ComfyUI — an open-source, node-based graphical interface for working and constructing generative AI workflows for content material creation — revealed main updates this previous month, together with as much as 40% efficiency enhancements for NVIDIA RTX GPUs, and help for brand spanking new AI fashions together with Wan 2.2, Qwen-Picture, FLUX.1 Krea [dev] and Hunyuan3D 2.1.
NVIDIA additionally launched NVIDIA TensorRT-optimized variations of in style diffusion fashions like Secure Diffusion 3.5 and FLUX.1 Kontext as NVIDIA NIM microservices, permitting customers to run these fashions in ComfyUI as much as 3x sooner and with 50% much less VRAM.
Plus, an replace to NVIDIA RTX Remix — a platform that lets modders remaster basic video games — launched immediately, including a complicated path-traced particle system that delivers beautiful visuals to breathe new life into basic titles.
ComfyUI v3.57 Will get a Efficiency Enhance With RTX
NVIDIA has collaborated with ComfyUI to spice up AI mannequin efficiency by as much as 40%. To place this in perspective, GPU technology upgrades usually solely ship a 20-30% efficiency increase.
Builders excited by optimizing efficiency and effectivity of diffusion fashions on their apps can learn extra on how NVIDIA helps speed up these workloads within the developer discussion board.
State-of-the-Artwork AI Fashions, Accelerated by RTX
Unbelievable fashions for AI content material creation have been launched within the final weeks, all of which are actually out there in ComfyUI.
Wan 2.2 is a brand new video mannequin that gives unimaginable high quality and management for video technology on PCs. It’s the newest mannequin from Wan AI — a inventive AI platform providing a powerful lineup of AI fashions, together with Textual content to Picture, Textual content to Video, Picture to Video and Speech to Video. GeForce RTX and NVIDIA RTX PRO GPUs are the one GPUs able to working Wan 2.2 14B fashions in ComfyUI with out main delays in output. Try the instance beneath created with the one immediate: “A robotic is cracking an egg, however by chance hits it exterior the bowl.”
Qwen-Picture is a brand new picture technology basis mannequin from Alibaba that achieves important advances in advanced textual content rendering and exact picture modifying. It excels at rendering advanced textual content, dealing with intricate modifying and sustaining each semantic and visible accuracy in generated photos. The mannequin runs 7x sooner on a GeForce RTX 5090 vs. Apple M3 Extremely.
Black Forest Labs’ new FLUX.1 Krea [dev] AI mannequin is the open-weight model of Krea 1, providing sturdy efficiency and educated to generate extra real looking, various photos that don’t comprise oversaturated textures. Black Forest Labs calls the mannequin “opinionated,” because it affords all kinds of various, visually fascinating photos. This mannequin runs 8x sooner on a GeForce RTX 5090 vs. Apple M3 Extremely.
Hunyuan3D 2.1 is a totally open-source, production-ready 3D generative system that transforms enter photos or textual content into high-fidelity 3D property enriched with bodily primarily based rendering supplies. Core elements embody a 3.3-billion-parameter mannequin for form technology and a 2-billion-parameter mannequin for texture evaluation to shortly generate extra real looking supplies. All of it runs sooner on Blackwell RTX GPUs.
Get Began With Superior Visible Generative Methods
Visible generative AI is a strong device, however getting began could be tough even for technical specialists and studying to make use of extra superior methods usually takes months.
ComfyUI makes it simple to get began with superior workflows by offering templates or preset nodes that obtain a selected job, like preserving a personality fixed all through totally different generations, adjusting the sunshine of a picture or loading a fine-tuning. This permits even non-technical artists to simply use superior AI workflows.
These are 10 key methods to get began with generative AI:
- Information video technology by defining the beginning body and finish body: Add a begin and finish body, in addition to how a video clip ought to start and finish. Wan 2.2 can then generate a easy, animated transition, filling the in-between frames to create a coherent animation. It’s best for making animations, scene shifts or defining poses.
- Edit photos with pure language: Use FLUX.1 [dev] KONTEXT to edit particular textual content sections of a picture.
- Upscale photos or movies: Take picture or video at a decrease decision and enhance its decision and element high quality by including real looking, high-frequency particulars.
- Management area composition: Assume extra granular management over picture technology by controlling the association and format of visible parts inside particular areas of a picture.
- Restyle photos: Use FLUX Redux to create totally different variations of photos whereas preserving core visible parts and particulars.
- Faucet image-to-multiview-to-3D fashions: Use a number of photos of an object captured from totally different angles to create a high-fidelity, textured 3D mannequin.
