Dropped Friday, September 5 — RefMods are roughly 1 MB conditioning adapters for MiniMax H3 that let you encode a character once and then load it like a LoRA, with no fine-tuning run at all. For anyone burning hours on character-consistency LoRAs, this collapses the identity pipeline from a training job to a file. Shipped alongside an 11 MB H3 Semantic Bridge that ports SenseNova U1.5 semantics into H3's conditioning space for better prompt adherence.
OpenAI's Sora API reaches end of life on September 24, 2026, following the web and app shutdown back on April 26. Any production pipeline still calling Sora endpoints has under three weeks to cut over to Veo, Kling, Runway, or Seedance. Budget for prompt rewrites, not just endpoint swaps — Sora prompt structure does not transfer cleanly to Veo 3.1 or Kling 3.0.
Released September 2, this workflow pairs H3 Motion Context with a latent-saver node to chain clips into arbitrarily long videos with lip-sync, and it runs on a 12 GB GPU. That's the single biggest unlock for narrative and faceless-channel creators who've been stuck stitching 6-to-15-second fragments in post. Consumer-hardware long-form is now a workflow problem, not a VRAM problem.
A 7B joint audio-video generator using gated cross-modal attention and progressive joint training, plus a separate 5B SR-DiT autoregressive one-step refiner that pushes output to 2K. Critically, it's Apache 2.0 — no revenue caps, no territory exclusions, none of the MiniMax H3 community license baggage that locks out US, EU, UK, and Korea weights users and gates anyone above $20 million in revenue. Audio and video are modeled jointly, not bolted on in post.
Three separate drops last week attacked H3 latency from different angles. September 4 brought an 8-step 768p Ref2VA distillation with audio; September 1's VDN-H3 hybrid attention renders a 14.4-second 768p clip in 8 steps; and September 3's H3 CLIPCached node cuts conditioning from 29.9 seconds to 1.1 seconds via disk caching. Stack all three and iteration cycles stop being the bottleneck.
Friday's 0.34.5 release added Comfy Cloud beta nodes running curated Flux 2, Z-Image Turbo, Mage Flow, MiniMax H3, and Music 3 workflows on hosted GPUs. Version 0.34.4 the day prior added Meta Muse Image text-to-image and edit nodes with up to 10 input images and 14 aspect ratios, Recraft V4 Styles Pro, and a MiniMax H3 Max Turbo option. Hybrid local-plus-cloud graphs are now first-class.
Current data puts AI-generated YouTube Shorts at roughly $0.03 to $0.10 per 1,000 views, so a million-view video nets $30 to $100 at a 45 percent creator share. Mid-tier faceless channels doing 10 to 50 million monthly views land in the $200 to $2,000 range. Long-form in the same niche runs $3 to $20+ RPM — up to a 60x gap — which is the strongest argument yet for pointing that new unlimited-length workflow at long-form.
Released September 2, Provcheck adds an offline C2PA signing node with neural watermarks for both image and audio plus Content Credentials bound to Bluesky. Timing is not coincidental — EU AI Act Article 50 transparency provisions became enforceable on August 2, 2026, requiring clear labeling of AI-generated content wherever it's published. TikTok, Instagram, and YouTube all now permit monetized AI content, but only with proper disclosure labels applied.
A community blind-test leaderboard went live September 2 comparing 23 different MiniMax H3 acceleration variants across roughly 4,000 generated clips. This is the kind of independent evidence that's been missing — most speedup LoRAs ship with cherry-picked demos and no quality-loss accounting. Check it before you commit a production pipeline to any given distillation.
A 65.6M-parameter LoRA plus a directed-attention patch maps WASD and camera keys to action-controlled video from a single starting frame — that's 0.199 percent of the 33B backbone, trained on 8,000 gameplay clips. Character keys are WASD, camera is IJKL, F for fast moves. Fair warning: it needs the attention patch, so loading the LoRA into a stock H3 pipeline won't reproduce the results. Apache 2.0 checkpoint.
