ComfyUI News & Open-Source AI Releases

JD Open Source releases JoyAI-Video-Edit, a 16B autoregressive diffusion model that edits live video streams in real time at 30 FPS 720p.

Apache-2.0

A classic OpenPose ControlNet LoRA for Krea 2 Turbo: feed a DWPose skeleton map and the model follows the pose while the prompt defines appearance.

Apache-2.0

Download taeh3.safetensors (MiniMax H3 TAE) and preview videos fast in ComfyUI. A tiny 2D autoencoder used via ModelPreviewOverride in ComfyUI-KJNodes.

Install xmarre's ComfyUI-Spectrum-MiniMax-H3 node: spectral feature forecasting skips transformer evaluations for ~30% faster MiniMax H3 generation in ComfyUI.

Download MiniMax H3 community quants for ComfyUI: GGUF Q2-Q5 (MiniMax-H3-FL2VA-Q3_K_M/Q4_K_M, text encoder GGUFs), pruned INT4/INT8 and NVFP4 files with bundled workflows.

minimax-h3-community-license-agreement

Train MiniMax H3 LoRAs with Ostris AI Toolkit: T2V and I2V support, NVFP4 quantization and VAE gradient checkpointing for consumer GPUs.

Open-Weights

Download MiniMax H3 open weights for ComfyUI: FL2VA and Ref2VA checkpoints, bf16, INT8, pruned and NVFP4 files, plus six official workflow templates (T2V, I2V, R2V).

Open-Weights

SenseTime's SenseNova-U1.5-8B-MoT-Preview brings 4K text-to-image and stronger image editing in one unified model, with ComfyUI support via the SenseNova_U1 custom node pack.

Apache-2.0

Image-editing LoRAs that convert a character image into clean TripleView and QuadView reference sheets, with bundled ComfyUI workflows for FLUX.2 Klein 9B and Krea 2.

Civitai-Model-License

Lodestones releases Kroma v0.1, a rank-256 ComfyUI-ready LoRA for Krea 2 with full fine-tuned norm and modulation deltas. Works on Krea 2 base and Turbo.

MIT

MiniMax launches H3, an omni-modal video model generating 2K 15-second clips with native stereo audio, multimodal context understanding, and instruction-based editing.

Open-Weights

A 1.4B, 6-layer distillation of SeedVR2-7B brings one-step 4x/8x upscaling to low-VRAM machines, with a drop-in ComfyUI custom node and ready-made workflow.

Apache-2.0

PrunaVAED is a drop-in replacement VAE decoder for LTX-2.3, achieving 1.7x faster decoding and ~50% lower peak VRAM with near-original visual quality.

ltx-2

Eyeline-Labs ID-V2V, a SIGGRAPH Asia 2026 identity-preserving video-to-video model, now has ComfyUI support through Kijai int8 convrot weights and a core PR.

NVIDIA releases Qwen-Image-Flash, a four-step DMD2-distilled Qwen-Image that keeps the base architecture for fast text-to-image generation.

nvidia-open-model-license

Black Forest Labs unveils FLUX 3, a unified multimodal foundation model generating 20-second video with native audio, images, and action prediction for robotics.

API-Only

Microsoft Asia releases Mage-Flow -- compact 4B model for text-to-image and editing, native resolution up to 2048, RL/4-step Turbo variants, MIT license.

MIT

NVIDIA releases Cosmos3-Edge, a 4B Mixture-of-Transformers world model generating video, images and audio from text, images and action trajectories, with ComfyUI community nodes.

OpenMDW-1.1

Damkohler released JLC Flux2 ControlNet v1.0.0, a custom node bringing FLUX.2-dev Fun ControlNet Union into ComfyUI with multi-branch control, reference conditioning, and in/out-paint.

Oasis Suite v1.5 expands Image Oasis into a three-node pack, adding Video Oasis Viewer for in-node preview and LTX2.3 Oasis for all-in-one LTX 2.3 video generation.