Creative AI Briefing: Friday, September 4, 2026
Drop one still image into ComfyUI, type a short action timeline using W, A, S and D, and get back five seconds of video in which the character actually walks forward and the camera actually pans left, with a synchronized soundtrack. The workflow file that does this is dated yesterday. It is the clearest example of the week's real story: the prompt box stopped being the only steering wheel. Three separate releases now let you drive a generated shot with keys or camera strings instead of adjectives. The catch is that almost none of this is a new model. Each one is a thin control layer strapped to a generator that shipped weeks ago, which means you inherit the base model's license, and in the two most interesting cases that license does not cover you.
New models
H3-World is the release worth your morning. It is a LoRA, not a model, and that is the point. The weights landed on Hugging Face August 31 and the paper went up September 1, and together they turn MiniMax H3, a 33-billion-parameter video generator, into something you steer with keystrokes. Each action becomes a structured pairing of a character instruction and a camera instruction, aligned to specific temporal video latents. A technique the authors call temporal attention routing then confines each instruction to its own time interval, which is what stops "walk forward, then turn left" from smearing into a single confused motion. The paper's argument is that MiniMax H3 could already respond to natural-language descriptions of character behavior and camera motion; H3-World just makes that response temporally precise instead of vague.
The efficiency numbers explain why this arrived so fast. 8,000 gameplay samples, 10,000 LoRA optimization steps, and 0.199% of the base model's parameters trainable. The shipped file is 131,227,832 bytes, rank 32, 104 A/B pairs. That is a 131 MB download that adds a control surface to a 33B model.
Two catches, both large. First, the word "interactive" in the paper is not the word you want in your head. The first release is a fixed contract, not a game: one first-frame image, 832x480, 124 frames at 24 fps, which is about 5.17 seconds. You author an action timeline across 37 action latents, queue it, and wait. Nothing responds to you while it renders. Second, the LoRA carries an Apache 2.0 license and the base weights do not. MiniMax H3 ships under the MiniMax H3 Community License Agreement, whose applicable territory excludes the European Union, the United Kingdom, South Korea and the United States. The node pack's own README states this plainly in its license section, which is more than most ports bother to do (paper, weights, ComfyUI package, hosted demo).
Image
Nothing shipped in image generation this window that qualifies as a model release. Both notable items are the same small idea arriving twice: more reference images. Midjourney's alpha changelog dated September 2 adds a lightbox editor where you can attach up to four reference images to a plain-language edit, see every edit from a session in one place, and try an experimental "Change Style" tool. It is alpha only, at alpha.midjourney.com, and Midjourney describes it as in progress (changelog).
Six days earlier, xAI made the same move. Its release notes for August 28 raise Grok Imagine Image 2.0's reference-image ceiling from three to five, add 21:9 and 5:2 cinematic aspect ratios, and move the default quality value from medium to auto. That is outside this window, and it only surfaced here because a September 2 migration notice restates it while redirecting an older model slug, so read it as context rather than news (release notes).
Two vendors independently deciding that the useful knob is "how many pictures can I hand you" says something about where control is actually moving.
Video
Echo-WM Flash and Base got Diffusers ports on September 3, and this is the second keyboard-control story of the week. Echo-WM is a four-step causal audiovisual world model from the JoyAI-Echo project: give it a starting image, a text prompt, and a camera-action sequence, and it generates video and synchronized audio together, keeping memory bounded with a sink-plus-FIFO KV cache instead of an ever-growing context.
The camera syntax is the interesting part, and it is legible in a way most video controls are not. Each segment is <keys>-<duration>. W and S move forward and back, A and D strafe, I and K pitch, J and L yaw, and none holds still. String them with commas: w-60,a-60,w-60,d-60. Flash uses four distilled denoising steps, requires num_frames = 1 + 24 * n, and defaults to 241 frames from 31 latent video frames. Because it is guidance-distilled it refuses a negative_prompt entirely.
The base checkpoint shipped August 23. What landed this week is the conversion that lets you load it through the Diffusers Modular Pipeline API rather than the project's own inference code, which is genuinely the difference between reading about a model and running one. The license is the whole story though: LTX-2 Community License, and the card says it directly. Academic research and non-commercial use only. No hosted demo exists (Flash weights, Base weights, paper).
