Creative AI Briefing: Tuesday, September 1, 2026
Six hundred and twenty-five frames of 1080p video, about twenty-six seconds at 24 fps, rendered on a 16 GB graphics card, with the same character still recognizable at the end as at the start. That is one creator's posted result from a node pack that went up over the weekend, and it is the shape of everything worth reading this morning. Three unrelated projects landed in the same 72 hours, and all three are aimed at the same weakness: generated video forgets. It forgets the room when the camera turns around, it forgets the face across a cut, and it forgets the soundtrack it was building. The consequence for anyone making things is that the unit of AI video is sliding from the shot to the sequence, and the skill that pays is continuity management rather than prompt polish.
New models
Matrix-Game 3.5 is the one to look at, though its release history needs unpicking. Riemann Dynamics' code, technical report and project site arrived at the end of August, with the paper dated August 30, but the weights had been sitting on Hugging Face since July 18 for the base models and July 26 for the distilled one, with nothing to run them. The repository carries no dated release tag, so treat August 30 as the report date rather than a confirmed code-drop timestamp. What you get now is a camera-controllable world model: you hand it one anchor image, a camera trajectory, and a text prompt, and it generates 720p video that walks through a scene in first person or third person and remembers the geometry when you come back. The mechanism it calls Patch Memory lifts what it has already generated into a 3D memory and retrieves it based on what should be visible from the new camera position, which is why a wall stays the same wall on the return trip instead of becoming a slightly different wall. Third-person mode takes up to four protagonist reference crops, and with them the character's identity is locked; without them the model invents the protagonist freely.
Both base models are 5B and built on Wan2.2-TI2V-5B, so the download pulls Wan as a dependency along with a Depth-Anything-3 checkpoint for metric depth. Apache 2.0 on the project, with the bundled first-person samples under CC BY 4.0. The catch is hardware and it is not close: one NVIDIA GPU with at least 40 GB of VRAM, Linux, at least 64 GB of system RAM, Python 3.10. The README states 704x1280 generation peaks around 40 GB. There is a second catch worth naming, because the phrase "minute-long real-time" attaches to this project everywhere and belongs only to the distilled three-step causal checkpoint; the base model runs 25 denoising steps at guidance 5.0 and generates in blocks of 80 frames, one block by default. The third catch is authorship of the input: the camera trajectory is a .npz of (N,4,4) camera-to-world matrices plus intrinsics in pixels. There is no way to type a camera move. You need a tool that emits camera paths before you can drive this thing at all (repo, 149 stars, base weights, distilled weights, project site). No hosted demo exists.
Image
Nobody shipped an image model in this window, and the interesting image story is what people are doing with one that shipped in June. Of the sixteen most recently created text-to-image repositories on Hugging Face at 08:17 UTC this morning, eight were LoRA adapters trained on Krea 2, split between krea/Krea-2-Raw and krea/Krea-2-Turbo. A re-check hours later put it at seven of sixteen, so read this as a live ratio hovering around half rather than a fixed number. Krea 2 went up June 18 and is a 12.8 billion parameter model behind a click-through gate and the Krea 2 Community License. That gate has not slowed adapter production at all, which is the actual signal: two and a half months after release, Krea 2 is where a large share of hobbyist fine-tuning attention now sits, ahead of the Flux and Stable Diffusion lineages in that particular sample.
What makes it possible is that the training is hosted. multimodalart/krea2-lora-trainer is a running Gradio Space that captions your images with Gemma 4 31B and then launches the training run itself, and it is worth understanding the arrangement before you click: the Space asks for OAuth scopes including jobs, because the training does not run on the Space. It runs as a Hugging Face Job under your own account. The interface is free; the compute is yours (trainer Space, Krea 2 Raw).
