Creative AI Briefing: Friday, August 28, 2026

A five-second video clip now takes about three seconds to generate. Not on your machine, on a rack of rented Blackwell chips, but that number crossed a line yesterday: generation is now faster than playback. Meanwhile a Pixel 8a can edit a photo from a text instruction with the model running entirely on the phone, offline, no account, no per-image charge. That second thing has been true since July. Almost nobody noticed.

That gap is today's story, and it is not the one the release feeds tell. The vendor news this week is thin and it is all hosted, all metered. The interesting work, the models small enough to run on a phone or a MacBook, was finished weeks ago and is sitting in public repositories waiting for someone to use it. If you have been tracking announcements, you have been looking in the wrong place.

New models

H3 Max landed from fal on August 27, and the number that matters is three seconds. That is how long it takes to generate a five-second clip, roughly 35 times the throughput of the official MiniMax H3 endpoint it was built from. fal took the open-weights H3 model, retrained it for prompt adherence and visual quality, then rebuilt the serving stack around it on NVIDIA GB200 hardware. In their own human preference study against twelve models including Veo 3.1, Kling 3, Wan 3.0 and Seedance 2.5, H3 Max ranks first on overall quality, prompt understanding, and aesthetics. Those evaluations are fal's, though Artificial Analysis and Design Arena report the same top placement independently. Hosted only, priced per generation, 50% off for the first week (announcement, text to video). At three seconds a generation you stop planning shots and start playing with them, which is a real change in how the work feels even though you are renting it.

That is the entire list of new model releases in the window. The rest of this briefing is about what is already available and underused.

Image

Adobe shipped a Photoshop update on August 27 with two features worth knowing. Instruct Edit with Masks lets you describe an edit in words while everything outside the mask stays untouched, so faces, logos, and brand assets are protected while the model works on the rest. Markup lets you draw on the image to show the model what you mean, sketching an arrow for position or brushing a rough shape for a new element, instead of trying to say it in a prompt. A new AI Assisted Editor sits alongside the Pro Editor in beta for anyone who wants a prompt bar rather than a layer stack (Adobe).

The more interesting image story is older and free. Google's LiteRT community has three diffusion models converted to run entirely on a phone, and all three publish their measurements. FLUX.2 klein 4B (posted July 9, card refreshed August 27) does generation and editing on-device, split into 21 graphs that load one at a time so peak memory is a single 912MB chunk rather than the 6.2GB total. It runs on a Pixel 8a's Mali GPU, a phone with no dedicated video memory at all. The catch is honest and large: 256 by 256 pixels, four steps, roughly five minutes per image (klein, Z-Image Turbo).

Bonsai Image 4B (July 29) is the one to watch. It uses ternary weights, meaning every value in the main network is one of three numbers, which compresses to 2.11GB with no rounding loss at all. Working set under 4GB, and on an iPhone 17 Pro it makes a 512 by 512 image in about 64 seconds on CPU alone, bit-identical to the same run on a Mac. Apache 2.0, so output is yours to sell (model card, hosted demo).

Also worth a look if you shoot: Canter (July 28, updated through August) is a 2-billion-parameter text-to-image model built for photography, trained by one person on a single GPU. It runs 1024 by 1024 on 8GB of VRAM and ships its own web interface (python app.py --in-browser). The license is unusual, ModelGo Attribution-ShareAlike 2.0, which permits commercial use and hosting but requires attribution and source disclosure, so read it before building on it. The author is blunt: text rendering is unreliable and it has seen almost no illustration (model card, demo).

Video

FastMetal is the most useful thing on this list if you work on a Mac, and it has been sitting there since August 10 with almost no attention. FastVideo published three text-to-video models, all shipped already compressed to INT8 so there is no conversion wait at startup, and distilled down to a handful of denoising passes instead of fifty. The three are not simply small, medium and large: the 1.3B and 14B are built on Wan2.1 at 448 by 832 and 77 frames, while the 5B is built on Wan2.2 and targets 720p. Match the tier to your Mac's memory, 1.3B for 16GB machines, 14B for 24GB and up with 36GB comfortable. Apache 2.0 throughout. Full setup in the workflows section below (14B, 5B, 1.3B).

The same team posted a four-step distillation of MiniMax H3 on August 21 that generates video and matching audio in one pass. They label it clearly as checkpoint 1400 of a planned 4000, so it is worth knowing about and not worth installing yet (preview).

Different in kind, and the most fun thing here: Alibaba's AMAP lab published an interactive world model on July 9 that runs at 720p and 16 frames per second on a single RTX 5090, responding to your input as it generates rather than playing back a finished clip. Apache 2.0, 19GB of memory, 1.2 seconds of lag, and a free browser demo that takes ten seconds to try (ABot-World, try it).