The other video item is a straight compression job with a real consequence. LTX-2.5 GGUF quantizations for ComfyUI appeared September 1, covering Q2_K through Q8_0 of Lightricks' 22B video-and-audio model. The original bf16 transformer is 39 GB. Q4_K_M is roughly 11.5 GB, which the card pairs with a 16 to 20 GB card, and it is explicit that you need another 4 to 8 GB on top for the execution context, VAE and text encoders. Read the card's own warning before you download the small ones: Q2 and Q3 "produce poor quality results and are generally unusable for final renders." Licensing is the friendliest of anything in this briefing. Under $10 million in annual revenue, the LTX-2.x Community License permits commercial and production use at no cost (weights, license).
Audio and music
Thin week, and one item worth knowing about precisely because it is not new. On September 1 someone published an inference mirror of MOSS-SoundEffect v2.0, the text-to-sound-effects model from OpenMOSS at Fudan University. The model itself shipped May 25. It generates 48 kHz sound effects up to 30 seconds from English or Chinese prompts, covering environments, urban scenes, creatures, human actions and short percussive clips, using a 1.3B Diffusion Transformer trained with flow matching, a DAC VAE, and a Qwen3-1.7B text encoder. Code and weights are both Apache 2.0, which makes it one of the very few genuinely unencumbered generators anywhere in today's briefing.
The mirror did two small useful things: it converted the upstream VAE from a pickled .pth file to safetensors, and it stores the transformer in the bf16 inference dtype rather than the fp32 upstream ships. That is second-day work, done so a DAW could fetch the checkpoint on demand, and it is a fair reminder that "shipped in May" and "usable in your tool" are different dates. There is a free hosted Space if you want to hear it before you download 11 GB (weights, hosted demo, mirror).
Open and local
The local story is that provenance became a node. Provcheck v1.4.0, per the project's own release history, shipped August 31 and included a free Apache 2.0 ComfyUI node; the ComfyUI Wiki wrote it up September 3, which is how most people will have seen it. One node, called "Provcheck (Free C2PA Watermark)," takes an image, audio or video socket and either writes a watermark plus a C2PA signature or reads an existing mark. TrustMark-B marks images, silentcipher marks the audio track. It runs offline, no account, and if the CLI binaries are missing it passes your media through untouched rather than crashing the render.
Star totals below are cache-busted shields.io reads taken September 4.
- T8mars/comfyui-minimax-h3-audio-T8: the 292-node ComfyUI pack that makes H3-World's WASD control actually runnable, plus long video, lip sync and 8-step acceleration routes (934) (repo)
- CreativeMayhemLtd/provcheck: watermark and C2PA-sign your generated media inside the graph, offline, tied to your Bluesky handle if you want it (11) (repo)
- Danzer1xxxxChan/H3-World: the upstream code for the action-control LoRA, if you want to train your own action set rather than use the shipped one (repo)
- jd-opensource/JoyAI-Echo: source and example camera trajectories for Echo-WM, including the
wm_causal_casesinputs the Diffusers card uses (repo) - OpenMOSS/MOSS-TTS: Apache 2.0 code for the sound-effects model, including the Gradio demo and fine-tuning recipes (repo)
- Lightricks/LTX-2: the license text that decides whether your LTX-2.5 renders are commercially usable, which for most readers here they are (repo)
Creative workflows
1. Steer a character and a camera through five seconds of generated video with WASD, in ComfyUI.
Node pack at T8mars/comfyui-minimax-h3-audio-T8. LoRA at t8star/Minimax-H3-World-Comfy. Workflow file: examples/workflows/26-h3-world/2026-09-04_H3_World_I2VA_832x480_124f_50step_Advanced.json.
The steps. Install the node pack through ComfyUI Manager by searching MiniMax H3 Audio T8, then fully quit and relaunch ComfyUI rather than just restarting the server. Download the LoRA with hf download t8star/Minimax-H3-World-Comfy --include "loras/**" --local-dir ComfyUI/models, which must land at ComfyUI/models/loras/minimax/H3-World/step-10000.safetensors. You also need the full MiniMax H3 base at models/diffusion_models/minimax_h3_fl2va_int8_convrot.safetensors, the Qwen3-VL text encoder, and both the video and audio VAEs. Make sure ffmpeg is on your PATH, because the final safe-save step needs it. Open the workflow above, replace the first frame and the models, and leave the generation contract alone: 832x480, 124 frames, 24 fps, 50 Euler steps, CFG 1.0. Pick a preset from forward, back, strafe left or right, tilt up or down, pan left or right, and fast pan, or choose custom and write your action segments as JSON.
How it works. The LoRA encodes each action as a paired character-and-camera instruction and binds it to specific positions in the 37-slot action latent timeline. Temporal attention routing then restricts each instruction to its own interval, so an action you scheduled for the second half does not leak into the first.
