Creative AI Briefing October 31, 2026 morning edition 10 min read

The community port now arrives within a day of the model, so the decision a creator faces is no longer whether a new model runs on their machine but which unofficial port to trust and which licence it carried along.

Idol ComfyUI-Idol Phosphene NVIDIA PiD MOSS-Image HairDeck 3d-gen local-creative-ai comfyui open-weights licensing-provenance image-gen

Creative AI Briefing: Saturday, October 31, 2026

A game artist with an eight-gigabyte graphics card can now turn one photo of a prop into a 3D object they can orbit, export and drop into a scene, for free and under a licence that allows commercial work. That model, Idol, landed on Hugging Face on Wednesday. By Friday it had a ComfyUI node pack and an Apple Silicon build, and Pika's six-step video model from Thursday was already running on a Mac. The pattern: the community port now arrives within a day of the model. The question for a creator stops being "will it run on my machine" and becomes "which unofficial port do I trust, and what licence did it bring along."

New models

Idol: one picture in, a 3D object out. speridlabs/idol (created October 28, 15:55 UTC, per the Hugging Face API; research page, code, 252 stars per shields.io) takes a single image of an object (a photo, a render or a generated picture), removes the background and builds a 3D Gaussian splat, a cloud of soft coloured points that renders as a solid object from any angle, including the sides the picture never showed. It writes a standard .ply file with up to 1,048,576 points, and there are two modes: synthetic for renders and generated images, real for photos. The model card lists a peak of 7.2 GB of graphics memory and roughly 3 to 4 GB of weights; the research page says the model has 0.66 billion parameters and runs in under 8 GB. Speed depends entirely on the card: Speridlabs' own table runs from 3.4 seconds on a B300 to 54.6 seconds on an L4 for a 262,144-point object at 50 steps, so treat the card's "under 2 seconds" headline as a data-centre number. In Speridlabs' own benchmark against TripoSplat, TRELLIS and LGM, Idol scores best on the views it had to invent; LGM is faster (1.1 seconds) and edges it on one input-view score, which Speridlabs says reflects blurry output. The weights are Apache 2.0, so you can use what you make commercially. The catches: it needs an NVIDIA card with CUDA for the official code, and you must accept the terms on two gated Hugging Face models before first run (Meta's DINOv3 image encoder and FLUX.2-dev, of which only the image decoder is downloaded), each with its own licence. Splats are not polygon meshes; most game engines and Blender need a splat plugin or a conversion step before you can light or rig the result. Free demo: the speridlabs/idol Space, on Hugging Face's shared ZeroGPU hardware.

Image

HairDeck: game-ready hair from one character drawing. hyz317/HairDeck (created October 27) takes one character image and generates stylised hair as GLB meshes with UV maps and spring-bone chains, the setup that lets hair swing when an avatar moves, then exports a VRM avatar. Each run gives four variations. Checkpoints are Apache 2.0, though SAM 3 and DINOv3 carry their own licences. There is a free demo, the hyz317/HairDeck Space on shared ZeroGPU hardware; local use wants Linux and a CUDA card. For VTuber and indie game artists, it skips the most tedious part of a stylised avatar.

MOSS-Image straight to 4K, eight steps total. lvladikov/MOSS-PiD-SuperFast-HighRes (created October 30, 23:50 UTC) is a pair of ComfyUI workflows that render a small MOSS-Image picture in four steps and hand it to NVIDIA's PiD decoder, which paints the final image at four times the width and height in four more steps, with no separate upscaler. The default turns 704 by 384 into 2816 by 1536. The catch is in the README: the PiD files sit under NVIDIA's NSCLv1 licence, and NVIDIA's model card says they "may only be used for non-commercial (research or evaluation) purposes." Workflow below.

A two-step MOSS add-on for lettering. akhaliq/MOSS-Image-IDM-2-Steps (created October 30) targets text rendering in two passes instead of the base model's fifty. The authors call the gain over the existing four-step add-on "incremental," trained it for about two hours on 249 text-heavy prompts, and recommend their own rapid-moss node because stock loading cannot apply its sampler. Apache 2.0.

