Creative AI Briefing: Sunday, August 30, 2026

Five seconds of video with a synchronized audio track, generated on a MacBook with 24GB of memory and no graphics card from NVIDIA in the building. Also a full image model that fits in a gigabyte and renders on an iPhone's neural chip, and a globally illuminated 3D render that finishes in four seconds where Blender takes fifty-four. None of the underlying models are new. What moved this week is the size of the machine you need to run them, and almost all of that movement points at Apple Silicon.

No lab shipped a creative model in the last twenty-four hours. Both of the genuinely new artifacts came from individuals with Hugging Face accounts, and both are conversions rather than new capability. That is the honest shape of the day, and it is also the interesting part, because the conversion layer is where the hardware barrier actually falls.

New models

Core ML AbsoluteReality, 6-bit landed August 29 from a single maintainer, a conversion of the AbsoluteReality photorealism checkpoint aimed at Apple's Neural Engine so it runs on an iPhone or a Mac without touching the GPU. The whole thing is 1.01 GB: a 618 MB UNet at 6-bit palettization, a 235 MB text encoder, two small VAE pieces. Palettization sorts the weights into a handful of shared values instead of storing every number separately, which is how a multi-gigabyte model becomes a phone download. Fixed at 512 by 512, no other resolution. The card is candid about where compression stops: only the UNet is palettized, because the text encoder holds CLIP's negative-infinity causal mask and the clustering algorithm cannot handle it. License is CreativeML OpenRAIL-M inherited from the source checkpoint, with use restrictions that travel with the weights. Zero downloads and zero likes this morning, so nobody has stress-tested it but the author (model card, toolchain).

Anima 2.9B arrived August 29, an anime and illustration model built by taking the Circlestone Labs Anima base and growing it from 28 transformer layers to 40, then training on an additional 1.7 million samples. That is layer expansion rather than a fine-tune: new capacity gets inserted into the middle of the network and trained while the original layers hold still. Full weights are 5.84 GB, with an int8 build at 3.08 GB. Start at 28 to 50 steps, CFG between 3.5 and 5, and 812 by 1216 or 1152 by 1536; the card calls 1536 square "iffy" and does not pretend otherwise. Two catches. The license is Circlestone Labs Non-Commercial, so personal or spec work only. And it needs ComfyUI 0.33.1 plus a dedicated custom node, with the card warning it "sometimes may not work with other custom nodes." Short prompts produce bland backgrounds, so write long (model card).

Image

Nothing new shipped from a vendor in image this week, so the two items worth your attention are both above: the 1 GB Core ML build and Anima 2.9B. If you want an image model you can try in a browser this morning without either, the SenseNova U1.5 8B unified generate-and-edit demo is live and free (Space).

Video

The MLX layer around LTX-2.5 consolidated this week and is now the most practical route to local video on a Mac. MLX is Apple's own array framework, built around the unified memory that Apple Silicon shares between the processor and the graphics chip, which is why a 64GB Mac holds a model that would need a datacenter card on the PC side.

SceneWorks/ltx-2.5-mlx went up August 27 and solves the annoying part: a pinned, pre-converted bundle of the whole LTX-2.5 pipeline at q4, q8 and bf16, including the distilled transformer, the rank-450 adapter and a staged Gemma 4 prompt rewriter. The card's phrasing is the selling point, that a fresh install "downloads these ready-to-run files; no in-app conversion or access to the gated source repositories is required." No hour of quantizing before you generate anything. Tiers are all-or-nothing: q8 applies to every quantizable stage and mismatched layouts get rejected. LTX-2.x Community License dated August 11, 2026, which carries commercial conditions and must travel with anything you redistribute (bundle).

FrankDase/LTX-2.5-MLX (August 20) is the leaner alternative, an 8-bit conversion of the official distilled checkpoint at roughly 43.6 GB, tested by its author on an M4 with 64 GB unified memory. Frame counts follow 8n+1, and the example run is 320 by 448 at 97 frames, 24fps (model card). One warning from the same neighborhood: exeterminal/Exe-Motion-Dynamic-V1-MLX describes an assembled 55 GB LTX-2.3 build, but its repository contains only a README and five images totaling 119 KB. The weights are not there.

