Creative AI Briefing: Wednesday, September 9, 2026
Run one git clone, open localhost in a browser, and a fifteen-second 864×480 clip with its own generated stereo soundtrack comes out of a single Radeon card sitting in a desktop. That package went up at 06:56 UTC this morning. It is the fourth release in two days whose entire pitch is speed: OpenAI cut image generation latency by up to half, a compiler team cut open video generation by 16.66 percent on AMD hardware, a Wan LoRA removed half the transformer calls without touching the step count, and Z-Image Turbo moved into a browser tab. Not one of the four publishes a quality-parity proof against the thing it replaces. The ceiling did not move this week. Your patience threshold did.
New models
ChatGPT Images 2.5 shipped September 8 to every ChatGPT tier on desktop, mobile and web, and the number OpenAI leads with is the one that matters: people already make more than 3 billion images a week across ChatGPT Images and the GPT-Image API models. The model claims sharper detail, more natural lighting, better preservation of subjects from your reference photos, and more reliable multi-turn editing where an earlier change survives the next five. Latency is down as much as 50 percent against Images 2.0.
Three product features arrived with it. Sketch lets you draw a rough layout directly in ChatGPT and hand it over as the visual guide, invoked by typing @Sketch. Templates give you a starting frame for formats like Poster and Merch instead of an empty prompt box. And you can share the prompt alongside a finished image, so someone else can run your setup against their own photos.
Two API models replace one. GPT-Image-2.5 Flare is described as the default choice for most applications, pitched at higher quality than GPT-Image-2 at 50 percent lower latency. GPT-Image-2.5 Sunburst trades generation time for tighter control across edits. Manus reports Flare running at two to four times the speed of GPT-Image-2 in its own evaluations, and Adobe confirms the 2.5 models are live inside Firefly, which is the practical route for anyone who wants this next to layers. C2PA metadata and invisible watermarking carry through (announcement, system card).
MiniMax H3 now runs on one AMD card, which gives Radeon owners their first documented single-GPU path to this model after a week in which every other port targeted Blackwell or Apple Silicon. EliovpAI/MiniMax-H3-W4A8-Paiton-RDNA4 went up at 06:56 UTC today for one 32 GB Radeon AI PRO R9700 on gfx1201. The default workflow produces a continuous 15.08-second clip at 864×480, 24 fps, with native 32 kHz stereo audio generated jointly. Install is one command that clones the plugin repo and runs launch.sh, which pulls a versioned GHCR image, verifies checkpoints against a lock file and opens ComfyUI with the workflow already wired.
The measurements are the reason to trust it. Paiton takes 332.85 seconds against 399.40 stock for a matched request, 16.66 percent lower latency and 19.99 percent more clips per hour, with fresh conditioning and MP4 writing inside the clock. All three paired long clips have equal video and audio latents and byte-identical MP4s, so the engine comparison itself is exact. Then the card says something almost nobody says out loud: "The 20% end-to-end latency engineering target remains unmet; the 19.99% throughput gain is a different metric." The same admission appears again in the older September 8 section, phrased as an objective rather than a target.
Read the component table before you plan around this. Conditioning falls from 21.718 seconds to 10.821. Sampling falls from 317.987 to 262.296. Video decode does not move at all, 42.521 against 42.473, and neither does audio decode or the mux. Roughly one eighth of your wall clock is a VAE that nobody optimized, and it is now the largest untouched block in the pipeline. Host requirements are modest by the standards of this category: the long-clip run was tested on a 16 GB system with 4 GB of swap, process RSS peaking at 12.09 GiB and up to 1.24 GiB of process swap sampled, with 24 GB or more recommended once you add a browser. Allow 60 GB of disk (release, full benchmarks).
One structural note worth carrying: the required Qwen3-VL-32B intermediate encoder here uses NVFP4 storage with FP16 compute and, in the card's own words, "is not replaced by a smaller model." That is a different engineering choice from the two-card Blackwell pipeline covered yesterday, which substituted an abliterated encoder. Same base model, opposite decision about what you are allowed to swap.
