fal dropped H3 Max Turbo in preview, and the numbers are interesting. It runs at twice the speed of H3 Max, costs half as much, and sits at the 97th percentile of H3 Max quality on internal evals. The promotional price is one cent per second for 768p output. They also extended the H3 Max discount another 75% through September 7th, bringing text-to-video and image-to-video down to two cents per second at 768p.
Also from fal: Fibo Edit 1.5 from Bria is now live on the platform. It takes up to four reference images at once, which opens it up for product placement, virtual try-on, background swaps, and multi-product pack shots in a single pass.
Runway had a packed release cycle. GWM Worlds 2 is their general world model, and it generates interactive real-time video at 720p, 24 frames per second, with 48 kHz audio. You can feed it text actions targeting specific subjects or entire scenes, move the camera continuously, and there's no fixed session length. That last part matters for anyone building interactive experiences on top of it.
They also shipped Runway Dev MCP, which lets coding agents work directly with the developer platform. Setting up Model Routers, looking up tasks, managing the platform, all through an agent. And inside Runway Ruby, ACEScg 1.3 and 2.0 EXR sequence export is now available in scene-referred half-float format, which is a meaningful addition for anyone doing color-managed post work.
Moving over to ComfyUI, there's a lot happening natively now. Recraft V4 Styles landed with built-in support, letting you build reusable styles from anywhere between one and ten reference images. Texture, composition, rendering characteristics, all of it carries through. Recraft also published new V4 Styles directly: Low Poly Render Icon, Crochet Lace, and Inflated 3D Gradient Icon.
ComfyUI also folded in Trellis.2 and Pixal3D at the core level. No custom nodes, no extensions. You get load, preview, and save nodes for 3D assets, mesh post-processing, and PBR texturing with normal map and ambient occlusion baking. That's a significant jump for 3D workflows inside the graph.
Magnific made Seedance 2.0 and 2.5 available with side-by-side prompt comparisons so you can see how the two versions handle the same input. Krea opened beta access to Krea Agents, positioning it as a platform built around AI agents for professional creative work.
Now to the open-source side of video generation. MiniMax released FastH3 publicly, integrated with Reactor. It streams 720p video with synchronized audio at 0.7 cents per second and supports first-frame image conditioning. The haoailab team paired it with vLLM-Omni on NVIDIA hardware and rendered a full 10.1-second MP4 with both video and audio in 8.7 seconds. Faster than real time.
Someone at fal wired Minimax H3 Max to a live chat interface and let viewers inject prompts mid-stream. The result was continuous video that snapped between scenes on command, described as interdimensional cable. The platform banned the stream more than once because the content was uncontrolled. It's a compelling proof of concept of what live prompt injection into a video stream looks like in practice.
Orbis is worth watching. It's presenting itself as the first video model that accepts live prompt edits while generation is streaming, and it holds scene state after the camera moves away for 30 seconds without regenerating what was already there. Scene state persistence is the part the field has been chasing.
And on a different note entirely, President Trump posted an AI-generated video depicting the destruction of Kharg Island. That's now part of the public record.
That's your AI digest for 03 Sep 2026.