Audio and music

Since July 25 there has been a 640MB bundle that runs Stable Audio 3 Small music generation entirely inside a browser tab, no server and no install, using 4-bit weights and plain WebAssembly. On an M-series Mac it takes 60 to 120 seconds for a 10-second clip. The author measured it at about 0.88 envelope correlation against the original, meaning the same musical structure with slightly more high-frequency grit, and says so plainly. WebGPU would be much faster and was deliberately skipped so the bundle works from any static file host. Stability AI Community License, which restricts commercial use above a revenue threshold (bundle).

From the same week, posted anonymously because it is under peer review: an LTX-2 based model that takes a text prompt plus reference voice clips and generates a multi-speaker scene, binding each voice to its speaker through the text. Roughly 8.5GB, and it needs Google's Gemma 3 12B separately as its text encoder. Interesting for dialogue and radio drama, but anonymity means no support and no roadmap (weights).

Open and local

The pattern, stated honestly: the on-device creative stack is finished and idle. Three phone-capable image models, three Mac-native video models, and a browser-based music generator all exist right now, all with permissive or near-permissive licenses, and their download counts are in the low thousands. These are not previews. They ship measurement methodology and failure modes, which is more rigor than most vendor launches manage. The bottleneck is not capability any more, it is that nobody has written the guide.

Star totals verified via shields.io:

• comfyanonymous/ComfyUI: the node graph almost every local model here targets; if you run creative AI on your own machine, this is the surface (130k stars) (repo)
• Lightricks/LTX-2: video and audio generated together in a single pass, and the base for the anonymous multi-speaker audio model above (9.3k stars) (repo)
• Lightricks/ComfyUI-LTXVideo: the official nodes and example workflows for running LTX inside ComfyUI (4.1k stars) (repo)
• hao-ai-lab/FastVideo: makes video models run on hardware you own, and the source of the three Apple Silicon checkpoints (4k stars) (repo)
• city96/ComfyUI-GGUF: loads compressed model files in ComfyUI, which is how most people fit large video models on consumer cards (3.9k stars) (repo)
• google-ai-edge/LiteRT: the runtime behind the three on-device image models (3.3k stars) (repo)
• kijai/ComfyUI-KJNodes: the utility node pack the community LTX workflows depend on; must be current for LTX-2 support (2.9k stars) (repo)
• amap-cvlab/ABot-World: real-time interactive world generation on one desktop GPU (2.5k stars) (repo)

Creative workflows

1. Generate video on a Mac in three passes instead of fifty. Start at the FastMetal collection and pick the tier your machine can hold.

The steps. Pick FastMetal-1.3B-QAD for a 16GB Mac, FastMetal-5B-QAD if you want 720p, FastMetal-14B-QAD if you have 24GB or more (36GB is comfortable). Install with pip install torch transformers mlx safetensors av imageio imageio-ffmpeg, clone github.com/FastVideo/FastVideo, then run examples/inference/basic/mlx_wan_prompt_to_video.py pointed at the downloaded folder. Each repo also carries an ema/ subfolder holding a smoothed variant of the same weights; point --mlx-checkpoint there if your outputs look jittery.

How it works. Two compressions stack. The model was distilled, meaning a student network learned to reach a finished frame in three passes where the teacher needed fifty. It was also compressed during that training rather than after, so the network learned to produce good frames while already living at INT8 precision, which is why it does not degrade the way a bolted-on conversion does. The weights arrive already compressed, so there is no wait at startup.

Why it is good. Video generation on a Mac has been an emulation story until now. This is a native path under Apache 2.0, which means output goes into client work without a license review.

Where it breaks. Apple Silicon and Python 3.11+ only, so no Windows and no Intel Macs. The 14B is locked to 448 by 832 and 77 frames with no audio track. Its weights are about 15GB, so a 16GB Mac will not hold it no matter how patient you are. These have been public since August 10 with very few downloads, which means you will be an early tester and should expect rough edges nobody has reported yet.

2. Run LTX-2.5 in ComfyUI using the community workflow pack. The files live at RuneXX/LTX-2.3-Workflows, which despite the name now covers 2.5.

The steps. Grab the models from Lightricks/LTX-2.5, which already ships as split files that load without repacking. You need gemma4-12b-with-proj-ltx-2.5-bf16.safetensors as the text encoder, a video VAE and an audio VAE, and the spatial upsampler. For the higher-detail path you also need ltx-2.5-22b-ic-lora-pixel-spatial-upscaler-x2-1.0.safetensors from the separate IC-LoRA repo, at a fixed strength of 0.5. Install ComfyUI-KJNodes and ComfyUI-GGUF, update both, then update ComfyUI itself. Start at 121 frames.

How it works. The distilled pipeline generates a fast, lower-resolution result. The DFR pipeline then adds a second stage that upsamples in latent space (the compressed form the model works in, before it becomes pixels) and applies a detailing adapter, which is why a two-stage run at 1088 by 1920 beats a single-stage run at the same size.

