Creative AI Briefing: Sunday, September 6, 2026

Type vh5tape, badly damaged VHS tape with heavy tracking errors in front of a scene description and MiniMax H3 hands back a 1980s local car commercial with tracking bands, chroma bleed, head-switching noise at the frame edge and matching muffled mono audio, free in a browser. That LoRA also carries the sentence organizing this entire week: do not load it on a rotated or quantized base, because it will load with no error at all and then give you melting limbs. Nobody released a generator in this window. What shipped was compression, five or six independent efforts to make last month's models fit on the card you already own. The compression works. It also quietly breaks the adapters you already have, and the point where it starts breaking is different for every model.

New models

AbsorbQuant is the most useful artifact of the window, and it is not a model at all. It is seven pre-quantized checkpoints, uploaded September 5, that push six popular open image models down to four bits in both weights and activations. FLUX.1-dev and FLUX.1-schnell arrive at 6.6 GB each, SDXL base and SDXL-Turbo at 1.7 GB, SANA-1.6B at 822 MB, PixArt-Sigma at 325 MB, plus an OCP MXFP4 PixArt build at 308 MB. Four-bit weights are ordinary now. Four-bit activations are not, and they are what buys the speed rather than only the disk space.

The method explains why the file survives the squeeze. Each layer splits into a rank-32 fp16 branch on the raw input plus a GPTQ-quantized NVFP4 residual carrying everything else. The full-precision low-rank part absorbs the outliers that normally poison a four-bit grid, so the residual quantizes cleanly, and it dispatches to nunchaku's real gemm_w4a4 kernel (3.9k stars, cache-busted read September 6) rather than a dequantize-and-pretend path. The licenses are stated correctly, which most quantizers skip: code is Apache 2.0, while every checkpoint inherits its base, so FLUX.1-dev is Black Forest Labs Non-Commercial and the SDXL files are CreativeML OpenRAIL++-M.

Now the catch, and it is a large one. Every install instruction starts with git clone https://github.com/chenjiaj109550158/AbsorbQuant. That repository does not exist: a cache-busted shields.io read returns "repo not found" and the GitHub API returns nothing. The download.py, generate.py and verify.py scripts the provenance section leans on are not public. You have 27 GB of real, dated, correctly licensed weights and no documented way to load them (weights, nunchaku).

Image

NanoDiffuser went up September 5 and squeezes a working text-to-image pipeline into roughly 382 MiB of model assets that run entirely in a browser tab on WebGPU. One denoising step, classifier-free guidance off, 512 by 512 out, PNG download. CLIP sits at 63.36 MiB with asymmetric Q4 weights, the U-Net at 312.71 MiB mixing Q4 matrix weights with per-output-channel INT8 convolution storage, and a TAESD decoder at 2.36 MiB in FP16.

The card refuses to oversell itself, which is rare enough to reward. INT8 reduces convolution storage while the convolution still executes in floating point, it says so. The sub-three-second target "has not been verified across devices." And its cosine-similarity gates, 0.96 for CLIP and 0.90 for the U-Net, are labelled "export acceptance thresholds, not measured benchmark scores or a general visual-quality rating." OpenRAIL++ inherited from the source, free hosted demo (weights, demo).

Video

Yesterday's lead story got 46% smaller overnight. VDN-H3, the hybrid-attention branch that makes MiniMax H3 render faster than it plays, had a 4.3 GB linear branch. On September 5 someone shipped an INT8 ConvRot build of just that branch at roughly 2.15 GB, claiming about 4.7 GB less peak VRAM at load. ConvRot means Hadamard Convolutive Rotation at group 256, the same outlier-spreading trick everything else this week is built on. It drops into ComfyUI/models/VDN_Minimax_H3/stage-dmd-step-250/linear_branch/ and inherits the MiniMax H3 Community License with its territory clause. Treat the numbers as the author's own: the card offers a YouTube walkthrough where a benchmark table should be (weights).

vh5tape is the style LoRA of the week and the best-written model card in the batch. Rank 32, 5,000 steps on fal's H3 trainer at 4:3, joint video and audio objective, trained on hand-curated four-second clips of period broadcast material: sitcoms, soaps, commercials, news, station sign-offs. The damage is a prompt dial with three trained levels, from "lightly worn VHS tape with slight analog noise" to "badly damaged VHS tape with heavy tracking errors and distortion." Because H3 trains audio jointly, the hiss and warble arrive without asking. The author reports 5,000 steps beating the 2k, 3k, 4k and 6k siblings in blind A/B, and that past 5k the damage got tamer rather than heavier.