Image
The Krea 2 second-day wave arrived all at once. Krea 2 Turbo itself is not new, with a Hub creation date of June 18, but on September 8 four separate packages landed against it: an Apple Silicon MLX conversion at roughly 10.5 GB (kruatech/Krea-2-Turbo-MLX, README currently empty, described here only from repository metadata), a GGUF set for llama.cpp-style runtimes (ldov/Krea-2-Turbo-GGUF), modular-diffusers reference-image blocks (OzzyGT/krea2_reference_blocks, Apache 2.0), and a deliberately paired style-LoRA ablation (shigure0110/krea2-style-lora-pair).
The MLX conversion is the one to read even if you never download it, because its gate text states the Krea 2 Community License in plainer English than the PDF does. Commercial use is permitted only below one million US dollars in company-wide annual revenue on a trailing twelve-month basis across all affiliated entities. You must implement content filtering in any deployment end users can reach. And if you redistribute, you pass the agreement on, name the derived model beginning with "Krea," and state that it was modified. That naming clause catches people.
Delighting became a category overnight. Two releases today attack the same problem from opposite ends. vLAR/RelightFormer (07:16 UTC) is a SIGGRAPH Asia 2026 feed-forward relighting transformer fine-tuned from Wan 2.1, trained on a new open Laval-Objaverse Dataset of 90,545 assets under 39,008 illumination conditions, with multi-view consistency as the selling point. It is CC BY-NC-SA 4.0, so it is a research tool, not a client tool, and training took 4× H200 (weights). DirtScan/trail-delighter (06:56 UTC) is the opposite: a small Apache-2.0 LoRA over FLUX.2 Klein 4B that removes harsh sun, cast shadows and lens flare from mountain bike trail photographs. Its method is the interesting part and it is covered in the workflows section below (adapter).
Z-Image Turbo runs in a browser tab with nothing installed. lwanming/Z-Image-Turbo-opt (September 9, 04:39 UTC, Apache 2.0) is an ONNX conversion targeting WebNN and WebGPU through ONNX Runtime Web, split into seven sub-models with the precision of each one published: text encoder and transformer at Q4F16, and the scheduler step, VAE preprocess, VAE decoder, safety preprocess and safety checker at FP16. Microsoft hosts a live demo you can open right now (weights, demo). And elismasilva/ltx2.3_image_custom_blocks (September 8, Apache 2.0) turns LTX-2.3 into a still-image model, with an AutoBlocks pipeline that picks text-to-image or image-to-image from whichever inputs you pass, at eight steps and 704×1280 (blocks).
Video
A CFG-only distillation LoRA landed at 05:37 UTC this morning, and it is refreshing because it refuses the fashionable trick. Perflow-Shuai/Wan2.1-T2V-14B-CFG5-to-CFG1-50Step-LoRA-r128-iter500 is Apache 2.0 over Wan 2.1 T2V 14B and states its position in the first paragraph: "This is not a few-step, DMD, SFP, or autoregressive checkpoint. It keeps the native 50-step schedule and removes the unconditional inference branch." Because Wan computes v_cfg = v_uncond + g * (v_cond - v_uncond), running at guidance 1.0 is exactly the conditional prediction, which means one transformer call per step instead of two. You keep all fifty steps and halve the compute.
The honesty is equally direct. Validation is eight held-out prompts across three arms with matched seed, noise and scheduler state, and the result is a mean gap closure of 23.29 percent toward the CFG=5 reference, which the card itself calls "a small diagnostic suite, not a broad benchmark." The run migrated from 16 data-parallel ranks to 8 at step 92, and the card says the continuation is sample-equivalent rather than bitwise identical. One finding to flag: the card cites AndysonYs/LongLive-LoRA as the public training code, and that repository returns "repo not found" from a cache-busted shields.io read, an empty response from the GitHub API and nothing from raw.githubusercontent, checked three ways this morning (adapter).
Smaller items: PocketAiHub/Sulphur-2-Distilled-MLX-Q8 (September 8, 08:33 UTC) is an 8-bit MLX pack of an LTX-2.3 derivative for Apple Silicon at roughly 29.7 GB, shipping a MODIFICATIONS.md and SHA256SUMS alongside the weights. OpenVDN/vdn-minimax-h3 reads 389 stars on a cache-busted check this morning, against 366 yesterday and 261 on September 5.
