Creative AI Briefing: Wednesday, September 2, 2026
Cut a moving subject out of a shot by clicking a few points on it, on a laptop, with no account, no upload and no subscription. The models download when you ask for them and run on your own machine through OpenCV. That shipped Sunday in a free video editor, and it sits at one end of the week's real story: the price of getting finished work out of a generator moved in both directions at once. Free and local tools absorbed jobs that used to cost a monthly fee, while the hosted platforms started metering the download button and the most ambitious new thing anyone announced went straight behind a sales form.
New models
Solaris is the most interesting thing Runway has published in months, and you cannot use it. Announced August 31, it is the first of what Runway calls Interface World Models: instead of generating a video clip, it generates a running interface, one frame at a time, responding to your clicks and drags as they happen. There is no HTML underneath, no component tree, no code. You give it a starting frame, which Runway suggests composing from real product photography, and text prompts that define what a click or a drag means inside that particular scene. Click a cat and your next clicks paint its fur onto whatever you touch. Drag a shirt off a rack onto a photo of yourself.
The engineering is a direct descendant of the video work. Solaris is built on Gen-4.5, converted to real time in three moves: autoregressive frame generation so each frame depends only on what came before, distillation of the many-step denoising process down to a few steps, then training the fast model on its own outputs so quality holds over a long session. A separate language model sits above it deciding what the interface should do next and writing the prompts that tell Solaris how to render it. Reasoning and rendering are split.
Runway published a user study rather than a benchmark, which is the right call for this: 250 participants, 30 interaction examples, close to 7,500 pairwise judgments, comparing Solaris against a coded interface generated by Claude Opus 5 from the same starting image and the same request. Participants preferred Solaris 61% to 24% on following the instruction and 71% to 21% on behaving naturally in the scene. Read that as the authors' own evaluation of their own model, because it is.
The catches are stated openly in the post and they are large. Text is still unstable, which is a problem for a medium made almost entirely of text, and Runway's own suggested fix is a hybrid where image models render text-heavy views during pauses. Long sessions still drift. Accessibility and screen reader support are named as unsolved. And the availability catch swamps all of it: there is no pricing, no API, no public launch date, and the only door is an early-access request form asking for your company size (post).
Image
No image model surfaced across this run's sources in this window, and a fact-check pass could not turn one up either, so read that as nothing found rather than nothing existing. The one image item worth naming is a tooling update: the (Deno) Ideogram Director node got a small extension on August 31, letting its bounding-box output feed both standard BBOX consumers and BOUNDING_BOX inputs on multi-LoRA region nodes. That matters more than it sounds, because Ideogram Director is a visual JSON prompt builder where you draw labeled regions directly on the ComfyUI canvas instead of describing a layout in a sentence, and this change lets those regions drive per-region LoRA assignment. The repo's own warning is worth reading first: the region-row count follows the box count but does not track box identity, so delete or reorder a box in the middle and your LoRA assignments silently shift (repo).
Video
Deno Custom Nodes v0.7.100 landed August 31, and the version worth caring about is the run of releases behind it. On August 26 Alibaba's PAI group published MiniMax-H3-Acc-LoRAs, two official acceleration adapters that let MiniMax H3 generate video and audio in eight sampling steps instead of its normal count, using a technique called Parallel Decoding Distillation. Alibaba shipped them as Diffusers scripts requiring diffusers >= 0.40.0. Useful if you write Python. Useless if your work lives in a node graph.
Five days later they run natively in ComfyUI. The (Deno) MiniMax H3 Acc LoRA Loader takes the official safetensors unconverted, applies the static LoRA weights plus the checkpoint's 32 time-dependent output heads, then reads the actual sigma boundaries your scheduler produces and fuses the right heads for those intervals at sampling time. You wire the loader's single model output into a stock BasicScheduler at simple/8 steps and a stock KSamplerSelect at euler. Sampler and step control stay in normal ComfyUI nodes, which is a better design decision than it sounds, because the earlier versions of this node returned its own sampler and sigmas and broke every workflow saved against them.
