It differentiates between image-edit and image-gen models, emphasizing the distinct roles of each in style transfer processes. The guide provides insights into the creation and application of custom LoRAs, underscoring the significance of high-quality datasets and the nuances involved in model training. It also mentions ready-to-use solutions like Recraft, which excels in abstract and artistic style transfer, and the importance of choosing the right model based on the desired level of control and specificity.

Reve is now integrated into ComfyUI via Partner Nodes, providing users with a powerful tool for generating high-quality, aesthetically pleasing images at resolutions up to 4k. The Reve Create node excels in producing dynamic images with unique lighting, textures, and a fashion editorial style, while the Reve Remix and Edit nodes allow for rapid iteration and enhancement of image outputs. Reve's capabilities include generating complex lighting effects, capturing motion blur, and providing cinematic quality, all of which are demonstrated through detailed examples such as high-fashion portraits and high-speed car shots. The system's flexibility is highlighted by its ability to incorporate elements from reference images, allowing users to remix and edit images seamlessly while maintaining aesthetic consistency. Users can access these features through Comfy Cloud by updating ComfyUI to its latest version and exploring the Reve nodes available in the Node Library.

HY 3D's advanced processing features are now integrated into ComfyUI, providing production-ready 3D post-processing capabilities that streamline workflows for 3D creators by addressing the common challenge of transforming raw meshes into assets ready for texturing, animation, and production pipelines. The update introduces three key features: 3D Parts Decomposition, which allows for the division of models into editable components; UV Unwrapping, which automatically generates UV layouts for easier texturing; and Smart Topology, which converts dense geometry into cleaner, production-friendly meshes. These capabilities facilitate modular workflows, optimize assets for game engines and real-time rendering, and support a smoother transition from raw model generation to final production use.

ComfyUI has introduced a new suite of tools designed to transform any ComfyUI workflow into a shareable, no-install application, making it accessible to users without a node graph background. The key features include App Mode, which provides a streamlined interface for running workflows, and App Builder, which allows creators to configure which node inputs and outputs are visible to users. These tools enable the sharing of workflows via a single URL, making it easy for anyone to access and use the apps without installing additional software. Additionally, ComfyHub serves as a platform for sharing and discovering community-created apps and workflows, providing a space for creators to showcase their work and for users to find and utilize new tools. The availability of these features on both Comfy Cloud and Comfy Local ensures optimized deployment across various systems, reinforcing ComfyUI's commitment to user control and community collaboration.

ElevenLabs has integrated its advanced voice AI capabilities into ComfyUI through Partner Nodes, enabling users to create voice-enhanced workflows directly within the node graph interface without needing external tools. This integration supports various audio functions such as text-to-speech, speech-to-speech transformation, speech-to-text transcription, voice isolation, text-to-dialogue generation, and text-to-sound effects creation, making it ideal for applications like podcasting, voiceovers, dialogue isolation, and creative sound design. The addition of ElevenLabs to ComfyUI facilitates the creation of comprehensive multimodal content pipelines, allowing users to seamlessly integrate voice with image, video, and text generation processes, enhancing the versatility and efficiency of content creation.

Kling Motion Control 3.0 has been released via Partner Nodes in ComfyUI, enhancing the capabilities introduced in Kling 2.6 Motion Control with a new feature called Element Binding. This system ensures facial consistency, allowing characters to maintain stable and recognizable features across various angles, emotions, and occlusions in generated videos. The upgrade offers improved expression fidelity by allowing users to upload multiple face reference images, ensuring accurate representation of complex emotions. Additionally, the new version provides face restoration through partial occlusions, maintaining character identity despite obstructions like hats or hands. It also ensures consistent face quality during camera movements such as zooming, panning, or tracking. To start, users can update ComfyUI to the latest version and search for the "Kling Motion Control" node to utilize these enhanced features.

LTX-2.3, the latest iteration of Lightricks' open-source audio-video generation model, is now supported in ComfyUI, bringing significant quality improvements to the field of open-source media creation. This update enhances the model's capabilities in several areas, including sharper textures and cleaner edges due to a new latent space and updated VAE, improved quality for vertical portrait videos suited for social media, and better audio quality with reduced noise for dialogue and music. The image-to-video conversion is more seamless with consistent motion and fewer glitches, while the text encoder has been upgraded for smarter prompt understanding and more accurate text rendering in videos. Users can easily access these improvements by updating ComfyUI to the latest version and downloading the necessary models through the Comfy Cloud platform, allowing for enhanced creative workflows in audio and video generation.

Comfy Cloud, an extension of ComfyUI, has officially exited its beta phase, marking a significant milestone for its community of creators and developers who contributed to its development. This platform offers cloud-based access to custom nodes, which are central to many local workflows, enabling users to run their most-used pipelines without the need for installation. Comfy Cloud's pricing is tailored to charge only for active GPU usage, encouraging experimentation without incurring idle time costs. It provides instant access to a wide array of open-source and closed-source AI models, and supports custom LoRA uploads for personalized output. Equipped with NVIDIA Blackwell RTX 6000 Pro GPUs, it offers high-performance capabilities for demanding workflows. Additionally, Comfy Cloud has introduced a free tier to attract new users, with paid plans offering extended features. Looking ahead, the platform plans to support API deployment, parallel workflow execution, and team collaborations, while expanding its model library and custom node offerings. To celebrate its launch, Comfy Cloud is hosting the #comfyrewind challenge, inviting creators to produce retro-style reels for a chance to win cloud credits.

Comfy Cloud has introduced a Free Tier offering 400 free credits each month to users without a subscription, allowing them to explore a wide range of image, video, audio, and 3D models without the need for payment or a credit card. Users can sign up using Google, and their accounts are automatically upgraded to this tier, enabling access to top image models such as Nano Banana Pro and Seedream 5.0, as well as video generation tools like Seedance and Kling. The platform supports over 900 models and offers 350 built-in templates, with workflows running on RTX Pro 6000 GPUs. While credits do not roll over and refresh monthly, users can still access their generated outputs and inputs, and they can import and run custom workflows similar to those available in paid tiers. A subscription option is available for those who want to continue using the platform beyond the free credits, with plans starting at $20 per month.