Creative AI Briefing: Wednesday, October 28, 2026

A 3D artist can photograph a ceramic mug, upload it to a free Hugging Face Space, and get back the three texture maps a renderer asks for: base colour, roughness and metallic. That add-on appeared last night, a few hours before a port that takes Google's live music model, an instrument that has only played in real time on Apple Silicon Macs since June, to NVIDIA and other PC graphics cards in about 2 GB of memory. The pattern today is the second-day port: community builders are moving tools onto hardware and into jobs their makers never shipped for. The machine you already own decides less this week. The licence on the original decides more.

New models

Magenta RealTime 2, ported to PCs. audio-cpp/Magenta-RealTime-2-GGUF (created October 28, 05:33 UTC, per the Hugging Face API) repackages Google DeepMind's Magenta RealTime 2 for the audio.cpp engine. The model makes 48 kHz stereo instrumental music from a text prompt or an audio example, steerable with notes, and streams it as it generates. Google's own apps and AU plugin (Jam, Collider, MRT2) run in real time only on Apple Silicon Macs, per the project README. The port's own benchmark on an RTX 5090: the 230M Small model runs at a real-time factor of about 0.08 (a 30-second clip in under three seconds) in about 2,150 MiB (roughly 2 GB) of graphics memory, against a 24,626 MiB peak (about 24 GiB) for Google's Python code on the same card; the 2.4B Base model peaks at about 4,200 to 6,300 MiB. The card does not say how much of the Python figure is preallocation. Streaming mode hands back audio in blocks of up to 160 ms. Licence, plainly: the weights are CC-BY 4.0, so commercial use is allowed with attribution, plus Google's usage terms, which bar outputs that infringe others' rights; Google "claims no rights in outputs you generate." The catch: the card's linked model guide was unreachable this morning, and audio.cpp's latest release (v0.9.0, September 30) does not mention Magenta, so the commands on the card may need a newer build than you can download today. Instrumental only, and every speed figure is the porter's own, from one GPU and four requests per setting.

Photo to texture maps for 3D. Mike0021/qwen-image-edit-2511-pbr-lora (created October 27, 19:23 UTC) is an add-on for Qwen-Image-Edit-2511 that turns a lit photo of an object into one PBR map per request, the maps game engines and renderers use to decide how a surface reacts to light. Three fixed instructions on the card ask for albedo (colour with lighting removed), roughness (black for glossy, white for rough) or metallic (white for bare metal). It was built for a projection-texturing tool, where each view of a 3D model is decomposed and the maps are projected and baked onto the mesh. Try it free in the projtex-pbr Space (created October 27, running on ZeroGPU). Licence: Apache 2.0 over an Apache 2.0 base; training renders used CC0 and CC-BY materials. The honest catch: trained only on synthetic Blender renders and tested on 42 views of 14 sample assets. On that test the recommended v2 file (pbr2/step-4215) scores 0.74 metallic accuracy (0.75 for v1), against 0.72 for a map that simply predicts no metal anywhere; the card does not define the metric. The card says v2 stops reading glossy paint as metal, though its own table does not show that fix, and its roughness error is about 10 to 14 percent worse than v1's.

Image

A 12 MB photo denoiser. IvmeLabs/Ivme-Denoise-v1 (created October 27, Apache 2.0) cleans noisy sRGB photos: about 70 ms for a 1024-pixel square frame on an RTX PRO 6000, while large photos need the supplied tiling (6.5 seconds for a 12.3-megapixel frame), and a CPU takes seconds per megapixel. The card is candid: it trails the larger DRUNet and Restormer models by 0.22 to 0.54 dB, has no real-sensor benchmark, wants gamma-encoded images rather than RAW or HDR, and may drift hue in saturated areas. Free Space.

Krea 3 at 4K through the API. Krea's October 27 changelog adds a 4K setting for Krea 3, billed as a 2K image plus a 2x upscale, unavailable with ultra or Zero Data Retention.

Video

Seedance without a ByteDance account. The October 28 changelog puts Seedance 2.5, 2.0, 2.0 Fast and 2.0 Mini behind a Krea API token. Per the Seedance guide, 2.5 runs 4 to 30 seconds at up to 1080p and 2.0 reaches 4K; inputs include start and end frames, reference images, video and audio. Failed and cancelled jobs are not billed. The catch: the guide prints no prices and points to a pricing page, and the October 27 entry says the Seedance 2.5 unlimited promo has ended.

