Creative AI Briefing: Thursday, September 24, 2026
Wrap your prompt in one sentence about alpha channels and Qwen-Image-2.1 hands back a PNG with a transparent background, cut by the model rather than by you. The exact wording sits inside the official ComfyUI template that got its workflow links published yesterday, and it runs on weights on your own drive. Hold that next to the other thing happening today: OpenAI switches off the Sora API, ElevenLabs has already pulled both Sora models from its picker, ByteDance has dated Seedance 1.5 Pro for November 11, and Adobe finished buying Topaz Labs on Wednesday. Two video models are going dark on somebody else's calendar, and a tool many photographers bought outright now belongs to a subscription company. The files on your disk did not change.
New models
FLUX 3 Action landed Tuesday, September 22 (HF createdAt 2026-09-22T10:36:43Z), with Black Forest Labs' write-up following Wednesday. This is not an image model, and BFL shipping it is the interesting part. You give it one camera frame plus a written instruction, and it predicts the next stretch of video frames and the control inputs that produce them, jointly. The creator-facing edge is the worked example in the repo: the same checkpoint plays two browser games, a shooter and a racer, and the caption you write is what selects which game it thinks it is playing. Action dimensions are four and four (move, strafe, turn, fire; steer, throttle, nitro with one dimension padded out). A drone entry named rotor reads its instruction straight out of each episode's index.json.
Licence is the FLUX Kommunity License, named in the card and sitting in LICENSE.md in the repo, so read it before you plan anything commercial. Two policy repositories, DROID and SO-101, share one frozen encoder set from the base repo, and the DROID policy ships in six packages including a one-step variant. The honest catch is scale: the published training command is torchrun --nproc_per_node 8, eight cards, and the games themselves (simulator, bot, recorder, player) live in separate repositories that are not in this one, so the example is documented rather than one-click. Free hosted demo exists and is worth ten minutes: a simulated SO-101 arm on ZeroGPU. Code: black-forest-labs/flux-action.
Image
The official ComfyUI packaging of Qwen-Image-2.1 picked up its workflow template links yesterday morning, September 23, which is the moment a model stops being a pile of files and starts being something you can run in fifteen minutes. The repository itself went up September 15. What arrived since is more interesting than the templates. On September 20 Kijai uploaded two separate prompt-expander encoders, qwen3.5_9b_qwen_image_2.1_pe_t2i.int8_convrot.safetensors and its pe_i2i sibling, plus qwen3vl_8b_w4a8.safetensors, a four-bit version of the prompt reader. So the thing that reads your sentence is now a swappable part with three sizes, which is exactly the lever Monday's briefing said mattered most for memory.
Watch the gap, though. The repository README lists all three and says they are for prompt enhancement through a separate text-generation node, but they appear in neither the workflow's own model-links note nor Comfy's documentation page for these templates, and nothing in the three shipped graphs is wired to use them. You can download them, file them correctly, and still have no workflow that touches them. Licence on the packaged files is qwen-research, per the card, pointing at Qwen's own LICENSE. That is the same research-and-evaluation restriction that has been hanging over every transparent PNG the community has made with this model all week, and repackaging it for ComfyUI did not change it. Community Spaces are running it if you want to judge output first, several of them linked off the repo page.
Video
The Sora API is discontinued today. OpenAI's help page is unambiguous: "The Sora API will be discontinued on September 24, 2026," with the web and app experiences already gone since April 26. If you made anything in Sora, export it now at sora.chatgpt.com/sunset, which is the Export button OpenAI points you at. The reason for the hurry sits on the help page rather than the export page: after any final export window passes, OpenAI says it will "permanently delete any data associated with your use of Sora." No successor is named.
The downstream version of that is more instructive than the announcement. ElevenLabs' changelog for September 23 records both Sora models removed from its Image and Video picker, and adds the sentence worth pinning above your desk: flows and templates that use a Sora node "need to be switched to another video model... before they can run again." Your saved generations survive. Your saved pipelines do not. The same entry dates Seedance 1.5 Pro for retirement on November 11, no longer available for new generations as of now, with Seedance 2.0 as the replacement.