Two things to know before you queue anything. The adapters are 8-step by training, and the pack lets you pick anything from 4 to 12 in the Simple scheduler, but only 8 is the configuration Alibaba trained and recommends. And the license on the weights is the MiniMax-H3 Community License Agreement, not Apache, not MIT. The node pack itself is GPL-3.0 (node pack, 159 stars; weights, created August 26 per the Hugging Face API).
Audio and music
The music story is a bill coming due. Suno introduces download limits tomorrow, September 3, announced August 10. Free accounts get 7 lifetime trial downloads, which do not reset monthly and carry personal-use rights only. Pro gets 20 per month and Premier 60 per month. Premier subscribers who also use Suno Studio, the browser production suite, face no download limits at all. Additional downloads will be purchasable.
The part that will surprise people is retroactive: the limits apply to every download starting September 3, including songs you made months ago. Your library stays playable and shareable on Suno, but pulling a file out of it now costs against a quota. Suno's stated reason is that limiting downloads makes it harder to mass-export music, and the same date brings new Terms of Service with revised commercial-use language and an updated arbitration agreement. Commercial rights still attach to anything downloaded on a paid plan (post, new terms).
Also on the board at one line: Luma announced on September 2, at LEAP in Riyadh, that Luma Agents will get a fully native Arabic interface covering navigation, menus and workflow tools rather than just prompting and output. It is the platform's first non-English interface. Availability is "later this year," so there is nothing to try yet (announcement).
Open and local
The local story is OpenShot 4.0, released August 30, and it is the clearest example this week of a free tool eating a paid feature. The headline for this audience is the new Object Mask effect: click a few positive points on a subject and a few negative points around it, get a detailed selection preview, and the mask follows the subject through the clip. Add more prompts on frames where the subject changes shape or gets occluded. The finished mask then becomes a reusable source for Blur, Pixelate, Color Grade and other effects, so you can treat the subject and the background separately.
It runs on your machine. OpenShot maintains a separate repository, OpenShot/openshot-onnx, that exports three model families to ONNX so libopenshot and OpenCV can execute them without Python or PyTorch: YOLO for detection and instance segmentation, EfficientSAM for the prompted seed mask, and Cutie for propagating that mask across frames. The export code is MIT; the README is explicit that the upstream weights keep their own terms, which is the honest disclosure most projects skip.
The rest of the release is not AI and matters anyway: a full color view with wheels, four editable curves, .cube LUT support with an intensity blend, and Luma Waveform, Histogram, RGB Parade and Vectorscope scopes with a region tool. Screen, webcam, microphone and system audio recording, each captured as its own separate clip. Ten new effects including audio-reactive visualizations, beat sync, film grain and a displacement map. GPLv3, Windows, macOS and Linux, free (post).
Star totals below are cache-busted shields.io reads taken September 2.
- OpenShot/openshot-qt: free GPLv3 editor that now does local subject masking, color grading and screen recording without an account (6.4k) (repo)
- OpenShot/openshot-onnx: the export utility that turns YOLO, EfficientSAM and Cutie into ONNX files OpenCV can run on your machine (MIT code, upstream weight terms apply) (repo)
- Deno2026/comfyui-deno-custom-nodes: 8-step MiniMax H3 sampling, a visual bbox prompt builder for Ideogram 4, and a text-encoder unload node for tight VRAM (159) (repo)
- aigc-apps/VideoX-Fun: Alibaba PAI's video toolkit, where the official Acc-LoRA prediction scripts live (2.2k) (repo)
- comfyanonymous/ComfyUI: the node graph everything above loads into, and the source of the native MiniMax H3 support these adapters attach to (131k) (repo)
- hradec/ComfyUI-HR-Endless-Sampler: arbitrary-length H3 video on a 16 GB card with a language model directing continuity between chunks (74, up from 72 yesterday) (repo)
- Kosinkadink/ComfyUI-VideoHelperSuite: video load and combine nodes that most of these workflows depend on for audio and frame handling (repo)
Creative workflows
1. Generate MiniMax H3 video and audio in eight steps instead of many, inside ComfyUI. Node pack at Deno2026/comfyui-deno-custom-nodes. Weights at alibaba-pai/MiniMax-H3-Acc-LoRAs. Note on versions: the README states a ComfyUI 0.30.0 floor for the pack's two MiniMax H3 reference-image nodes, and states no version floor at all for the Acc LoRA Loader itself, so you need whatever build gave you native H3 support.
