ComfyUI News & Open-Source AI Releases

A rank-64 edit LoRA gives MOSS-Image instruction-based editing with one to four references, and ComfyUI core merged native edit-LoRA support the same day.

Taobao Content AI's PTD per-hop LoRAs take MiniMax H3 Ref2VA down to 8 or 4 steps, with a RefAsymOPD sibling and validated ComfyUI nodes.

Tencent Hunyuan released Hy-Nomas, an 8-step interactive world model that generates 720p video under text and camera control, with a ComfyUI port already out.

WarmBloodAban's MiniMax H3 Singularity V2 INT8 brings wider action coverage, tighter prompt following, a fixed audio reference path and stronger style rendering to ComfyUI.

Comfyanonymous published a LoHa that trains MOSS-Image-Base for 4 MP and larger canvases, shipping four checkpoint versions with 4096x2304 comparison samples.

ComfyUI-RMBG v3.3.0 adds a Canvas compositing studio, VRMBG video matting, FeyNoBg fine-hair matting, YOLO26 segmentation, and an offline Florence-2 rebuild.

FastVideo released FastH3 Omni, an 8-step MiniMax H3 distillation that turns image, video and audio references into one clip with stereo sound, plus ComfyUI weights.

Bria released Fibo 2.0, an 8B open-weights model that unifies image generation and editing in one set of weights, plus a 4-step Edit Turbo checkpoint.

CC BY-NC 4.0

Kromoss is a full finetune of MOSS-Image-Base, released as bf16 and int8 ConvRot single files for ComfyUI, with an 8-step DMD LoRA and conversion scripts.

Apache-2.0

KohakuMoss 0.1.0 trains LoRA, LyCORIS or full finetunes on MOSS-Image-Base and SFT, runs from 6 GB cards, and exports adapters in ComfyUI's key layout.

KohakuMoss License 1.0

Alissonerdx releases MiniMax H3 Character Swap v1.1, an 80 MB dynamic-rank SVD extraction with closer likeness, plus two ComfyUI workflows, one on native nodes.

BFS-Character-Swap-Community-1.1

Iris-3B Turbo, a 3B pixel-space diffusion transformer with no VAE, now runs through FastSDCPU's OpenVINO build on CPU and a Metal engine on Apple Silicon.

Two community LoRAs cut MOSS-Image-SFT from 50 steps to 4 or 8 network evaluations, and the 4-step rapid-moss adapter ships a ComfyUI node with native workflows.

Apache-2.0

Fizgig v8.4.0 adds LoRA Isolation, which pairs DOP with fused trigger tokens and negative-pair training to cut stranger face leakage from 48/50 to 0/50 on Krea 2.

ComfyUI-MangaScreen adds Manga Finish, Manga Panel and Manga Levels nodes that print halftone screentone manga pages, with a Qwen-Image 2.1 Turbo and a lite workflow.

Krea 2 Identity Edit V2 and General Edit V2 ship as separate LoRAs, adding reference placement to the ComfyUI-Krea2Edit node pack for ComfyUI.

Run MOSS-Image in ComfyUI: download moss_image_sft_bf16 or int8_convrot, load the qwen3.5_9b text encoder, and control layout JSON with the MOSS-Image Layout Prompt node.

Plenio 1.0 turns ComfyUI into a free local song studio: a language model writes the lyrics, YuE2 or MiniMax Music 3 performs them, and a new browser player plays and shares them.

Comfy Org's Open Source Higgsfield challenge puts $25,000 in prizes behind open-source rebuilds of apps like Genjutsu on the Comfy Developer Platform, open to October 30.

Black Forest Labs released FLUX 3 Image Klein: a 5.6B open weights image model for text-to-image, editing and multi-reference generation, with day-one ComfyUI support.