Video
HR Endless Sampler is the weekend's most useful object, and it is one person's ComfyUI node pack with 72 stars. It replaces SamplerCustomAdvanced and renders video of any length by splitting one long latent into chunks that MiniMax H3 can actually hold. The part that separates it from ordinary chunking is that it runs Gemma 4 12B between chunks as a director. Before chunk one, Gemma reads the whole prompt and plans the timing of every shot with full knowledge of where the physical chunk boundaries will fall. Then for each chunk it looks at the previous chunk's frames sampled at 2 FPS plus the exact final frame, and writes a fresh short prompt for that chunk that keeps dialogue intact inside <d>...</d> tags and recalculates cut timecodes onto the local clock. H3 never sees Gemma's notes, only the final description.
The memory trick underneath is that Gemma runs in a separate process. H3, Qwen and the video VAE are unloaded before Gemma starts, and the Gemma process exits before sampling resumes, which is how a 12B language model and a video model share a 16 GB card. The author's own numbers: 625 frames at 1080p with chunk_frames=56 and video_continuation=22, using KJNodes' Low VRAM Attention node set to 4. Without that node, drop to 39 frames per chunk (repo).
Also on the board, at one line because there is nothing to run: DreamX-Creator 1.0 from Alibaba's DreamX team put up a paper on August 31 and initialized its repository on September 1, promising a 7B model that denoises video and audio jointly rather than bolting sound on afterward. The repo's own roadmap still has weights, inference code and evaluation as open checkboxes (repo).
Audio and music
thepatch/acestep-1.5-mlx landed August 31 at 22:07 UTC: an f16 MLX conversion of ACE-Step v1.5 XL base, MIT licensed, about 11.6 GB of storage across the transformer and its conditioning stack. This is the unquantized base rather than the turbo variant, so expect the full step count and a slower wall clock in exchange for headroom. Read the provenance files in the repo before you commit disk to it (model card).
The music news that actually changes something is a rule, not a model. Australia's chart eligibility change took effect with the ARIA Chart dated Monday, August 31, announced August 25. Wholly AI-generated tracks are no longer eligible for the ARIA Charts or the ARIA Awards; recordings that use generative AI in a supporting role stay eligible if they are "substantially human made," judged against the labelling standard the global music community announced on July 10. ARIA can remove a record retrospectively, adjust chart positions, withdraw accreditations and ask for a number-one award back. There is a documented disputes process (ARIA).
Open and local
The local story this week is that the long-video problem got solved on consumer hardware by a person, not a lab, and the solution was to put a language model in charge of continuity. That pattern is going to repeat. Star totals below are cache-busted shields.io reads taken September 1.
- hradec/ComfyUI-HR-Endless-Sampler: arbitrary-length MiniMax H3 video on a 16 GB card, with Gemma 4 planning shot timing between chunks (72) (repo)
- Riemann-Dynamics/Matrix-Game-3.5: walk a camera through a generated 720p scene and have the scene still be there when you turn around (149) (repo)
- comfyanonymous/ComfyUI: the node graph every workflow on this page loads into (131k) (repo)
- kijai/ComfyUI-KJNodes: utility pack whose Low VRAM Attention node is a hard dependency for the 1080p recipe above (3.2k) (repo)
- city96/ComfyUI-GGUF: loads GGUF-quantized diffusion checkpoints, the standard low-VRAM path (4k) (repo)
- Wan-Video/Wan2.2: the Apache 2.0 video family Matrix-Game 3.5 is built on top of (17k) (repo)
- leejet/stable-diffusion.cpp: image and video generation as one C++ binary, no Python environment (6.9k) (repo)
Creative workflows
1. Render a 1080p video longer than your model's ceiling, on 16 GB. Node pack at hradec/ComfyUI-HR-Endless-Sampler. MiniMax H3 only for now; LTX-2.5 is described as planned.