Then the part to read twice. The card names the bases that break it: *_int8_convrot, nvfp4, w4a8. Those store weights in a rotated or quantized basis, so the LoRA "loads without any error and then produces warped faces, melting limbs and disappearing objects." And the default checkpoint in the official ComfyUI H3 tutorial is minimax_h3_fl2va_pruned_int8_convrot, which the author says outright you must change. True of MiniMax H3 adapters generally, not just this one (weights, demo).

Approach one carefully. Minimax-h3_Singularity landed September 5 with 28 likes and 55 GB of int8 fusion checkpoints, claiming HDR clarity, distant-face restoration and better martial-arts motion. No method detail, no comparison, no benchmark, an affiliate invite code on its only demo link, and a card declaring license: apache-2.0 on a derivative of MiniMax H3 weights that are not Apache anything (weights).

Audio and music

Thin window. Gepard 1.1 arrived September 5 under a real Apache 2.0: 555 million parameters, voice cloning in English, Spanish, Portuguese and Dutch, trained with GRPO on top of gepard-1.0 and LAION's Emolia set. It ships a BENCHMARK_REPORT.md and a technical report PDF beside the weights, more accountability than most small TTS drops offer, and there is a free hosted Space if you want to hear it before downloading half a gigabyte (weights, demo).

One date correction, because it will fool a scanner. A Breeze TTS 2 repo appeared September 6 with a full card and a leaderboard chart. It is a mirror. The real release is BreezeBlue/Breeze-TTS-2, API createdAt August 25, now at 451 likes and 6,357 downloads. The license split is the thing to know: source Apache 2.0, weights and self-hosted outputs research and non-commercial only, commercial rights only through the paid hosted platform (upstream, demo).

Open and local

The local story this week is that a handful of people started measuring quantization instead of shipping it and hoping. Star totals are cache-busted shields.io reads taken September 6.

• nunchaku-tech/nunchaku: the W4A4 kernel almost every four-bit image and video build here dispatches to, including AbsorbQuant and the rotated NVFP4 H3 line (3.9k) (repo)
• JoaoZaokk/comfy-quant-bench: the harness behind the three-build Wan comparison, and the only place this week where somebody wrote a prediction down before measuring and then published being wrong (2) (repo)
• OpenVDN/vdn-minimax-h3: yesterday's hybrid-attention release, up from 261 stars on September 5, all four training stages open (291) (repo)
• T8mars/comfyui-minimax-h3-audio-T8: the H3 node pack this ecosystem's ComfyUI side routes through (956) (repo)
• breezeblue-ai/breeze-tts: Apache 2.0 inference code for Breeze TTS 2, voice design and voice direction included (285) (repo)

Creative workflows

1. Shoot a 1980s broadcast day, with the correct base checkpoint. LoRA at KennethFal/vh5tape-vhs-lora-minimax-h3. Files: vh5tape.safetensors, or vh5tape-comfyui.safetensors for the version with explicit alpha tensors. Prompts and seeds for every sample: PROMPTS.md.

The steps. Drop the file in ComfyUI/models/loras/. Load a base that is not rotated: minimax_h3_fl2va_bf16.safetensors is the recommendation, minimax_h3_fl2va_pruned_fp8_scaled.safetensors is acceptable. Wire UNETLoader into LoraLoaderModelOnly into MiniMaxH3ImageToVideo. Strength 1.0, useful range 0.7 to 1.2. Prompt order is trigger, then damage tier, then scene. Render 480P at 4:3, roughly 20 to 25 full-precision steps.

How it works. Trigger and damage phrase were both trained through H3's own caption format, so the tier phrases behave like a dial rather than keywords you hope land. Because H3 generates audio in the same pass, describing the sound ("muffled mono audio", "canned laughter", "cheesy jingle") gets you tape-worn audio for free, and short quoted dialogue comes through close to verbatim.

Why it is good. No conversion step, all 104 modules load without warnings, and the aesthetic is one phrase you can turn up or down mid-project instead of a grade you apply afterward.

Where it breaks. Stacked with a turbo LoRA, drop this one to 0.4 to 0.6: the two perturbations add and a four-step path has no headroom, and four-step turbo softens the tape grain, which is the opposite of the point. Stay near 1 MP, single shot per prompt. The base-checkpoint rule above is not advice, it is the failure mode.

2. Find your own four-bit threshold before you trust a converted checkpoint. Bench at JoaoZaokk/comfy-quant-bench. Worked example with all three builds published: Wan2.1-VACE-1.3B-W4A4-ConvRot.