Adobe closed its acquisition of Topaz Labs on Wednesday, September 23. Adobe says the Topaz brand stays, the apps and models "continue to be available as standalone products," and that Topaz technology is already available in Firefly and Photoshop. Topaz's Neurostream lets you process on your own machine or in Topaz's cloud, which is the detail that decides whether this stays useful to people who work offline. Adobe's post names no price; the roughly $340M figure circulating comes from press coverage, not from Adobe. Topaz CEO Eric Yang joins Adobe's Digital Video and Audio team. If you restore archival footage, clean up low-light shots, or push AI-generated video up to delivery quality, this is the upscaler you already use changing owners.
Smaller and free: Adobe Premiere mobile is now on Android as of September 22, at no cost, about a year behind the iOS app.
Audio and music
Basis Conversations 1500 opened to access requests Wednesday, September 23. It is a speech dataset, and the specific thing it has that the usual open corpora do not is overlap. Conversations seat two to four people talking at once, each one on their own separated channel at 48 kHz, across 22 languages, with a median conversation length of 33 minutes and 2,645 distinct speakers. If you are building anything that has to survive real human turn-taking, a dubbing tool, a podcast cleanup pass, an interview transcriber, a voice agent, the failure mode is always people interrupting each other, and this is training and evaluation material aimed squarely at that.
Two practical notes. Access is gated with manual human review, and the request form asks for institution and work email, so it is not a click-and-download. Licence is basis-data-license-1.0, sitting in the repo as LICENSE, and the gate text requires you to agree not to identify, locate or contact any speaker, and not to use the data to impersonate a specific person. You can hear samples first without requesting anything at the project's listening page. The repository was created September 15; the announcement and the open gate are Wednesday's.
Open and local
The local case this week is not about speed or quality. It is about the fact that a model file on your own drive cannot be removed from a picker. Every hosted item above has a date attached to it or a new owner; the Qwen and FLUX weights do not. That is not an argument for running everything locally, which is slower and fussier and eats your afternoon. It is an argument for knowing which of your finished deliverables depend on a model you do not possess, and it is a cheap thing to audit: open your five most-used graphs and count the API nodes.
- comfyanonymous/ComfyUI: 135k stars. The node graph most local image and video work runs inside. Trending because it is where every one of this week's weights actually gets used.
- Comfy-Org/workflow_templates: 991 stars. The official template repo, and the source of the two Qwen-Image-2.1 graphs linked from the model card yesterday. Small star count, disproportionate usefulness: these are the reference graphs, in plain JSON, that you can read before you download 90 GB.
- Comfy-Org/ComfyUI-Manager: 16k stars. Installs and updates the custom node packs the community releases weekly. Worth naming this week because a pack that breaks is the second most common reason a working graph stops working; a retired hosted model is now the first.
- huggingface/diffusers: 35k stars. The Python path for people who script generations rather than wire them. Relevant now because most of the weights above publish Diffusers-format code before anyone writes a Comfy node.
- huggingface/lerobot: 28k stars. The robot-control library FLUX 3 Action ships configs for. Listed here not for robotics but because it is the harness for the games example below, which is the most creator-adjacent thing in the release.
- black-forest-labs/flux-action: 25 stars. Two days old, which is the whole story of the number. Standalone training, export and inference for the new model, with the separate
transformer.pyandtransformer_inf_bf16.pysplit written, the README says, so that readers and coding agents can follow one path without tracing inheritance.
Creative workflows
1. Transparent PNGs out of Qwen-Image-2.1, locally, with no selection tool. The graph is image_qwen_image_2_1_t2i.json in Comfy's official template repo, and the edit graph beside it is image_qwen_image_2_1_image_edit.json.
The steps. Drop qwen_image_2.1_int8_convrot.safetensors in ComfyUI/models/diffusion_models/, qwen3vl_8b_int8_convrot.safetensors in ComfyUI/models/text_encoders/, and qwen_image_2.1_vae_bf16.safetensors in ComfyUI/models/vae/. Load the template. Then write your prompt in this exact frame, which is quoted verbatim from the note node inside the graph: "This is an RGBA format image with transparency. [your description]. The image has an alpha channel and a transparent background." Save as PNG. For full 2K output, set the ResolutionSelector to 1:1 and 4 megapixels rather than the default 1.
