OpenAI's Videos API and all Sora 2 model aliases are removed on September 24, 2026, two days from now. OpenAI's own deprecation table leaves the "recommended replacement" column blank — there is no in-house successor, so every migration path goes to a rival vendor. Veo 3.1 is the common landing spot at roughly $0.03/sec for 720p up to $0.60/sec for 4K with audio; anyone with saved generations should export at sora.chatgpt.com/sunset before the lights go out.
Video Rebirth dropped HyperFlow, a data-free self-distillation LoRA that collapses H3 sampling from 49 transformer forwards down to 8, and it hit Banodoco's #minimax_h3_chatter on September 18. Two ComfyUI ports landed almost immediately — drbaph's conversion and Saganaki22's node pack, which rides ComfyUI's native MiniMax-H3 model with no custom sampler and no core file edits. drbaph's pitch is that it holds closer to true H3 quality on consistency, camera control and detail than the older turbo LoRAs.
Saturday's v0.37.0 adds native Qwen-Image 2.1 support with three official templates, MoGe 3 geometry estimation, faster Qwen3.8 text encoders, and automatic fast disk loading. MoGe 3 arrives via kijai's PR as MoGeModelV3 — MoGe-2's architecture plus a sparse 3D UNet that iteratively refines predicted log-depth — shipping as moge_3_vitl_fp16 and moge_3_vitg_fp16 with a new refine_steps control from 0 to 8, defaulting to 3.
Alibaba released Qwen-Image 2.1 on September 20: a 7B visual generator with native RGBA transparency through a 64-channel VAE, editing from up to 10 reference images, and native 2K output — all in one model that fits a consumer GPU. Diffusers, ComfyUI, vLLM-Omni, SGLang and LightX2V all had support on release day. The catch is the Qwen Research license — open weights, not open source, so you cannot legally ship it commercially without a separate grant from Alibaba.
Cseti's CrossView-Warp landed Sunday as a Ref2VA LoRA plus a ComfyUI node that re-renders existing footage from a different camera position using a MoGe depth warp. That's a genuine coverage tool — you shoot once and pull a reverse or a wide from the same take instead of re-rolling the generation and praying for character consistency. It ships under the MiniMax-H3 Community license.
Sunday's Hugging Face trending snapshot has MiniMax-H3 at 5,532 likes and 4.06 million downloads, with Lightricks' LTX-2.5 at 4,567 likes and 1.61 million. A community repack, Minimax-h3_Singularity, is already at 574 likes and 242K downloads billing itself as a drop-in upgrade with improved motion coherence. LTX-2.5 is the 22B dual-stream audio-video model that does a 10-second clip from an image in roughly 6.8 seconds on NVIDIA superchips, free under $10M ARR.
TikTok has re-weighted the For You Page to deprioritize anything its synthetic media classifier flags, and AI-generated content is explicitly barred from the Creator Rewards Program for failing the originality test. YouTube's automatic AI detection went live in May 2026 — no disclosure plus detected photorealistic AI equals an auto-applied label. Instagram renamed its "AI Creator" badge to "AI-Generated Profile," and unlabeled synthetic personas are seeing throttled reach.
ComfyUI-MiniMax-H3-SPEED shipped V2 on September 20 with support for Euler, Heun, DPM2 and RES Multistep, plus a sampler-aware Sigma Harvest for progressive-resolution runs. It also patches the image-to-video keyframe resizing that was quietly cropping conditioning frames. Pair it with the September 19 video VAE kernel work in 0.36.0 and the September 6 sparse attention node, and long H3 runs get materially cheaper on VRAM.
SupraLabs posted Supra2-IMG yesterday — a 105-million-parameter DiT text-to-image model trained entirely from scratch in ten hours on a single H100, with samples at 256 by 256, under Apache-2.0. Nobody's shipping client work at 256 pixels, but the recipe is the point: a genuinely reproducible from-scratch training run at hobby budget, fully permissive, with no research-license asterisk.