The steps. Convert your model at more than one promotion threshold. Render the same prompt at the same six seeds through each build and through the unquantized reference. Look at the pictures. Run probe_quant_dispatch.py --forward-only to confirm the quantized math is executing rather than silently dequantizing.

How it works. "Effective error" is measured per layer, on the real activations that layer saw during sampling, against the format that layer actually received. For Wan 2.1 VACE 1.3B, a build at 0.0546 (2 of 300 layers at four bits, 2.15 GiB) matches the FP16 reference; 0.0793 keeps the subject and smears everything else; pure W4A4 at 0.1602 is coloured wreckage. Z-Image v2 at roughly 6B tolerates 0.1421. The HunyuanVideo 1.5 family at roughly 13B tolerates 0.1837.

Why it is good. It kills the idea that a converter default is safe. The author's version: --promote-error 0.15, chosen once on a 6B model and carried everywhere since, "writes a file here that loads cleanly, dispatches natively, passes every structural check, and renders a smear."

Where it breaks. One prompt, three to six seeds, 480x480, one scheduler, one card, no perceptual metric, and the size monotonicity rests on three data points. The page also carries two dated self-corrections from September 1, and they are the most instructive part: the earlier renders ran at cfg 1.0 with no shift and no negative prompt, a distilled model's operating point, which drifted the FP16 reference too; and an earlier threshold table credited a break to the wrong model family.

Worth testing

• vh5tape in the hosted Space, no install, the fastest way to hear jointly-trained audio degradation. Tradeoff: the Space runs someone else's base, so it is the ceiling, not what your local rotated checkpoint gives you (demo).
• NanoDiffuser in a WebGPU browser, zero install, zero account. Tradeoff: one step, no guidance, 512 by 512, speed target explicitly unverified (demo).
• Breeze TTS 2 for voice design, describing a speaker in a sentence instead of supplying reference audio. Tradeoff: research and non-commercial weights, with commercial rights only through the paid platform (demo).
• Rotated NVFP4 FastH3 if you own a 16 GB Blackwell card: 12.8 GB single file, zero offload, a 480p 124-frame clip in about 25 to 30 seconds of denoise at four steps. Tradeoff: custom loader node, a specific nunchaku wheel, Linux, and euler at sigma shift 12, because the author reports res_multistep strobing on neon and rain (weights, loader).

What actually matters from today's signal

Rotation is the technique of the week and almost nobody outside these model cards is naming it. The trick is a Hadamard transform applied before the four-bit matrix multiply, spreading a layer's outlier values across many channels so no single large number dominates the grid. It works. The rotated NVFP4 H3 build reports roughly 70% better on a Laplacian-variance sharpness proxy against identical un-rotated weights, at about 15% cost per step. AbsorbQuant does the same job with a rank-32 full-precision branch instead. Both get you a model that fits where it did not fit last month.

The bill arrives on the adapter side. A rotated base stores its weights in a different basis than the one your LoRA was trained against, and nothing in the stack checks. The LoRA loads. The layer count matches. The render comes out wrong in a way that looks like a bad prompt. If you spent this year collecting style LoRAs and also switched to a smaller checkpoint to fit your card, you may have already lost half your library without seeing an error, and the officially documented default is the rotated one.

The fix costs nothing but attention. Keep one full-precision or fp8 base for adapter work, use the four-bit build for renders where no LoRA is involved. And treat every converter default as unproven on your model: the Wan bench shows an error level that destroys a 1.3B model sitting comfortably inside the safe band for a 13B one. There is no threshold of the format, only one per model, and the person who found that wrote down a wrong prediction first and then showed you the pictures. That is the standard the rest of this space should be held to.

---

Source access notes: blog.comfy.org served a cached July 2024 page and was skipped. The Hugging Face daily-papers API and the ComfyUI GitHub releases API both returned responses over the fetch size limit and were not read. github.com/chenjiaj109550158/AbsorbQuant returned "repo not found" from both a cache-busted shields.io read and the GitHub API, reported above as a finding rather than a fetch failure. Bucket A was quiet: OpenAI's latest creative-relevant post predates the window, Adobe's September 4 item is a Times Square marketing story, Replicate last posted August 4 and fal August 27, so no vendor launch is omitted. Writing is omitted because nothing qualified. Every date above comes from the Hugging Face API createdAt field, not a listing timestamp; the Breeze TTS 2 entry documents a mirror repo that would otherwise have read as a September 6 release. An adversarial fact-check subagent ran against this draft and caught one error: an earlier version claimed Gepard 1.1 had no hosted demo, when its own card links a live Space.