The steps. Install the pack through ComfyUI Manager, search Deno Custom Nodes, restart. Download one of two files depending on your pipeline: MiniMax-H3-FL2VA-Acc-8Step.safetensors for first-and-last-frame or text-to-video work, or MiniMax-H3-Ref2VA-Acc-8Step.safetensors for reference-image work. Both are rank 64, network alpha 64, BF16. Put the file in ComfyUI/models/loras/ or the dedicated ComfyUI/models/minimax_h3_acc_loras/ folder. Add (Deno) MiniMax H3 Acc LoRA Loader, connect your native H3 diffusion model to its model input, pick the matching adapter, and send the loader's single model output to your guider. Build the sampling lane from stock nodes: BasicScheduler at simple, steps 8, into KSamplerSelect at euler, into SamplerCustomAdvanced. Leave the native H3 video and audio sigma shifts at 12.0 and 3.0. Leave LoRA strength at 1.0.
How it works. The adapter carries static LoRA weights plus 32 time-dependent Parallel Decoding Distillation output heads, one set per training interval. At sampling time the node reads the descending sigma boundaries your scheduler actually produced and fuses the heads that cover those intervals, so the adapter adapts to your schedule instead of the other way around.
Why it is good. It uses Alibaba's official file unconverted, so you are not trusting a community requantization. Sampler, scheduler and step count stay in stock ComfyUI nodes you already know. And the repo hosts side-by-side result videos against both the un-accelerated baseline and the community Minimax-h3-Turbo 4-step checkpoints, so you can judge the quality cost before downloading anything.
Where it breaks. Three places. If you run a Comfy-Org *_pruned_* checkpoint, which is exactly what low-VRAM users grab, the loader needs a matching non-pruned checkpoint present under models/diffusion_models/ to derive an AdaLN bridge; without one it warns once and silently skips 50 of the adapter's updates in a documented compatibility mode. Step counts from 4 to 12 are selectable but only 8 is trained. And workflows saved with the older v0.7.92 to v0.7.94 three-output loader migrate once on the canvas and then need a manual resave, with raw API prompt JSON not migrating at all.
2. Roto a moving subject with local models, no account and no cloud. OpenShot 4.0, free, GPLv3.
The steps. Install OpenShot 4.0. Drop your clip on the timeline and add the Object Mask effect. Click positive points on the subject and negative points on the surrounding area. Review the generated selection preview, then process it through the clip. On frames where the subject deforms or disappears behind something, add more prompts. Then point Blur, Pixelate or Color Grade at the finished mask, or invert it to work on the background instead. The models download on demand from OpenShot/openshot-onnx the first time you ask for them.
How it works. Three separate models with three jobs. EfficientSAM turns your clicked points into a seed mask on one frame. Cutie propagates that mask forward through the video. YOLO handles detection and instance segmentation for the separate Object Detection effect, which also gained downloadable ONNX models and segmentation masks in this release. All three are exported to ONNX so libopenshot and OpenCV can execute them directly, which is why no Python or PyTorch install is required.
Why it is good. Your footage never leaves the machine, there is no AI account, and there is no subscription gate on the feature. The mask is reusable rather than baked, so one roto pass feeds several effects.
Where it breaks. OpenShot publishes no VRAM or timing figures for masking, and says plainly that performance depends on your computer and the chosen model. The point-prompt loop is interactive, so a subject that changes shape often will cost you a lot of clicking. The openshot-onnx README also notes that generated model binaries are not in Git and that upstream model licenses govern the weights, so check terms before shipping commercial work built on a specific checkpoint.
Worth testing
- The 8-step Acc-LoRA in a browser, before you download a 1.37 GB safetensors file. Three community Spaces run these adapters hosted, including
mezroui/MiniMax-H3-Acc-LoRAs-demo, which is up and running. Honest tradeoff: these are user-built Spaces, not Alibaba's, so queue times and uptime are whatever the owner is paying for (Space). - OpenShot 4.0's Object Mask, if you have ever paid for a roto tool. Tradeoff: no published performance figures, and the interactive prompting gets tedious on complex motion (download).