Audio and music

SeaOtter 0.5 Beta appeared August 28 and it generates the wrong thing on purpose, which is what makes it interesting. Every music model in this briefing for the past month has produced audio. SeaOtter produces MIDI: a 120 MB LoRA adapter over a 3B instruction model, trained on MidiCaps, emitting REMI tokens that convert straight to a .mid file. That file opens in your DAW as editable notes, so you can change a chord, move the melody up a third, swap the instrument, quantize the timing. None of that is possible with a rendered waveform. Apache 2.0, so the output is yours to sell.

The card is honest about how far one training epoch gets you: "After 1 epoch, the model mainly writes single note lines and rolling arpeggio patterns. It does not reliably play 3 or 4 note stacked chords at the same position yet." Treat it as a melody and figure generator, not a composer. Two repository problems will stop you cold if you clone it blind: the inference snippet points at a different repo id than the one hosting the card, and the required miditok_config.json is missing from the file tree. Prompts need a BPM, a key and a mood (model card).

Open and local

Three separate stacks arrived at the same place this week, and no single announcement made that obvious.

The community route is rapid-mlx, which serves nine MLX-converted video checkpoints behind an OpenAI-compatible API on a Mac: LTX-2.5 at q8, LTX-2.3 at q4, four Wan 2.2 variants and three CogVideoX-Fun 5B quants. The floor is 24 GB of unified memory. The number that matters, quoting the docs, is that LTX-2.5 q8 "is a 67.7 GB download but also serves at 24 GB via its low-RAM distilled path," and it is "the only checkpoint that returns synchronized audio." Wan and CogVideoX-Fun are video only. The counterweight is speed: on an M3 Ultra, CogVideoX-Fun q4 took about 338 seconds to produce one second of 672 by 384 video. Overnight rendering, not iteration (docs).

The framework route is FastVideo, which as of its own dated changelog entry on August 19 runs FastMetal-QAD through an MLX runtime with uv pip install -e '.[mlx]', in 1.3B, 5B and 14B tiers.

The rendering route is the surprise. RenderFormer replaces a path tracer with a transformer: hand it a triangle mesh with materials and lights and it produces a globally illuminated image directly, with reflections, soft shadows and indirect bounce, no per-scene training and no ray tracing anywhere in the model. The released V1.1 Large is 483M parameters under MIT, and the repository states it is tested on Apple Metal GPUs with an 8 GB minimum. Bounded: up to 8 light sources, roughly 4,096 triangles with 8,192 usually still working (repo, model).

Its successor surfaced this morning and you cannot have it yet. RenderFormer-V2 (ECCV 2026, Stanford and Microsoft Research) handles 128K primitives at 2048 by 2048 with refraction, volumetric scattering and environment lighting, and reports 3.92 seconds per 2048px frame against Blender Cycles at 53.8 seconds, all timings the authors' own on an A100. No V2 weights and no V2 code are published; the repository still serves V1. Watch it, do not plan around it (project page).

Star totals verified via cache-busted shields.io:

• ml-explore/mlx: Apple's array framework, the thing every Mac model on this page runs on (28k stars) (repo)
• apple/ml-stable-diffusion: the conversion toolchain that puts an image model on an iPhone (18k) (repo)
• comfyanonymous/ComfyUI: the node graph, and it does run on M1 through M4 (131k) (repo)
• Lightricks/LTX-2: the audio-and-video model the MLX ports are ports of, though its own repo mentions Apple Silicon nowhere (9.3k) (repo)
• hao-ai-lab/FastVideo: 50-step video models cut to 4 steps, now with an MLX path (4.2k) (repo)
• city96/ComfyUI-GGUF: loads compressed weights, with a documented MacOS Sequoia torch version caveat (4k) (repo)
• microsoft/renderformer: neural global illumination, MIT, runs on Metal at 8GB (971) (repo)
• james-see/ltx-video-mac: a native macOS app wrapper around LTX on Apple Silicon (392) (repo)
• dgrauet/ltx-2-mlx: pure MLX port of LTX-2, MIT, int4 path down to 16GB (106) (repo)

Creative workflows

1. Generate video with synchronized audio on a Mac, no NVIDIA card. Go to the rapid-mlx video docs and pick ltx-2.5-mlx-q8, which the docs call "the recommended starting point" and the only one of the nine checkpoints that returns audio.