Audio and music
The audio story this window is a defect, and it is sitting in a benchmarks file rather than a headline. The Radeon H3 package measured its own generated audio and reported that a four-step pouring clip produced two channel samples at or above full scale, peaking around 1.00053, while the eight-step version of the same prompt produced five, peaking around 1.02007. The card states plainly that "the native waveform reaches its decoder clamp" and that "no normalization, audio replacement or mixing hides these shared limitations."
That is a real problem for anyone cutting native-audio video into a timeline. Clipping at the decoder means the distortion is baked into the file you receive rather than introduced by your export, and the only reason we know is that one team bothered to measure peak sample values and print them.
A second detail from the same package deserves attention. Everything in that pipeline is quantized except one component: the denoiser is W4A8, the encoder is NVFP4 storage with FP16 compute, the video VAE is INT8 weights with BF16 compute, and the audio VAE is left at FP32. Somebody decided audio was where the bits had to stay. The clip arithmetic does not line up cleanly either, with 165,600 audio samples at 32 kHz giving 5.175 seconds against 5.1667 seconds of video, plus AAC padding to 165,888 on decode. The card explicitly warns against calling the result exactly five seconds or sample-aligned.
Open and local
The pattern holds from yesterday and sharpened: the artifact that changes what you can do is a runtime, not a checkpoint. Star totals are cache-busted shields.io reads taken September 9.
- Eliovp-BV/paiton-vllm-plugin: the ROCm container stack and launch script behind today's single-Radeon MiniMax H3 release, and the same repo that shipped the RDNA4 FLUX.2 Klein package two days ago (0) (repo)
- antirez/h3.c: from-scratch C and Metal MiniMax H3 for Apple Silicon, still the best-documented speed-quality tradeoff table in the category (2.6k) (repo)
- OpenVDN/vdn-minimax-h3: hybrid-attention H3 with all four training stages published, climbing steadily all week (389) (repo)
- microsoft/webnn-developer-preview: hosts the free browser Z-Image Turbo demo and the reference WebNN pipeline behind it (69) (repo)
- leejet/stable-diffusion.cpp: the runtime the new image GGUF drops target, including this window's Krea 2 Turbo quantizations (6.9k) (repo)
- ostris/ai-toolkit: still the trainer under most community video LoRAs shipping this week (12k) (repo)
- kohya-ss/musubi-tuner: named by name in several of the Krea 2 character LoRAs published September 8 (2k) (repo)
Creative workflows
1. Generate fifteen seconds of video with native audio on one AMD card, then verify you are not being sold a number.
Package at EliovpAI/MiniMax-H3-W4A8-Paiton-RDNA4. Files: launch.sh, run.sh, comfyui/workflows/turbo8-15s.json, checkpoints.lock.json, BENCHMARKS.md.
The steps. Clone github.com/Eliovp-BV/paiton-vllm-plugin, then run ./models/MiniMax-H3/launch.sh. Open 127.0.0.1:8190/?paiton=1, which loads the Turbo8 fifteen-second workflow with the engine toggle already in the graph. Edit the prompt, fix the seed, click Run. For terminal work, stop the interface with ./launch.sh --stop, then ./run.sh download --profile turbo8 and ./run.sh generate --engine paiton --preset turbo8 --frames 362. To check the vendor's claim yourself, run ./run.sh benchmark once per engine into separate fresh output directories, then ./run.sh report.
How it works. The compiled denoiser fuses Turbo residuals, paired SwiGLU and RDNA4 INT8 attention into one artifact rather than replacing the model. Both engines load the exact same pinned W4A8 checkpoint and Turbo adapter, so the comparison isolates the runtime. Frame counts snap upward to H3's 17k+5 grid, which is why 360 becomes 362 and the clip is 15.0833 seconds rather than 15.000.