The steps. Install the node pack, then run ~/comfyui/tools/python.sh -m pip install -r requirements.txt, which pulls llama-cpp-python==0.3.35 from the CUDA 12.5 wheel channel plus huggingface-hub>=0.34.0 and av>=16.0.0. Swap SamplerCustomAdvanced for HR Endless Sampler and wire five extra inputs: clip from the model clip, vae from the H3 video VAE, images from your ref2va reference images, prompt from your full prompt text, and fps at 24. Add HR Endless Sampler Preview in the model path ahead of the guider. Set chunk_frames=56 and video_continuation=22, install KJNodes and set MiniMax H3 Low VRAM Attention to 4. First run downloads Gemma 4 12B QAT Q4 GGUF plus a 465 MB Q8 MTP assistant into models/llama_cpp/gemma-4-12b-it-qat-q4_0/. Write prompts in MiniMax shot format: first shot has no timecode, later shots use strictly increasing cut times such as [Shot 2] At 00:02.833, the camera cuts inside the temple.
How it works. Each chunk carries 22 finished frames forward as a synchronized video and audio reference, plus the previous chunk's last five frames as a boundary keyframe. Gemma sees the tail, rewrites the prompt for the next chunk, and hands H3 one description at a time.
Why it is good. You keep the identity references across every chunk, you can hover a chunk in the preview to read the exact prompt Gemma wrote for it, and debug_start_chunk lets you re-render one late shot without resampling everything before it.
Where it breaks. Four things, all stated by the author. Chunk two costs more VRAM than chunk one because it carries the continuation tail, so a setting that survives the first chunk can still die on the second. Chunking only reduces the temporal part of the memory cost, so one full-resolution sampling step must still fit. There is no workflow JSON template yet. And the most consequential one: "Gemma observes generated video frames, not generated audio," so on a long chain nothing is supervising the soundtrack.
2. Drive a camera through a generated scene with a locked character. Repo at Riemann-Dynamics/Matrix-Game-3.5.
The steps. Create a Python 3.10 environment, install PyTorch for your CUDA version, then pip install -r requirements.txt. Download three things into checkpoints/: RiemannDynamics/Matrix-Game-3.5-Base, Wan-AI/Wan2.2-TI2V-5B for the T5 encoder, VAE and DiT scaffold, and depth-anything/DA3NESTED-GIANT-LARGE-1.1 for metric depth. Then run python infer.py --person third --image input.png --camera camera.npz --prompt-file prompt.txt --refs <folder of protagonist crops>. Drop --refs and the model invents the protagonist instead. Outputs land in outputs/third_person/<timestamp>/ as result.mp4 alongside memory_visualization.mp4, a two-row diagnostic panel showing the generation against the mosaic memory.
How it works. Depth-Anything-3 gives metric depth, the frustum engine reprojects past observations into a 3D memory, and camera projection matrices are folded into the positional encoding so the model reasons about view geometry and time together.
Why it is good. memory_visualization.mp4 is the rare debugging artifact that tells you why a generation drifted rather than just showing you that it did.
Where it breaks. 40 GB VRAM floor, Linux only, 64 GB system RAM. Each block is 80 frames and consumes 84 camera poses, and a trajectory shorter than 1 + 84 x num_blocks poses is padded by holding the last pose, which will read as a dead camera at the end of your shot. You still need something upstream that produces .npz camera extrinsics.
Worth testing
- Eddie AI's review round trip shipped August 30 and is the most quietly practical thing in this briefing for anyone who cuts for clients. Send a review link, the reviewer needs no account, and their notes come back as markers on the timeline at the exact frame when you open the export in DaVinci Resolve. Also new: caption export as SRT, VTT or plain text at no cost, and credits purchasable without a subscription. Honest tradeoff: this is a hosted service and your client footage goes to it, which is presumably why SOC 2 Type II and GDPR certification shipped the same week (post).
- Krea 2 LoRA training in a browser. Upload images, get captions from Gemma 4 31B, train. Tradeoff: the run executes as a Hugging Face Job on your account, not on the free Space, so check what that costs you before you queue a long run (Space).