One add-on for edges, depth and pose on LTX 2.3. zxyty/physicalai-ltx-2.3-ic-lora-union-control (created October 27) steers LTX-2.3 video from a reference clip's edges, depth or pose, reading the reference at half resolution to cut the work per frame. LTX-2 Community License; training data only partly listed.

Audio and music

A 1980s Eurodisco sound for YuE2, trained in 32 minutes on an 11 GB card. jamespetts/yue2-1980s-eurodisco-lora (created October 27, 18:52 UTC) changes how a YuE2 song sounds (chorused synth bass, gated toms, stacked pads) without changing song structure, so the model's other add-on slot, the one that shapes structure, stays free. Put eurodisco first in the style prompt. The author trained it on 11 of his own AI-generated Eurodisco tracks on an RTX 2080 Ti and published the training workflow file. The licence tangle: the card says CC0, but the YuE2 base is CC BY-NC 4.0, and the trainer's README says add-ons trained with it, and audio made with them, carry those non-commercial terms. A separate packager, LattifAI, writes that the upstream licence lets individual creators monetize outputs; we could not confirm that wording in the m-a-p model card this morning. The training songs being AI-generated adds another unknown. Ask before you sell anything made with it.

YuE2 in 2.6 GB. LattifAI/YuE2-3B-GGUF (created October 27) shrinks the song model to a 2.57 GB file plus a 60 MB decoder; on a 4090, a 22.3-second song took a median 9.09 seconds in about 5.6 GiB, per the packager. Tested only in its own EdgeSpeak app; the card calls llama.cpp loading unverified.

Open and local

The local story is reach. Google's real-time Magenta apps and plugin run only on Apple Silicon (its Python library runs offline on NVIDIA cards); a community port claims a PC path in a tenth of the memory. The YuE2 trainer recommends 24 GB; one musician trained a usable add-on on 11. A Qwen editing model that was built to change photos now outputs material maps. None of this is new science. It is people closing the gap between what a model can do and the machine sitting on a creator's desk.

0xShug0/audio.cpp: one local engine for speech, music and separation models on Windows, Linux and Mac, with a built-in web interface; the Magenta port targets it. 3.8k stars (shields.io) (repo).
audio-cpp/Magenta-RealTime-2-GGUF: Google's live instrumental model in 1.2 to 5.5 GiB files; created October 28 (card).
Mike0021/qwen-image-edit-2511-pbr-lora: photo in, albedo, roughness or metallic map out, Apache 2.0; created October 27 with a free Space (card).
Starnodes2024/ComfyUI-YuE2-Trainer: train a YuE2 sound add-on on 5 to 30 of your own songs inside ComfyUI; MIT code, 75 stars (shields.io) (repo).
IvmeLabs/Ivme-Denoise-v1: a 12 MB photo denoiser that runs on almost anything; created October 27 (card).
sekkit/UltraTex-Qwen-Image-2.1: multi-view mesh texturing on Qwen-Image 2.1, MIT; created October 27 (card).
LattifAI/YuE2-3B-GGUF: full songs with vocals in a 2.6 GB file; created October 27 (card).

Creative workflows

1. Turn a product photo into a three-map material set. Source: qwen-image-edit-2511-pbr-lora card and the projtex-pbr Space.

The steps. Shoot the object evenly lit, no hard reflections. In the Space, or locally with Qwen-Image-Edit-2511, the pbr2/step-4215 file and the Lightning 4-step add-on at 768 pixels, run the card's three instructions one at a time: albedo, then roughness, then metallic. Load the three maps into a principled material in Blender (albedo to Base Color, the other two as non-colour data) and test under a different HDRI than the photo.

How it works. The add-on learned from Blender renders where the true maps were known for every pixel, so it learned to subtract lighting from a picture.

Why it is good. A prop artist gets a starting material from one photo in minutes instead of painting three maps by hand.

Where it breaks. Synthetic training, a small test set, a metallic map that barely beats guessing "no metal" on the author's benchmark, and 768-pixel output. Check the metallic map first, and repaint it by hand if it disagrees with what you know the object is made of.