Why it is good. Fifteen continuous seconds with a coherent subject and a generated soundtrack, on one card, with no per-generation billing and nothing leaving the machine. And the evidence is unusually complete: raw per-run timings, file hashes, sampled GPU energy, disk reads per phase and temperatures are all published, plus a rejected-experiments section that reports a 4 GiB host cache saving only about 1.3 seconds and causing swap, and a native RDNA4 INT4 WMMA probe that produced a coherent clip, changed image detail and composition, and still came out slower than the compiled W4A8 path at roughly seven seconds per denoiser evaluation. Lower bit width did not mean faster here, and it did mean different pixels.
Where it breaks. Linux with Docker and working /dev/kfd and /dev/dri only, one specific GPU, and 60 GB of disk before anything renders. Audio clips at full scale on some prompts, as measured above. Four-step pouring duplicates the bottle; eight-step pouring overfoams. The card states outright that "original full-precision quality parity is unproven," and that hosted context processing, 2K regeneration, image and reference conditioning, larger canvases and other durations are all outside the qualification.
2. Turn a hosted editing model you pay for into a small local LoRA you own. Worked example at DirtScan/trail-delighter, Apache 2.0 over FLUX.2 Klein 4B. Teacher endpoint: fal's Qwen Image Edit lighting-restoration LoRA.
The steps. Collect a set of your own source images that share the exact problem you keep fixing by hand. Run a selection of them through the hosted model that already solves it well, and keep both the input and the output for each one. That paired set is your teacher set. Train a LoRA on those pairs against a small base you can run locally, which in this case means FLUX.2 Klein 4B rather than the much larger Qwen model that produced the targets. Ship the adapter.
How it works. You are not distilling the teacher's weights, which you do not have. You are distilling its behavior on your specific distribution of images, through a dataset. The hosted model is a labeling service you rent for one afternoon. The student never needs the teacher's capacity because it only needs to reproduce one transformation on one kind of picture, and that is a far smaller function than general image editing. The author states the tradeoff in one line: using Klein as the base makes the result "smaller and easier to run."
Why it is good. It converts a recurring per-image API cost into a fixed one-time cost plus a file you keep. It works for any repetitive edit you already outsource, whether that is delighting, background cleanup, colour matching or product straightening. And it gives you a local, offline path for work you cannot upload.
Where it breaks. Every error the teacher makes becomes a rule the student learns, and there is no correction step in this loop. The published example ships no dataset size, no rank, no learning rate and no evaluation, so the recipe is sound but the settings are yours to find. The result is narrow by construction: a delighter trained on sunny trail photographs is not a general delighter. And check the teacher's terms of service before you build a training set out of its outputs, because that permission is not universal. The adapter here is Apache 2.0 but explicitly "subject to the terms of the base FLUX.2 Klein model."
Worth testing
- Z-Image Turbo in your browser, no install, no account, at Microsoft's WebNN demo. Honest tradeoff: the transformer and text encoder are Q4F16, so this is a preview-quality path, and first load pulls the full model set over the network.
- ChatGPT Images 2.5 Sketch, free on any tier by typing
@Sketch. Honest tradeoff: you get no control over the model version, no local copy, and OpenAI applies C2PA plus invisible watermarking to everything, which is fine for most work and disqualifying for some clients. - Krea 2 Turbo on Apple Silicon via
kruatech/Krea-2-Turbo-MLX, roughly 10.5 GB. Honest tradeoff: the README is currently empty, so there is no documented usage path, and the licence caps commercial use at one million dollars of company revenue with a mandatory "Krea" prefix on anything you redistribute. - The CFG-only Wan LoRA for anyone already running Wan 2.1 T2V 14B at fifty steps. Honest tradeoff: iteration 500 of a research run closing 23 percent of the gap on eight prompts, validated only at the exact matched contract of 50 steps, CFG 1.0 and LoRA scale 1.0.
- LTX-2.3 as a still-image model through
elismasilva/ltx2.3_image_custom_blocks. Honest tradeoff: this is a video model's single-frame path, so expect its motion-trained priors to show up in your stills, and there is no published comparison against a dedicated image model.
What actually matters from today's signal
Everything shipped this window competes on time, not on output. That is worth taking seriously rather than dismissing as a plateau, because for most working creators the binding constraint has never been the ceiling of the best possible generation. It has been how many attempts fit between the brief landing and the deadline. Halving image latency, as OpenAI claims against Images 2.0, changes the shape of an afternoon more than another point of prompt adherence would. So does going from nine to eleven finished fifteen-second clips per hour on a card you already own.