- Rework sound to video: Create video clips or animations immediately from audio inputs akin to speech, music or environmental sounds.
- Management video trajectory: Mechanically information the movement of objects, digicam or scenes inside movies.
- Edit photos with inpainting: Fill in or alter lacking or undesirable elements of a digital picture in a visually seamless, contextually constant means with surrounding areas.
- Broaden the canvas with outpainting: Generate new picture content material to increase the boundaries of an present picture or video footage, add element to cropped sections or full a scene.
Comply with ComfyUI on X for updates to inventive templates and workflows.
Broaden Your ComfyUI Zone
ComfyUI plug-ins allow customers so as to add generative AI workflows into their present purposes. The ComfyUI neighborhood has began constructing plug-ins for among the high in style inventive purposes.
The Adobe Photoshop plug-in enhances Photoshop’s native Firefly fashions by permitting customers to run their very own flows and choose specialised fashions for particular duties. Native inference additionally allows limitless generative fill with low latency.
The Blender plug-in — featured within the NVIDIA AI Blueprint for 3D-guided generative AI — permits customers to attach 2D and 3D workflows. Artists can use 3D scenes to manage picture generations or create textures in ComfyUI and apply them to separate 3D property.
The Foundry Nuke plug-in — just like the Blender — allows a connection between 2D and 3D workflows so customers don’t should alt-tab and swap between purposes.
The Unreal Engine plug-in allows ComfyUI nodes immediately within the Unreal Engine consumer interface to shortly create and refine textures for scenes utilizing generative diffusion fashions. See the instance beneath.
Run Hyper-Optimized Fashions for NVIDIA RTX GPUs in ComfyUI
One of the best ways to make use of NVIDIA RTX GPUs is with the TensorRT library — a high-performance deep studying inference engine designed to squeeze most pace out of the Tensor Cores in NVIDIA RTX GPUs.
NVIDIA has collaborated with the highest AI labs to combine TensorRT of their fashions, akin to Black Forest Labs’ fashions and Stability AI’s fashions. These fashions are additionally out there quantized — a compressed model of the community that makes use of 50-70% much less VRAM and affords as much as 2x sooner inference whereas sustaining comparable high quality.
TensorRT-optimized fashions could be run immediately in ComfyUI by means of the TensorRT node, which at present helps SDXL, SD3 and SD3.5, in addition to FLUX.1-dev and FLUX.1-schnell fashions. The node converts the AI mannequin right into a TensorRT-optimized mannequin, after which generates a TensorRT-optimized engine for the consumer’s GPU — a map of methods to execute that mannequin with optimum effectivity for his or her specific {hardware} — offering important speedups.
Quantizing fashions, nevertheless, takes a bit extra work. For customers excited by working quantized and TensorRT-optimized fashions, NVIDIA affords preconfigured information in a easy container referred to as a NIM microservice. Customers can use the NIM node in ComfyUI to load these containers and use quantized variations of fashions like FLUX.1-dev, FLUX.1-schnell, FLUX.1 Kontext, SD3.5 Massive and Microsoft TRELLIS.
Remix Replace Provides a Path-Traced Particle System
A brand new RTX Remix replace launched immediately by means of the NVIDIA app, including a complicated particle system that permits modders to boost conventional hearth and smoke results, in addition to extra fantastical results, like these within the online game Portal.
With RTX Remix, legacy particles from basic video games could possibly be interpreted as path-traced, enabling them to solid real looking mild to boost the looks of many scenes. However in the end, these particles have been nonetheless over 20 years outdated, missing element, aptitude and fluid animations.
RTX Remix’s new particles have bodily correct properties and may work together with a recreation’s lighting and different results. This allows particles to collide, precisely transfer in response to wind and different forces, be mirrored on surfaces, solid shadows and assign their very own shadows.
For an entire breakdown of the brand new particle system, learn the GeForce article.
Every week, the RTX AI Storage weblog sequence options community-driven AI improvements and content material for these trying to be taught extra about NVIDIA NIM microservices and AI Blueprints, in addition to constructing AI brokers, inventive workflows, productiveness apps and extra on AI PCs and workstations.
Plug in to NVIDIA AI PC on Fb, Instagram, TikTok and X — and keep knowledgeable by subscribing to the RTX AI PC publication. Be part of NVIDIA’s Discord server to attach with neighborhood builders and AI fans for discussions on what’s attainable with RTX AI.
Comply with NVIDIA Workstation on LinkedIn and X.
See discover relating to software program product data.