The steps. Python 3.11 or newer, ffmpeg, 24 GB of unified memory minimum. The q8 checkpoint is a 67.7 GB download, so start it before bed. Serve it through the OpenAI-compatible Videos API. Frame counts must follow 8n+1 with a floor of 9, so 9, 97 or 121, not 100. Do not pass a guidance scale or a negative prompt: the docs say LTX-2.5 "takes neither." For Wan and CogVideoX-Fun the frame rule is 4n+1 with a floor of 5, at 50 diffusion steps by default.

How it works. The 67.7 GB figure is the download, not the memory ceiling. LTX-2.5 has a distilled path, a student network trained to reach a finished frame in a few passes instead of fifty, plus a low-RAM mode that streams weights through memory in stages rather than holding the whole model resident. Those two together are why a 24 GB machine serves a checkpoint six times its own memory. Audio comes out of the same forward pass as the picture, not a second model bolted on, which is why the sound lands on the action.

Why it is good. The only local path that gives you picture and sound together on consumer hardware, on a laptop you can close and carry. No rental, no queue, no uploading client footage to somebody's endpoint.

Where it breaks. Speed. The docs record CogVideoX-Fun q4 at roughly 338 seconds for one second of 672 by 384 video on an M3 Ultra. Even granting LTX-2.5's distilled path is faster, plan in minutes per clip. The 14B Wan checkpoints want 64 to 96 GB and are out of reach on most machines. And the LTX-2.x Community License carries commercial conditions, so read it before billing anyone.

2. Put a real image model on your iPhone. Start from apple/ml-stable-diffusion and grab a finished conversion at mahmudplx/coreml-absolutereality-6bit rather than converting anything yourself.

The steps. Four compiled Core ML files: Unet.mlmodelc at 618 MB, TextEncoder.mlmodelc at 235 MB, VAEDecoder.mlmodelc at 95 MB, VAEEncoder.mlmodelc at 65 MB, 1.01 GB total. Set attention to SPLIT_EINSUM_V2, pass --latent-h 64 --latent-w 64 for 512 by 512, and point compute units at cpuAndNeuralEngine, not the GPU. Apple's requirements table puts the floor at iOS 16.2 and macOS 13.1, with the memory improvements landing at iOS 17.0 and macOS 14.0, on A14 or M1 and up.

How it works. 6-bit palettization builds a lookup table of shared weight values and stores an index per parameter instead of a full number, shrinking the UNet roughly four times while the model still does arithmetic at its original precision. SPLIT_EINSUM_V2 reshapes the attention operation into the form the Neural Engine's fixed-function hardware executes fastest, which is why the compute-unit setting matters more here than on a GPU.

Why it is good. A gigabyte and a phone. No server, no credits, no network, nothing leaving the device.

Where it breaks. 512 by 512 and nothing else. The text encoder stays at fp16 because CLIP's causal mask contains negative infinity values a k-means palettizer cannot cluster, so a third of the bundle refuses to compress. Apple's docs note the runtime can spike over 2 GB of peak memory, which on an older phone means the process gets killed. And CreativeML OpenRAIL-M restrictions ride along with the weights.

3. Render a globally illuminated 3D scene on a Mac with 8GB. Clone microsoft/renderformer, install microsoft/renderformer-v1.1-swin-large (483M parameters), and set scenes up with the Blender extension.

The steps. PyTorch 2.0 or newer; Flash Attention is optional and the script falls back to torch scaled dot product attention on a Mac. Convert a scene with python3 scene_processor/convert_scene.py examples/cbox.json --output_h5_path tmp/cbox/cbox.h5, then render with python3 infer.py --h5_file tmp/cbox/cbox.h5 --output_dir output/cbox/. Start from examples/init-template.json. You get an HDR .exr and an LDR .png; pass --tone_mapper agx or filmic for something that looks finished. batch_infer.py --save_video turns a folder of frames into a clip.

How it works. No ray tracing in the model at all. Each triangle with its material becomes a token, camera rays become tokens, and a transformer converts the input sequence into pixel patches. A view-independent stage resolves how light bounces between surfaces; a view-dependent stage turns ray bundles into pixels. Because transport is learned rather than sampled, there is no noise to denoise and no sample count to tune, which is why it finishes in a fixed short time instead of converging slowly.