2. Train YuE2 on your own sound with an 11 GB card. Source: eurodisco card with its yue2-nar-lora-train.json, and the trainer README.

The steps. Install ComfyUI-YuE2-Trainer, download the yue2_3b_bf16 file (about 7.8 GB; the smaller INT8 file cannot be trained), and point it at a folder of 5 to 30 consistent songs you own. On a smaller card, follow the README: shorter clips (6 to 8 seconds; the Eurodisco run used 8) and the 8-bit optimizer. The Eurodisco run used 2,000 steps and took 32 minutes. Generate with the trigger word first and cot set to off, as the trainer README recommends.

How it works. YuE2 has a planner that writes the song and a renderer that makes it sound like something; the trainer only teaches the renderer, so melody and lyrics stay under your prompt's control.

Why it is good. A producer can teach the model a production palette from their own catalogue in under an hour on a 2018 graphics card.

Where it breaks. It is a sound and timbre add-on, not a voice clone, per the README. Some seeds drift into modern EDM gloss, per the card. And the licence question above is unresolved.

Worth testing

projtex-pbr Space: free texture maps from one photo. Tradeoff: 768-pixel maps from a synthetic-trained add-on.
Ivme-Denoise Space: drop in a noisy night shot. Tradeoff: sRGB only, and slightly behind heavier denoisers.
Magenta RealTime 2 apps: Jam and the AU plugin on a Mac. Tradeoff: Apple Silicon only; Base needs an M3 Pro or M2 Max class chip.
Krea Seedance API: one token for four Seedance versions. Tradeoff: prices are on a separate page and the unlimited promo is over.

What actually matters from today's signal

Ports are now where most creators meet new models. The lab ships a model shaped for one platform; a community packager makes it fit the rest. Magenta RealTime 2 is the clean case: four months of Mac-only live music, then a GGUF file that claims a PC path in 2 GB. That is good for anyone with a gaming PC and a song idea, and it means the useful question about a model is no longer "is it open" but "has someone ported it to my machine yet."

The cost is trust. Every speed figure in today's ports comes from the porter's own test on one card, and one port shipped ahead of the runtime that runs it. Run your own ten-second test before you plan a session around any of them.

And licences do not port cleanly. The Eurodisco add-on calls itself public domain while its trainer says anything made with it inherits a non-commercial base. The PBR add-on is Apache all the way down. Magenta is CC-BY with usage terms. Three releases, three answers to "can I sell this." Read the base model's licence, not the add-on's badge.

Source access notes: Quiet primary-lab window. Nothing new for creators on October 27 to 28 from OpenAI, Adobe (latest October 19), BFL, Stability, ElevenLabs, Suno, fal or Luma (covered yesterday). Runway news, Google blog and Replicate explore showed no dates. Midjourney via WebSearch only: nothing current. GitHub trending served a stale page (years-old repos) and was not used; Trendshift showed no star figures and no creative repos, so no deltas are cited. HN Algolia returned only October 6 to 12 stories. The GitHub API (ComfyUI and audio.cpp commits) was blocked, audio.cpp's commit page is disallowed by robots, and the Magenta model guide linked from the GGUF card failed to load. Civitai and blog.comfy.org were not attempted. Etsy's October 22 AI listing rules were omitted as covered in recent briefings. DongShanHu/Qwen-Image-Edit-2511-Multiple-Angles-LoRA (created October 28) was omitted as a duplicate of fal's earlier release. All Hugging Face dates are API createdAt. Star counts are shields.io totals. Adversarial fact-check ran (Sonnet subagent): it confirmed every Hugging Face date, licence, file size, star count, the Magenta speed table, the trainer's non-commercial statement and the Krea items, and caught eight wording errors, all fixed: a misread PBR metallic score (v2 is 0.74 against a 0.72 no-metal baseline), an unsourced roughness-cost attribution, MiB/GB unit mixing, an Ivme speed figure misstated per megapixel, the Eurodisco training songs being AI-generated, the 2080 Ti's age, unsupported wording about Google and Apple chips, and the Seedance promo date. It could not read the audio.cpp Magenta guide or confirm LattifAI's monetization claim, which m-a-p's card does not contain.