But notice what none of these releases contain. Not one publishes a quality comparison against the full-precision or full-guidance thing it stands in for. The Radeon package says parity with original H3 is unproven and means it. The CFG LoRA closes under a quarter of the gap to its own teacher and says so. The browser build quantizes the transformer to four bits and does not pretend otherwise. The speed is measured; the cost of the speed is asserted or simply left blank. When you adopt one of these, you are accepting an unmeasured quality debt in exchange for a measured time credit, and the only person who can price that trade is you, on your own material.
The counter-signal worth watching is the video VAE. On the AMD numbers, decode is 42.5 seconds of a 333-second job and did not improve by a single tenth of a second, because every optimization effort this year has gone into the denoiser. Sampling has now been attacked from four directions in two weeks: fewer steps, fewer guidance branches, cheaper attention, compiled kernels. Keep that up and the decoder becomes the whole story. The team that publishes a faster video VAE without changing a pixel will move creator wall clock more than the next distilled adapter will.
Source access notes: All Hugging Face dates are API createdAt values in UTC, never listing "Updated" timestamps. Star counts are cache-busted shields.io JSON reads taken September 9. GitHub file fetches used raw.githubusercontent.com with cache busting. Verified three ways and reported above as a finding: AndysonYs/LongLive-LoRA, cited by the Perflow-Shuai card as its public training code, returns "repo not found" from cache-busted shields.io, an empty body from api.github.com and an empty body from raw.githubusercontent.com. kruatech/Krea-2-Turbo-MLX has complete frontmatter and a zero-byte README, so it is described only from API metadata and its gate text. Blocked, stale or unusable sources this run: stability.ai/news served a cached listing whose newest items are from the Stable Diffusion 3.5 era, so no Stability claim in this briefing rests on it; blog.comfy.org remains a JavaScript wall; civitai.com and midjourney.com/updates were not reachable in a usable form and nothing here depends on them. Bucket A was otherwise quiet on creative work: ElevenLabs' most recent post is a September 2 chief revenue officer hire, Adobe's blog has published nothing on creative AI since September 4 and appears here only through OpenAI's own quotation of Adobe on Firefly, Google Labs' newest item is a consumer story-curation app, and Runway's most recent research post is GWM Worlds 2 on September 3, outside this window. Items found and deliberately excluded: ldov/ideogram-4-gguf (September 8) is a genuine quantization set for ideogram-ai/ideogram-4-fp8, but that base carries a non-commercial licence and the GGUF repository bundles a folder of unrelated explicit LoRAs, so it is noted here rather than recommended. Several accounts continued bulk-mirroring MiniMax H3 and Krea 2 derivatives overnight; sorting the Hub by creation date currently returns more re-uploads than releases. The "Writing" section is omitted because nothing qualified. Base models referenced as substrate rather than news: MiniMax H3, Krea 2 Turbo (Hub createdAt June 18, 2026), Ideogram 4 (May 30, 2026), Wan 2.1 T2V 14B, LTX-2.3, FLUX.2 Klein 4B and Z-Image Turbo. An adversarial fact-check subagent ran against this draft and caught five substantive problems, all corrected above. It found that the "20% target remains unmet" sentence appears verbatim only once, with the second instance reworded from "engineering target" to "objective," so the text no longer claims two identical quotations. It found that the bit-equal, byte-identical result covering ten paired clips belongs to the earlier September 8 short-preset round, while the 15.08-second test reports three paired long clips, and the number has been corrected. It found that the rejected INT4 WMMA probe "changed image detail/composition," which an earlier draft omitted while calling the probe correct. It removed an uncited superlative claiming a first-ever single-GPU path outside NVIDIA and Apple, which no source supports. And it caught that "15.5 GiB usable RAM" belongs to the September 8 round's host description, not the long-clip round, which states 16 GB system RAM with 4 GB swap. Every other load-bearing figure, date, quotation and star count survived the pass unchanged, including the audio full-scale peaks, the unmoved video decode times, the Krea 2 licence terms and the unreachable LongLive-LoRA repository.