Why it is good. Preview-quality global illumination in under a second per frame on hardware that cannot run a serious path trace, MIT licensed so the output is unencumbered.

Where it breaks. Hard limits, stated plainly in the repo. Up to 8 light sources. Training covered 4,096 triangles, with 8,192 "usually still works" at inference, so detailed meshes need remeshing first through the included scene_processor/remesh.py. Camera distance must sit between 1.5 and 2.0 from scene center at a 30 to 60 degree field of view, bounding box inside plus or minus 0.5, emission summing between 2500 and 5000. Outside that envelope quality falls off. Specular is white only. And the authors note the released weights differ from the paper's results, because a shader bug in the training data was found and fixed after submission.

Worth testing

• MiniMax H3 Turbo LoRA (free Space): video with a synchronized soundtrack, in a browser, no install. Tradeoff: the four-step distilled path, so fine motion and hands degrade in busy scenes.
• SenseNova U1.5 8B MoT (free Space): one model for generating and editing images, saving a round trip through a second tool. Tradeoff: unified models trail dedicated editors on precise local edits.
• MiniMax Music 3 Studio (free Space): full tracks from a caption plus lyrics. Tradeoff: you get a finished waveform, so to change one note, use SeaOtter above instead.
• RenderFormer V1.1 (model): the only item here that renders your own geometry. Not for you if your meshes are dense; the 4K triangle envelope is real.
• Anima 2.9B (model): the day's most capable new image weights. Not for you if you bill clients, or if your ComfyUI install is delicate.

What actually matters from today's signal

Spend your next hardware decision on memory, not on a graphics card. Every meaningful item on this page is a story about fitting inside a smaller box: a 67.7 GB video model serving from 24 GB of unified memory, a diffusion pipeline squeezed to 1.01 GB for a phone chip, a renderer that skips ray tracing entirely and clears 8 GB. The model side has been flat for days while the memory side moved constantly, and unified memory is the specific resource doing the work, because one pool serves both compute stages instead of forcing a copy across a bus. A 64 GB Mac now does work that a year ago meant renting time.

The counter-signal is that none of this is fast. 338 seconds for one second of small video on an M3 Ultra is the number to keep in your head, and it comes from the docs of the friendliest tool in the group. Local Apple Silicon generation is a batch process you start and walk away from, closer to a render queue than to the three-second hosted generations fal is selling. If your work depends on trying twenty variations of a shot in an hour, rent. If it depends on privacy, on not paying per generation, or on running offline, the local path just became real, and it will be slow.

The quieter risk is licensing drift. Three of today's items restrict what you can do with the output: LTX-2.x Community, Circlestone Non-Commercial, CreativeML OpenRAIL-M. Downloading weights and being allowed to sell what they make are now fully separate questions, and the conversion layer doing all this useful work is also where license text most often goes missing between the source repo and the copy you install. Check the original.

---

Source access notes: Vendor primaries were checked and were quiet: OpenAI News, Adobe Blog, fal, Replicate and ElevenLabs published nothing creative-relevant in the last 24 hours, most recent items dated August 27 or earlier and already covered. blog.comfy.org served a JavaScript subscribe wall. The diffusers releases page exceeded the fetch size limit. Reddit is on the fetch blocklist. Midjourney, Civitai, Kling and Seedance had no fetchable primary announcements this run. Hacker News Algolia returned results dated today and was used for discovery only. Every Hugging Face date here comes from the API createdAt field, not from listing-page "Updated" timestamps; nine of the twenty-six repos checked had recent modification times sitting over createdAt dates weeks or months old, and were excluded or correctly redated. GitHub HTML star counts were stale on four of ten repos checked, in one case by 25 percent, so all star totals come from cache-busted shields.io JSON. RenderFormer-V2 performance figures are the authors' own measurements on a single A100 and its weights are unreleased, flagged inline. exeterminal/Exe-Motion-Dynamic-V1-MLX is listed as a warning rather than a recommendation because its repository contains no weights. The Writing section is omitted: no text-AI release met the creation-relevant filter today.