- Matrix-Game 3.5 distilled checkpoint rather than the base, if you have the card. Three steps instead of 25. Tradeoff: first person only, and
infer_distilled.pyrequires an explicit--checkpointand a config file rather than inheriting defaults (weights). - ACE-Step 1.5 XL base on MLX, if you already ran the turbo conversion and want to hear what the undistilled model does with the same prompt. Tradeoff: 11.6 GB and no speed story (model card).
What actually matters from today's signal
The interesting engineering this week was not a bigger model. It was three different answers to the question of what a generator should be allowed to remember. Matrix-Game 3.5 answers it geometrically, holding a 3D scene and retrieving from it by visibility. HR Endless Sampler answers it linguistically, putting a small language model in the loop to keep writing down what just happened so the next chunk can act on it. DreamX-Creator proposes answering it across modalities, denoising sound and picture together so they cannot drift apart. Those are three attacks on the same wall, and the wall is the reason AI video has stayed stuck at clip length while image generation moved on years ago.
For your time and money, the practical read is that the bottleneck moved. It is no longer whether a ten second clip looks good. It is whether shot four still contains the person from shot one, and whether you can find out cheaply when it does not. That is why the two most valuable features in this briefing are unglamorous ones: Matrix-Game's memory_visualization.mp4 and HR Endless Sampler's per-chunk prompt tooltips. Both exist so you can see the failure instead of guessing at it. Tools that let you inspect continuity are worth more right now than tools that produce a slightly prettier frame.
The counter-signal arrived on the same days, and it is not subtle. On August 31 Instagram renamed its AI creator label to AI-generated profile and started limiting the reach of unlabeled profiles that feature an AI-generated person, with a documented path back through Account Status. On the same date ARIA's rule took effect and wholly AI-generated tracks stopped being eligible for the Australian charts and awards. Read the Instagram policy carefully, because it draws the line in a specific place: using AI tools in your process does not trigger the label, presenting an AI-generated person as a human does. Both bodies landed in the same spot, which is that AI in the making is fine and AI as the identity is not. The week you gained the ability to make longer, more coherent AI video is the week two large distribution channels started asking who is actually in it.
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 is a JavaScript wall and returned nothing. github.com/trending served a clearly stale page and was discarded. The GitHub REST API and the Hugging Face daily_papers API both returned empty bodies from this environment; dates were established from raw.githubusercontent.com READMEs with cache-busting query strings, Hugging Face createdAt fields, and primary vendor pages. runwayml.com/blog returned a JavaScript shell; the Runway changelog rendered but had nothing in window. midjourney.com/updates was not directly reachable and nothing was included from it. stability.ai/news returned what appeared to be a stale snapshot and was not used. petapixel.com category pages returned empty bodies. Two dates supplied by crawl agents were wrong and were corrected against primary sources before writing: the SoundCloud direct-purchases post is dated August 26, not August 29, and was dropped as out of window; the ARIA announcement is dated August 25, with August 31 being the chart date the rule takes effect from. Matrix-Game 3.5's weights predate its code release by roughly six weeks, verified through the Hugging Face createdAt fields, and the briefing states both dates rather than the repo's publication date alone. Sections omitted for having nothing worth saying: Writing.
An adversarial fact-check pass ran against this draft before publication and caught four things, all now corrected in the text above. The opening sentence originally attributed dialogue and room tone to the specific 625-frame demo clip; the README caption claims only 1080p, 625 frames, 16 GB, so the audio capability was being imported from MiniMax H3 generally and has been removed from that sentence. The August 30 date for Matrix-Game 3.5 could not be pinned to a dated primary page because the repository has no release tag and the GitHub REST API is unreachable from here, so it is now stated as the report date with that limitation named. The Krea 2 LoRA ratio moved from eight of sixteen to seven of sixteen between the first read and the re-check, and is now presented as a live ratio. The checker's re-fetch of the Instagram creators post returned an empty body on two attempts; the page rendered in full on the original fetch, dated August 31, 2026, and multiple independent outlets carry the same details on the same date, so the citation stands.