Why it is good. LTX-2.5 generates picture and sound in one pass, so footsteps in the shot are the footsteps in the audio. This is the shortest route from download to a working graph, and the models stay compatible with most LTX-2.3 adapters.

Where it breaks. The author has not shipped dedicated 2.5 workflows yet and tells you to reuse the 2.3 ones, so expect to fix a node or two. Both custom node packs must be on their latest versions for LTX-2 support, and a stale install fails in confusing ways. Commercial use is free only under a $10M annual revenue threshold.

3. Put an image model on a phone. The conversion recipe is public at john-rocky/hf-to-litertlm, and the finished graphs are at litert-community.

The steps. Fastest result: download the Bonsai Image graphs (dit_int4b32.tflite, textenc_int4.tflite, vae_dec_fp32.tflite), run pip install ai-edge-litert numpy pillow transformers, then python generate.py --model-dir . --prompt "a red fox sitting in fresh snow at sunrise". For klein on Android, stage the tokenizer/ folder alongside the graphs so the app can encode a typed prompt rather than a baked-in one, and set GpuOptions(precision = FP32).

How it works. These models are too big to load as one file on a phone, so each is cut into chunks that load, run, and unload in sequence. Peak memory becomes the largest single chunk rather than the whole model. Bonsai goes further: its weights are already ternary, so the 4-bit container stores them with no rounding at all.

Why it is good. No network, no account, no per-image cost, and output bit-identical to a desktop run. For anyone building a creative app, this is the difference between a server bill and a download.

Where it breaks. Slow. Bonsai takes about 64 seconds on an iPhone 17 Pro, klein five minutes on a Pixel 8a. Klein is fixed at 256 by 256 with one reference image, and changing either means re-exporting. Two settings silently corrupt output: compute must be FP32, since the modulation layers overflow at half precision, and the attention mask must be pre-expanded across heads.

Worth testing

• Bonsai Image on Hugging Face: the ternary phone model, free in a browser, no install. Tradeoff: 512 by 512 output, so this is for testing whether the quality holds up, not for final assets.
• ABot World interactive: steer a generated world with your own input in real time. Tradeoff: a research toy with no way to export a clean take.
• H3 Max on fal: the fastest quality video generation available today, at 50% off through roughly September 3. Tradeoff: hosted only, metered, and you cannot take the weights with you.
• Canter demo: photography-focused image generation from a one-person project. Tradeoff: it cannot render legible text, so nothing with type in it.
• MiniMax H3 inpainting: describe a change to part of a clip and leave the rest alone. Tradeoff: struggles when the region you name overlaps heavy motion.

What actually matters from today's signal

Two things are true at once and they point in opposite directions. Renting frontier speed keeps getting better, and H3 Max at three seconds a clip is not something your tower will do this year. And separately, the free local tier crossed the line from research demo to usable weeks ago and the audience has not arrived. Download counts on the Mac video models are in the hundreds. On the phone image models, low thousands. These are not obscure projects, they come from FastVideo and from Google's own edge runtime team.

The practical read: if you have been waiting for local creative AI to be ready before you invest time in it, the waiting is over and has been for a month. The stuff on your list to try later is already finished. That is also the argument against buying a GPU right now. The fast tier will be rented regardless, and the free tier increasingly runs on hardware you already own, which leaves the mid-range tower squeezed from both sides.

The counter-signal is that on-device remains slow and small. Five minutes for a 256-pixel image is a demo, not a workflow, and anyone claiming phones are ready for production is selling something. What makes these releases credible is that they say so themselves. The klein card names the two graphs that failed to run on the GPU delegate at all. The Bonsai card explains that its low similarity score is sampler drift rather than damage, with sharpness measurements to back it. The browser audio bundle says WebGPU would be faster and why it was skipped anyway. That is a higher standard of disclosure than the benchmark charts at the top of this briefing.

---

Source access notes: An earlier draft of this briefing dated nine community releases to August 27 and 28 based on the "Updated N hours ago" timestamps on Hugging Face listing pages. Those are modification times, not publication times. Every date here has been re-verified against the Hugging Face API createdAt field, and the affected items are correctly dated to July and mid-August. Only three items were genuinely published in the last 24 hours: fal's H3 Max, Adobe's Photoshop update, and HeyGen's TAVR paper (SIGGRAPH Asia, no public weights, omitted from the body). A MiniMax H3 compression I had planned to cover was dropped: the repository URL redirects to a differently-specified variant, so the figures could not be tied to the card. blog.comfy.org was unreachable and GitHub's releases page returned stale cache. Midjourney, Civitai, Kling and Seedance had no fetchable primary announcements in the window. The Writing section is omitted: no text-AI release met the creation-relevant filter. H3 Max preference rankings are fal's own human evaluations, labeled inline. Quantization quality figures are each author's own published measurements.