How it works. The transparency is not a post-process and not a mask node. The model's decompressor carries an alpha channel natively, and the sentence you wrap the prompt in is what tells it to use it. Everything downstream is ordinary: one loader per part, a single text encode node, a standard sampler.
Why it is good. A cutout that comes out of the model has no halo to clean up and no edge to feather, and for repetitive asset work, icons, product shots on white, sticker sheets, it removes an entire step you were doing by hand. The top level is five nodes, two of which are documentation notes, because the real work sits collapsed inside one subgraph. Three functional nodes is legible enough to modify without reading for an hour first.
Where it breaks. The template ships at 25 steps while its own note says the official pipeline runs about 40 to 50 with euler, so out of the box you are at roughly half, and thin or wispy edges are where you will see it. Keep cfg at 1 on this path; the note says the negative prompt does nothing while cfg is 1, so a negative prompt you type in will be silently ignored. Save as anything but PNG and the alpha is gone. Prefer dimensions in multiples of 32. And the licence on these files is qwen-research, so whatever you cut out, check the terms before it goes in a paid job.
2. Teaching one model to play two different things from recorded demonstrations. The recipe is examples/games/README.md with the settings documented in docs/embodiments.md.
The steps. Record episodes with each game's own scripted bot, which writes one index.json plus one .npz per episode holding frames at 256 by 256 and actions at 15 Hz. Point TrainConfig.dataset at examples.games.dataset:build. Edit the paths in configs/games/train.json, then run torchrun --nproc_per_node 8 -m flux_action.cli train --config configs/games/train.json. Export with flux-action export-checkpoint --checkpoint <run>/step-3000 --output <run>/export-3000-bf16 --dtype bfloat16. To run it, load the export, call policy.prepare_inference(), and keep num_inference_steps=1 and single_frame_encode=true from the config it shipped with.
How it works. One window is 33 frames and 32 actions: one frame you observed, then the frames your actions led to. The caption is the switch. Write the shooter's name and the model behaves like the shooter; write the racer's and it steers. Where the two games have different numbers of controls, the narrower one is zero-padded and those padded columns are excluded from the loss, so one output head serves both.
Why it is good. This is the clearest published example of caption-as-router that I have seen, and the idea generalises well past games. If you have recorded sequences of a thing happening and the inputs that caused it, you can train one model over several different kinds of that thing and pick between them with a sentence at run time. The repo also ships test_trainer_games.py, which runs the whole plug through the real trainer on synthetic data, so you can verify your own data loader before you spend a night of compute.
Where it breaks. Eight cards in the published command, and the docs warn that the default warmup phases of 1000/2000/1000 fit a 30,000-step run; at 3,000 steps you must switch to 200/600/200, because with the defaults, in the docs' own words, "the trunk never unfreezes" and stays frozen for the whole run. The games themselves are in separate repositories not included here, and the drone data is explicitly absent, so you cannot reproduce the demo end to end without supplying your own recordings. And the ceiling is honest and low: both games were recorded by a scripted bot, so the policy imitates that bot and will not discover anything the bot never did.
Worth testing
- The FLUX 3 Action arm simulator. Free, on ZeroGPU, no install. Tradeoff: it is a simulation of a desk robot, so what you are really evaluating is whether written instructions produce sensible motion plans, not whether this helps your day job.
- Your own Sora export, today. sora.chatgpt.com/sunset. Tradeoff: none, except that the page promises an email when the archive is ready rather than an instant download, so start it before you need it.
- Topaz's standalone apps while they are standalone. Adobe committed in writing to keeping them available as separate products and keeping the brand. Tradeoff: that commitment is a blog sentence, not a licence term, and Neurostream's on-device path is the part worth confirming still works the way you need before you rebuild a workflow around it.
- The Qwen-Image-2.1 transparency frame, on a community Space if you do not want the download. Tradeoff: community Spaces go down without warning and none of them is official, so a bad result may be the Space rather than the model.
