10/16/2026
MOSS-Image Drops 5B Open Weights
OpenMOSS ships MOSS-Image — a 5B DiT with Apache-2.0 weights, 4K-token prompts, and native JSON layout control. Run it today via Diffusers, Gradio, ComfyUI Manager, ModelScope, or HF Spaces.
Aurelia
265 Articles
Latest Contributions
10/16/2026
Krea 3 Lands in ComfyUI Partner Nodes
ComfyUI just shipped a hosted Krea 3 Partner Node — text-to-image and instruction edits with up to 8 refs, 1K/2K, and precise-edit mode, billed on Comfy credits.
10/16/2026
OmniHuman V2 on fal: 5-Minute Talking Videos
ByteDance's OmniHuman V2 is live on fal: one portrait plus up to five minutes of audio becomes a lip-synced talking video, at $0.105 per second, about a third cheaper than v1.5.
10/16/2026
ComfyUI v0.41.0: FLUX.3 Klein, HDR & Multi-Ref
ComfyUI v0.41.0 is tagged with native FLUX.3 Image Klein support, a new SDR-to-HDR node, a 100-slot reference-latent node for edit models, and MiniMax H3 LightVAE loading.
10/15/2026
FLUX 3 Image Klein: 5.6B Open Weights
Black Forest Labs open-weights FLUX 3 Image Klein — 5.6B params, native 4K, bounding boxes, up to 10 refs. Day-0 on ComfyUI, Diffusers, fal, and the BFL API from $0.009/1K.
10/15/2026
Kling 4.0 Fully Available on App, MCP & CLI
Kling says 4.0 is fully available on the refreshed kling.ai creator UI plus MCP and CLI. HDR output and the public/enterprise API are still Coming Soon — confirm your model in-app.
10/15/2026
Krea 3 API Hits Early Access: Specs and Pricing
Krea just documented the Krea 3 API: an instruction-following image editor with up to 8 references, pixel-exact regional edits, 1K/2K output, and pricing from $0.03 an image. Beta workspaces only, for now.
10/15/2026
ATAS: Screenplay-to-Film That Remembers Sets
ZIP Lab open-sourced ATAS — MIT code that turns a screenplay into a multi-shot film with persistent scene and event memory, defaulting to Wan 3.0 on Alibaba Cloud.
10/15/2026
Kling 3.0 and O3 Land on Runway's API, Up to 4K
Runway Dev now serves Kling 2.5 Turbo Pro, Kling 3.0, and Kling O3 through one API — multishot prompts, 7-image references, and native 4K from $0.09 a second.
10/14/2026
FLUX 3 Video: 4K Now ~Half Price, Up to 2x Faster
Black Forest Labs just made native 2K and 4K FLUX 3 Video up to 2x faster and cut the price — 4K text-to-video drops from $0.80/s to $0.41/s on the BFL API.
10/14/2026
FlashVSR+ One-Step 4K Upscale on One 4090
OpenImagingLab’s FlashVSR+ is open (Apache-2.0): one-step diffusion video super-res with untiled 4K on a single 24GB RTX 4090 — Turbo and Base weights on Hugging Face.
10/14/2026
Mirelo SFX 2.0 Keeps Dialogue, Replaces Music
Mirelo’s new video-to-sound model preserves speech, strips unwanted music, and generates action-synced SFX/ambience up to 10 minutes — live on Mirelo, fal, and Runway.
10/14/2026
EntroPack: Fit Qwen-Image & H3 to Your Exact VRAM
ModelScope open-sourced EntroPack, a calibration-free compressor that lets you pick any bits-per-weight, plus ready-made Z-Image-Turbo, Qwen-Image-2.1, and MiniMax-H3 packs from 2.3 to 8 bpp.
10/14/2026
ComfyUI v0.40.0: LoRA Stack, Cameras & Portables
ComfyUI v0.40.0 is tagged with Windows portables (including a new NVIDIA ARM64 build), native LoRA stack loaders, CameraAngle nodes, and a fat partner-node tray you can pin.
10/13/2026
FastH3-Omni: Real-Time H3 Video + Audio, Open Oct 20
FastVideo, Nuva Lab, and Reactor just previewed FastH3-Omni, a sped-up MiniMax H3 that makes video and synced sound from text, image, and audio references. It's on the Reactor API now, with open weights promised for October 20.
10/13/2026
fal Reframe: H3 Max Turns Any Video Into Any Ratio
fal just launched Reframe, an H3 Max endpoint that outpaints your footage to a new aspect ratio while keeping its timing and audio, from $0.05 per second for clips up to five minutes.
10/13/2026
Qwen-Image-2.1 Lightning: 4-Step LoRA for Gen + Edit
LightX2V just dropped Lightning for Qwen-Image-2.1: a 67 MB distilled LoRA that squeezes text-to-image and multi-reference editing into four steps instead of forty.
10/13/2026
Higgsfield Genjutsu API: Motion Transfer + Comfy Node
Higgsfield's viral Genjutsu motion transfer is now a public API at 50% off through October 31, and ComfyUI just merged a Genjutsu partner node: one source clip, up to eight references, a brand-new cast.
10/13/2026
LightVAE: Faster Wan & H3 Video Decode
ModelTC's LightX2V team just open-sourced LightVAE: slimmed video VAE decoders for Wan2.1, Wan2.2 and MiniMax-H3 that decode up to 2.94× faster, with a ComfyUI node and six workflows ready today.
10/13/2026
Iris-3B Turbo: 4-Step Pixel-Space Images in 0.7s
Speridlabs just dropped Iris-3B Turbo: a 4-step distilled, Apache-2.0 text-to-image model that paints pixels directly (no VAE) and lands a 1024px image in about 0.7 seconds on an H100.
10/12/2026
Helix World: Real-Time Video Worlds With Stereo Sound
Noiz AI and HKUST officially launched Helix World: drop in one image, steer with WASD, and get 24 fps video plus camera-aligned stereo sound that turns when you do. HelixBench lands with it today.
10/11/2026
Vivix W1: Real-Time AI Video From $0.003 a Second
Vivix opened its W1 video model to everyone with a playground, a public API, and a promo rate of about a third of a cent per generated second. Here's what it does, what it really costs, and how to call it.
10/10/2026
Midjourney Opens Test Group for a Midjourney MCP
Midjourney is recruiting a limited community to test a Midjourney MCP that lets AI assistants like Claude and ChatGPT drive its models, and its application hints at a paid MCP tier with bundled credits.
10/10/2026
HiDream-O1-Video Hits #6 on AA I2V With Audio
HiDream's native omnimodal video model just debuted at #6 on Artificial Analysis' image-to-video-with-audio board, with 1080p clips of 5 to 20 seconds at about $0.10 a second on the HiHarness API.