Why it is good. The file is 131 MB and byte-identical to the researchers' own upload, verified by a published SHA-256, so you are not trusting a community requantization of anything. It adds no new pip dependencies. And a blind test on a fixed parking-lot sample had users correctly picking out the continuously-forward version, which is a modest claim honestly stated rather than a benchmark chart.
Where it breaks. Four places. The contract is fixed, so 832x480 and 124 frames are not suggestions; a first frame at another aspect ratio gets proportionally covered and center-cropped, which cuts the edges off your subject without warning you. You cannot stack this with OpenVDN, SLA, VSA, Sol-Attn or BlockCache, because each of those takes over the model or attention path. H3-World requires the full non-pruned base, and the project's serial tests of that base on a 16 GB card left only 535 to 890 MB of VRAM free, so run one job at a time and drop resolution before you drop steps. And the base weights are licensed for a territory that excludes the US, UK, EU and South Korea.
2. Watermark and C2PA-sign a render without leaving the graph, offline. Node at CreativeMayhemLtd/ComfyUI-Provcheck, CLI at CreativeMayhemLtd/provcheck.
The steps. Clone the node into ComfyUI/custom_nodes or install "Provcheck" from ComfyUI-Manager. Download the provcheck and provcheck-kit binaries and put them on your PATH; check with provcheck-kit --version. For watermark only, wire LoadImage into the node with mode: write, watermark: on, c2pa_sign: off, and take the image output into a normal SaveImage. To sign as well, run the one-time provcheck-kit init, provcheck-kit login -u <handle>, provcheck-kit publish, then set c2pa_sign and embed_identity on. Note that the node's README writes that middle command with a bare handle and the CLI's own repo writes it with the -u flag, so check --help if one form is refused. Turn on record_provenance if you want the model, seed, sampler and prompts baked into the manifest.
How it works. The node is a thin Python wrapper. Each modality is serialized to a file, handed to the CLI's two-step stamp-then-sign path, and reloaded. The signature is bound to a key you publish to your atproto identity, so a verifier downstream can confirm the file really came from your handle.
Why it is good. It never leaves your machine, it needs no account, and it fails closed: a missing binary or a missing identity passes your media through with a console note rather than killing a long render. include_full_workflow is off by default so prompts and paths do not leak, and even when you turn it on, secrets are dropped and paths shortened to basenames.
Where it breaks. The image path needs ONNX Runtime 1.22.x, and if the runtime is missing or mismatched the image mark silently does nothing, which is the worst possible failure mode for a provenance tool. When signing, the artifact is the file written to output/, not the tensor that continues down your graph. Video marks ride the audio track and re-encode it to AAC, so a silent clip carries nothing. And the honest ceiling, stated by the maintainers: this is not a deepfake detector. It tells you whether a file carries a valid mark, never whether unmarked media was generated.
3. Generate 48 kHz foley from a text prompt, free, with a license you can actually use. Model at OpenMOSS-Team/MOSS-SoundEffect-v2.0. Try it first at the hosted Space.
The steps. For local use, make a clean Python 3.12 environment, clone OpenMOSS/MOSS-TTS, and install from moss_soundeffect_v2 with the CUDA extras. Load MossSoundEffectPipeline.from_pretrained("OpenMOSS-Team/MOSS-SoundEffect-v2.0", torch_dtype=torch.bfloat16, device="cuda"), call it with a prompt and a seconds value up to 30, and save with pipe.save_audio. The card's recommended settings are 100 solver steps, cfg_scale 4.0 and sigma_shift 5.0.
How it works. A 1.3B Diffusion Transformer trained with a flow-matching objective produces continuous latents, which a DAC VAE decodes to a 48 kHz mono waveform. A Qwen3 text encoder reads the prompt, in English or Chinese, with the duration prepended as a tag the way it was during training.
Why it is good. Apache 2.0 on both the code and the weights, which is rare enough in generative audio to be the headline. 48 kHz output is delivery quality, not preview quality. And the hosted Space means you can audition it before committing to an 11 GB download.
Where it breaks. 100 solver steps is a lot for iteration, so expect to wait between takes. The pipeline wraps the DiT in torch.compile plus a Triton CUDA graph, which means the first call can take several minutes to compile; set TORCHDYNAMO_DISABLE=1 if you hit Dynamo or Triton errors. And the upstream repo still ships its VAE as a pickled .pth, which is why a mirror had to exist at all.
Worth testing
- H3-World action demo on Hugging Face, if you want to see keystroke control before installing 33B of anything. Tradeoff: a hosted Space cannot show you the 16 GB VRAM squeeze or the territory restriction that decides whether you should be running this at all.