- DENO Video Compare in a browser, no install, for A/B-ing two renders with a slider, side by side, difference view or toggle. Tradeoff: it is a static GitHub Pages tool, so everything runs client-side and large files will tax your browser (tool).
- Suno downloads, today. If there is anything in your Suno library you actually want as a file, the free tier's lifetime allowance starts counting tomorrow. Tradeoff: none, this is just a deadline (policy).
What actually matters from today's signal
Two lines crossed this week. On one side, a free GPLv3 video editor added subject masking that runs on your own machine with downloadable ONNX models and no account, and a hobbyist node pack made an official Alibaba acceleration adapter loadable in ComfyUI five days after it shipped as Python scripts. On the other, Suno starts charging against a quota tomorrow for the act of downloading a file you already made, retroactively, and Runway published the most ambitious generative idea of the week behind a form that asks for your company size.
The practical read for your money is that the meter is moving to the export step. Generating has been getting cheaper for two years and it kept getting cheaper this week; eight steps instead of many is a straightforward cut in your own electricity bill and your own waiting. But the platforms that host generation have noticed that generation is not the scarce thing anymore. Possession of the file is. Suno's own explanation for the limit is that it makes mass export harder, which is honest, and the consequence is the same either way: on a hosted platform, plan around downloads, not around generations.
The counter-argument to my own framing is worth stating. Solaris is not expensive because Runway is greedy; it is expensive because generating every frame of a running interface genuinely costs more than serving a page that was built once, and Runway says so in the post. Early access forms are what new and costly things look like before they are cheap. The same was true of video generation three years ago. So the honest version is not that hosted tools are closing down. It is that the gap between what you can run yourself and what you have to rent keeps getting redrawn, and this week it moved in your favor on the desktop and against you in the cloud. If you make things, the lesson is boring and durable: keep the files.
Source access notes: Sandbox network egress is unavailable, so every fetch went through the managed fetch tool; no curl or wget was used. blog.comfy.org remains a JavaScript wall, and a direct post fetch resolved to a December 2025 article, so it was discarded. GitHub HTML release pages served stale content again this run; ComfyUI release dating was taken from docs.comfy.org/changelog instead, and all cited files were read through raw.githubusercontent.com with cache-busting query strings. github.com/trending, Trendshift, the Diffusers and llama.cpp release pages were not reached this run. Reddit returned no indexed results through search. midjourney.com/updates was not directly reachable and nothing was included from it. stability.ai/news returned content with no reliable date stamps and was not used. higgsfield.ai/blog exceeded output limits and was not confirmed clean for the window. blog.google/technology/ai/ redirected to an undated hub. One live discrepancy worth naming: the Hugging Face HTML page for alibaba-pai/MiniMax-H3-Acc-LoRAs showed 168 likes while the API showed 182, and the card's "downloads last month" read 27,009 against the API's 32,893 total; API values were used. Every Hugging Face date in this briefing comes from the API createdAt field, not a listing timestamp. Sections omitted for having nothing worth saying: Writing.
An adversarial fact-check pass ran against this draft before publication and caught five things, all corrected above. The Acc-LoRA repository's 2.8 GB usedStorage was being described as "weights"; the two safetensors are 1,372,450,680 bytes each and you download one, so the figure is now stated as 1.37 GB and the demo videos are no longer counted as weights. The Suno line narrowed an unlimited-downloads benefit into a stems-and-samples benefit; Suno's post says Premier subscribers who use Suno Studio face no download limits at all, and the text now says that. The ComfyUI 0.30.0 floor was attached to the wrong node: the README states it for the pack's two MiniMax H3 reference-image nodes and states no floor for the Acc LoRA Loader, and both mentions are now scoped correctly. The claim that nothing shipped in image generation could not be independently confirmed and is now framed as nothing found. The assertion that the OpenShot models run through OpenCV's DNN module specifically was an inference beyond what the openshot-onnx README says, and has been softened to what the README states. Everything else the checker attacked held, including all four Solaris user-study figures against the Runway post, every Deno README setting, the Hugging Face createdAt and license, the OpenShot model division of labor, the Suno tier numbers and retroactivity, and all five cache-busted star counts.