- The Basis listening samples. Free, no access request. Tradeoff: listening tells you whether the overlap is as messy as real conversation, which is the only question that matters, but the dataset itself is behind manual review so plan for a wait.
What actually matters from today's signal
Two hosted video models got death dates this week and one of them dies today. That is not a story about Sora being bad, and it is not a story about OpenAI. It is a story about what a pipeline is. ElevenLabs wrote the important sentence for everybody: templates with a Sora node cannot run until you swap the node. If you have spent the last year building repeatable processes on hosted models, and most working creators have because hosted models are better and cheaper than running your own, then you own a set of documents that stop functioning on a schedule you do not control. The generations survive. The recipe does not.
The Topaz acquisition is the same shape from a different angle. Nothing bad happened to Topaz users on Tuesday; Adobe said the right things and said them in public. But an app you bought is now a line item inside a company whose business is subscriptions, and the on-device processing path that made it useful to people working offline is now a strategic choice somebody else gets to make each year. Meanwhile the FLUX and Qwen weights that landed this week are governed by licences with real restrictions, which is a worse deal than "hosted and reliable" in almost every way except the one that came due today.
So the counter-signal, and I think it is the right one to act on: do not migrate everything local, because that is a huge tax for a risk you can price. Do the audit instead. Open your five most-used graphs and your five most-repeated jobs, and write down every model in them you do not have on disk. That list is your exposure. For each one, decide now whether you would rebuild it or abandon it, and write the decision next to the name. It takes an hour, it costs nothing, and today is the day it stops being theoretical.
Source access notes: All Hugging Face dates verified against createdAt from the API with cache-busting, never listing "Updated" timestamps. The basis-conversations-1500 response exceeded fetch limits before reaching createdAt, so its September 15 creation date came from the repo's HF object-id timestamp, a method that matched createdAt exactly on two other repos this run. Rejected on verified dates despite ranking high in "recently updated" feeds: three vpakarinen H3 LoRAs, Smite79/MiniMax-H3-Longvideos, OpenVDN/vdn-minimax-h3, JoaoZaokk/LTX-2.3-22B-distilled-1.1-W4A8-ConvRot, Lightricks/LTX-2.5, stabilityai/stable-audio-3-optimized. Several shared a lastModified to the minute, the signature of a batch re-upload rather than a release wave. Stars from shields.io; repository files read through raw.githubusercontent.com. Blocked or unusable: blog.comfy.org (JavaScript shell), ComfyUI GitHub releases HTML and atom (stale or empty), ComfyUI and diffusers CHANGELOG.md (do not exist), Ideogram news (contact stub), Recraft and PetaPixel (over fetch limits), Runway blog (JavaScript shell), Stability AI news (nothing dated 2026-09), Midjourney updates (unreachable), Google Labs and DeepMind (month-only dates in September, so nothing could be placed in the window and nothing from them is cited), MusicTech (undated listings). The Adobe/Topaz figure of roughly $340M is press-sourced only; Adobe's post states no price. Omitted sections: Writing, nothing in the window affected content creation. Adversarial pass: an independent fact-check agent re-verified every load-bearing claim against primary sources and found five errors, all corrected before saving: the Topaz close was written as Tuesday when September 23 was a Wednesday; FLUX 3 Action was described as three policy repositories when the README names two (DROID and SO-101) plus a separate base repo holding the shared encoders; the Qwen template was described as four visible nodes when the JSON has five at top level, two of them documentation notes; the "permanently delete" quote was attributed to the Sora sunset export page when it appears on OpenAI's help article; and the warmup-phase warning was overstated to "you train nothing" when the docs say only that the trunk never unfreezes. Every direct quote in this briefing was confirmed verbatim, and all star counts, licence names, dates, filenames, commands and the Basis dataset figures were confirmed unchanged. A later scoped pass on the derived articles caught one more error that applied here too and has been corrected in both: the prompt-enhancer and four-bit reader files are listed in the Comfy-Org README's own folder diagram, so only the workflow note and Comfy's documentation page omit them. That same pass also confirmed a third official template, image_qwen_image_2_1_background_removal.json, whose fixed instruction is Remove the background, and output a PNG image.