- MOSS-SoundEffect v2.0 for free foley in the browser. Tradeoff: it is a May model, and the 30-second ceiling and 100-step default make it a sound-effects tool, not a scoring tool.
- LTX-2.5 at Q4_K_M if you have a 16 GB card and have wanted video with native audio locally. Tradeoff: budget another 4 to 8 GB beyond the 11.5 GB of weights, and skip Q2 and Q3 entirely for finals per the card's own advice.
- The Provcheck node in read mode, pointed at files people send you. Tradeoff: it answers "is this marked" and nothing else, and the image detector goes quiet without the right ONNX runtime.
- Echo-WM Flash's camera strings, if you can work non-commercially. Tradeoff: no hosted demo, no negative prompt, and a license that stops at academic and non-commercial use.
What actually matters from today's signal
Control is the product now, and it is being sold separately from the model. H3-World is 0.199% of MiniMax H3's parameters and a 131 MB file. Echo-WM's camera language is a comma-separated string. Midjourney and xAI both spent their week raising a reference-image count. None of that is a new generator, and all of it changes what you can actually direct. If you have been waiting for the next big model, you have been watching the wrong layer.
The bill comes due on licensing, and it comes due in a specific way that is easy to miss. Both of this week's best control layers are permissive on their own terms and useless on their own. H3-World's LoRA is Apache 2.0 and needs base weights whose applicable territory excludes the US, the UK, the EU and South Korea. Echo-WM's Diffusers port is a convenience wrapper around a checkpoint restricted to academic and non-commercial use. The badge you see on the repo you downloaded describes maybe one percent of the bytes you are about to run. Read the base model's license, not the adapter's.
The counter-signal is that the two items here with clean, usable terms are the quiet ones. LTX-2.5's community license gives you free commercial use under $10 million in revenue and its GGUF quants now fit a 16 GB card. MOSS-SoundEffect is Apache 2.0 top to bottom and has been sitting there since May. If you want to ship work rather than read about it, that is where your week is.
Source access notes: Crawled September 4, 2026, covering roughly August 30 to September 4. Every Hugging Face ship date in this briefing comes from the API createdAt field with cache busting, never a listing "Updated" timestamp; the H3-World LoRA is August 31, its ComfyUI package September 4, the Echo-WM Diffusers ports September 3, the LTX-2.5 GGUF September 1, the MOSS mirror September 1, and MOSS-SoundEffect v2.0 itself May 25. Star counts are cache-busted shields.io reads. GitHub file contents were read via raw.githubusercontent.com with cache busting after this project was previously burned by stale blob pages.
Blocked or unusable this run: seed.bytedance.com (JavaScript wall, no body content, so no Kling or Seedance coverage), midjourney.com/updates (JavaScript shell only; worked around via updates.midjourney.com, and the Grok Imagine item is sourced to migration docs rather than an announcement post), runwayml.com/news (JavaScript-rendered index returns nothing; the changelog's most recent entry is August 24), blog.comfy.org and docs.comfy.org/changelog (responses exceeded fetch limits), fal.ai/models (same), and several api.github.com release endpoints (rate limited, so the Provcheck v1.4.0 date is sourced to the project's own release-history table rather than an API timestamp). Replicate's blog last posted August 4 and fal's August 27. The Writing section is omitted because nothing surfaced. Hugging Face's own blog published nothing about generative image, video, audio, music or 3D in the window.
An adversarial fact-check pass ran against this draft before publication and caught two errors, both now fixed. It found that the Grok Imagine reference-image increase, the 21:9 and 5:2 ratios and the auto quality default all shipped August 28 per xAI's own release notes, not on the September 2 date of the migration notice originally cited, so that item was rewritten as context rather than news. It also found that the 290 MB free-VRAM figure first attached to the H3-World workflow actually comes from the node pack's OpenVDN tests on the pruned base model, while H3-World requires the full non-pruned base whose tests reported 535 to 890 MB free; the figure was corrected. Everything else it checked held, including the arXiv abstract's training numbers, the LoRA byte count and rank, the workflow filename and generation contract, Echo-WM's frame math, all seven license claims, both verbatim quotes, and both star counts.
Two framing corrections were made during research, before the adversarial pass: the September 1 MOSS upload is an inference mirror of a May 25 model rather than a new release, and Provcheck's ComfyUI node shipped with v1.4.0 on August 31 rather than on the September 3 date of its wiki write-up. Note for future runs: Provcheck's own changelog announces a ComfyUI node twice, once at v0.9.0 on June 29 and again at v1.4.0, so the "new node" framing in the v1.4.0 entry should be treated with care.