ComfyUI News & Open-Source AI Releases

NVIDIA's LongLive-Plug distills few-step, CFG and long-context LoRAs once per backbone and reuses them across 54 downstream models, with weights for MiniMax H3 and Wan.

ComfyUI v0.38.0 adds native Ming-Image support with an official template, lands Wan ID-V2V, w6a8 quantization, HDR LogC3 and ACEScct color spaces, and speeds up SeedVR2 and YuE2.

Lightricks' Native Resolution brings LTX-2.5 4K and 8K video edits to ComfyUI through Tiled Fusion, with new Refine and Restore IC-LoRAs.

LTX-2.x Community License

MATLOW's ORB360 LoRAs give MiniMax H3 Ref2VA a full clockwise 360 camera orbit around a frozen subject, plus a CardSpin variant that turns the photo itself in space.

SenseTime's OpenSenseNova released Looped-DiT, which reruns shared transformer blocks inside each denoising step so a small model can beat a much larger one.

MIT

AcademiaSD LoRAlab Trainer Studio bundles LoRA trainers for Qwen-Image 2.1, FLUX.2 Klein, Krea 2, Z-Image, MiniMax-H3 and more behind one launcher, from 4 GB of VRAM.

MIT

AnyAngle is a Qwen-Image 2.1 LoRA that moves a photo to any camera angle: render the scene from a Gaussian splat or 3D model, then let the edit model carry the view over.

LanPaint 2.2.0 brings Qwen-Image 2.1 under its masked sampler: inpaint the text-to-image or edit model, and let transparency change shape along with the pixels.

Vlo 0.3 is a free, open-source video editor that drives ComfyUI from the timeline: MiniMax H3 and Qwen workflows, SAM2 masking, grading and frame-accurate export.

Turbo8 distills Qwen-Image 2.1 to 8 steps with no CFG, covering text-to-image, editing, RGBA output and subject extraction, with two ready ComfyUI workflows.

A community LoRA for Qwen-Image 2.1 keeps edits on the original frame: the same prompt and seed come back without drift or stray repainting, with two ComfyUI workflows behind it.

WarmBloodAban's Qwen-Image 2.1 LoRA pack pairs an anime character-consistency adapter with one that turns illustration into photographic characters.

Apache-2.0

A community LoRA for MiniMax H3 FL2VA turns one photo into a full 360 camera orbit that lands back on its own first frame, with ideal ComfyUI settings.

Fizgig v6.5 adds Qwen Image 2.1 LoRA and LoKR training with the studio's own training adapter, turbo previews and a new driver system, saving LoRAs ComfyUI loads.

A MiniMax H3 LoRA that morphs one picture into another: elements, materials and subjects transform into one another across a seamless first-and-last-frame transition in ComfyUI.

Akatz Labs' character swap LoRA for MiniMax H3 Ref2VA replaces one person in a reference video with a character image or sheet, and ships the weights and dataset publicly.

ComfyUI-Fizgig-H3-Still adds two nodes that build a true one-frame MiniMax H3 latent and decode it without banding, for clean stills from 3 MP up.

Draft mode previews a Seedance 2.5 clip at 480p and renders the 1080p final from the same draft. ComfyUI adds a Draft model option, a Draft to Final node and three templates.

API-Only

BFS Best Face Swap ships a Qwen-Image 2.1 head-swap LoRA with a ready ComfyUI workflow, plus experimental LTX-2 in-context head-swap weights.

MIT

A Krea-2-Raw LoRA that repaints the black-masked region of an image from a text prompt, in mild, default and strong variants, with four example ComfyUI workflows.