ComfyUI v0.28.0: Native Video Super-Resolution, 3D Support, and PixelDiT Architecture — A Complete Breakdown

As the most popular node-based workflow tool in the open-source AI image and video generation space, ComfyUI has just released version 0.28.0. This update delivers significant improvements across three dimensions: model support, node functionality, and partner integrations. The addition of native video super-resolution, the new PixelDiT architecture, and 3D data export capabilities further solidifies ComfyUI's position as an all-in-one generative AI creative platform.

New Open-Source Model Support

The most noteworthy highlight of this update is the native integration of several cutting-edge open-source models — users can call upon these capabilities directly within their workflows without relying on third-party plugins.

SeedVR2: Native Image and Video Super-Resolution

The addition of SeedVR2 gives ComfyUI native image and video super-resolution (upscaling) capabilities. Super-resolution technology has undergone a dramatic evolution from traditional interpolation algorithms to deep learning. Early models like ESRGAN (Enhanced Super-Resolution Generative Adversarial Network) used adversarial training to achieve breakthroughs in perceptual quality, but their frame-by-frame processing nature imposed inherent limitations in video scenarios. SeedVR2 represents the diffusion model approach to super-resolution, reconstructing high-frequency details through iterative denoising while establishing inter-frame temporal dependencies — a technical direction aligned with research like Google's Imagen Video and Stability AI's VideoLDM.

Compared to traditional ESRGAN-style super-resolution models, SeedVR2 is built on a diffusion architecture and performs significantly better on complex textures, facial details, and inter-frame consistency in video. Temporal consistency has always been the core pain point of video super-resolution — frame-by-frame independent upscaling often causes flickering. SeedVR2's native video support is expected to address this issue at an architectural level, making it practically meaningful for video creators.

PixelDiT PID 1.5: A New Diffusion Architecture

PID 1.5 introduces support for the PixelDiT (Pixel Diffusion Transformer) architecture. DiT (Diffusion Transformer) was proposed by Meta AI Research in 2022, with the core idea of replacing the traditional U-Net backbone in diffusion models with a Transformer. This architectural shift delivers better scalability — model performance improves more smoothly as parameter count increases — which is the foundation that enabled OpenAI's Sora and Stability AI's SD3 to achieve ultra-large-scale training, and underpins most mainstream generative models today.

What makes PixelDiT unique is that it performs diffusion directly in pixel space, rather than the currently dominant latent space approach. The latent space approach compresses images into a low-dimensional representation via a VAE before diffusion, which is more computationally efficient; the pixel space approach skips the encode-decode step, theoretically preserving more complete spatial information. This gives the community an alternative to latent space diffusion, enriching ComfyUI's model ecosystem.

int4 Quantization: Lowering the Hardware Barrier

This update also adds int4 quantization support for convrot, along with dedicated int8/int4 optimizations for NVIDIA Turing architecture GPUs. Model quantization is the technique of compressing neural network weights from high-precision floating-point numbers (typically FP16 or BF16) to low-bit integer representations: int8 quantization compresses each parameter from 16 bits to 8 bits, theoretically reducing VRAM usage by about 50%; int4 further reduces VRAM usage to roughly 25% of the original. The Turing architecture (corresponding to NVIDIA RTX 20 series) was among the first consumer GPUs to introduce Tensor Core hardware-level int8 acceleration. This targeted optimization means this now-aging GPU generation can achieve significant inference speedups — consistent with the philosophy of communities like llama.cpp and GPTQ, bringing frontier AI capabilities to a broader hardware base through software-level innovation. For users with limited VRAM, the hardware barrier to high-quality generation is substantially lowered.

Comprehensive Node Upgrades

ComfyUI's core competitive advantage lies in its modular node system. This update adds numerous practical nodes, particularly in the 3D and text processing domains.

3D Asset Export: Native Support for Gaussian Splatting and Point Clouds

The new Save 3D (Advanced), Save Splat, and Save Point Cloud nodes mark ComfyUI's official integration of 3D content generation into native workflows. Save Splat corresponds to the widely popular 3D Gaussian Splatting (3DGS) technique — introduced by a team from Heidelberg University in 2023, it represents scene geometry using explicit 3D Gaussian ellipsoids, rendering orders of magnitude faster than the previously dominant NeRF (Neural Radiance Fields), enabling real-time preview on consumer hardware. Its core data format, the .splat file, stores each Gaussian kernel's position, rotation, scale, color, and opacity attributes, and can interface directly with downstream 3D editing tools like SuperSplat and Luma AI. Save Point Cloud handles point cloud data export.

The addition of these nodes enables a complete 2D-to-3D asset generation pipeline to be closed-loop within ComfyUI, offering direct value to creators in game development, film production, and digital twin applications.

Enhanced Text Processing Nodes

The SaveText node now supports saving text output in .txt, .md, or .json formats; the Text Overlay node can overlay text on images and videos; and the previously deprecated built-in Text node has been re-enabled.

Though these basic features may seem simple, they can significantly improve automation in batch generation scenarios — automatically logging generation parameters, adding watermarks or subtitles, and reducing post-processing workload.

Partner Node Integrations

Beyond in-house features, ComfyUI continues to expand its connections to commercial AI services, adding several notable partner nodes in this release.

sync.so sync-3: Lip Sync and Talking Images

The sync.so sync-3 model brings native support for lip-sync and "talking image" generation. The core task of lip-sync technology is to establish a mapping between audio features and facial mouth shapes. Early approaches relied on rule-based phoneme-to-viseme mapping; in the deep learning era, models like Wav2Lip achieved more natural results through end-to-end training. sync-3 represents the current commercial frontier in this space, using a diffusion model to generate the mouth region while maintaining overall facial consistency and improving both lip detail and synchronization accuracy. Integrated into the node workflow, users can seamlessly embed lip sync into more complex generation pipelines — a significant step toward productizing high-growth applications like digital humans, virtual streamers, and video dubbing.

Seedream Thinking Control and Gemini Image General Availability

The Seedream node has added an option to disable "Thinking" (chain-of-thought reasoning), giving users the flexibility to trade off speed versus quality — disabling the reasoning process yields faster generation. Chain-of-Thought reasoning improves adherence to complex instructions by having the model perform explicit intermediate steps before outputting the final result, but it comes with additional latency overhead. This toggle gives users finer-grained control over efficiency. Additionally, the Gemini Image preview model has been upgraded to its general availability release, with officially recognized improvements in stability and usability.

Summary

ComfyUI v0.28.0 is not a single-point breakthrough — it's a broad, systematic upgrade. From native video super-resolution (SeedVR2), a new diffusion architecture (PixelDiT), and low-bit quantization optimizations, to 3D asset export, lip sync, and commercial model integrations, the trajectory is clear: ComfyUI is evolving from an "image generation tool" into a "unified workbench for multimodal generative AI."

For everyday users, quantization lowers the hardware barrier. For professional creators, 3D and video capabilities expand creative boundaries. For developers, the continuously open node ecosystem means more possibilities. Users with relevant needs are encouraged to update soon and pay attention to how quantized models perform on their specific GPU.

Key Takeaways