Video

Pika H3 Flex on a Mac, one day later. Mrbizarro/Pika-H3-Flex-MLX (created October 30, 07:54 UTC) converts Pika's first-frame add-on for Apple Silicon and runs inside Phosphene, a free MIT-licensed Mac app for local video, image and song generation (274 stars, shields.io). On the porter's M4 Max with 64 GB, a 5-second 640 by 384 clip took 5 minutes at a measured 25.7 GiB peak; 1024 by 576 took 16 minutes for 5 seconds and 55 minutes for 15. Compare that with Pika's own notes from Thursday, where 480p needed roughly 66 to 69 GiB on an H100. The port is marked preview, the multi-reference mode is not converted, and 15-second clips drift. The licence still governs: the MiniMax H3 Community License excludes the United States, the EU, the UK and South Korea, so most readers of this briefing cannot legally run it locally.

A second Mac video port, VVladTech/Kandinsky-6.0-Lite-distill-5s-MLX (created October 30), runs Kandinsky 6.0 Lite with sound on Apple Silicon under MIT. It is early: only a roughly one-second, 256 by 432 test size works end to end, and it needs 32 GB of unified memory.

Audio and music

No new audio model shipped from a lab or vendor we track in this window. The October 28 to 30 uploads were repacks: audio-cpp/Audio8-TTS-Realistic-0.6b-GGUF (created October 29) packages Edge0's Audio8 voice model (created October 14) for the audio.cpp runtime, and agentory's music GGUFs, covered yesterday, continued.

Open and local

The local story is speed of arrival. Idol shipped Wednesday afternoon with CUDA-only code. By Friday morning PedroMarinhoDev/Idol-ComfyUI (created October 30) had converted it to run on ComfyUI's built-in TripoSplat nodes at about 6 GB, and by Friday night ddalcu had a Mac build at 3.6 GB resident, about 40 seconds per object on an M-series Ultra. The same day, Pika and Kandinsky video reached Apple Silicon. These are one-person ports with little testing behind them, but they set the new floor: a model that appears on Monday is on your laptop by Wednesday.

speridlabs/idol-sdk: the official image-to-3D code and CLI; Apache 2.0 weights, model created October 28. 252 stars, shields.io (repo).
PedroMarinhoDev/ComfyUI-Idol: drops Idol into ComfyUI with a template, idol_image_to_splat; needs ComfyUI 0.43 or newer. Model repo created October 30; 1 star, shields.io (repo).
ddalcu/mlx-serve: a local Mac server that now serves Idol, returning a GLB of textured discs most viewers can open, or true splats on request. Idol builds created October 30, figures from the Hugging Face card; 2.2k stars, shields.io (repo).
mrbizarro/Phosphene: free Mac app for local video with sound, songs and images, now with an H3 Flex preview. 274 stars, shields.io (repo).
nv-tlabs/PiD: the decoder that turns a small render into a 4K image in four steps; code Apache 2.0, weights non-commercial. 1.1k stars, shields.io (repo).
mflux-community/qwen-image-2-1-turbo-mflux-q4: Qwen-Image-2.1 Turbo for Mac via MFlux, 10.0 GB at 4-bit (the full model's 4-bit repo is about 20 GB); repos created October 30, research licence inherited (card).

Creative workflows

1. Turn one product shot into an orbitable 3D prop in ComfyUI. Source: PedroMarinhoDev/Idol-ComfyUI and GitHub repo, template idol_image_to_splat.

The steps. Update ComfyUI to 0.43 or newer. Install the node pack ("Idol (image to 3D splat)" in ComfyUI-Manager, or comfy node install comfyui-idol). Download diffusion_models/idol_fp32.safetensors and vae/idol_vae_decoder_fp32.safetensors from the Idol-ComfyUI repo (fp16 versions at half the size give the same result at default precision), plus clip_vision/dino_v3_vit_h.safetensors, vae/flux2-vae.safetensors and background_removal/birefnet.safetensors from VAST-AI/TripoSplat. Open Templates, ComfyUI-Idol, idol_image_to_splat, load your image, keep euler, simple, 50 steps, CFG 4, and run. Use the Idol Input Camera node to set the starting view for a turntable.

How it works. The model reads your picture, guesses the hidden sides, and outputs a cloud of soft points that looks solid from any angle.

Why it is good. A concept artist gets a turnaround reference, and an indie dev gets a blockout prop, from one image on a mid-range card.

Where it breaks. One-person port with no published tests; the back of the object is invented, not reconstructed; splats need a plugin or conversion before most engines and renderers can use them; for photos, set mode to real on both Idol nodes (the default synthetic is for renders and generated images).

2. Draw a 4K picture in eight steps with MOSS-Image and PiD. Source: lvladikov/MOSS-PiD-SuperFast-HighRes, files MOSS-PiD-SuperFast-HighRes-ComfyUI-local.json and -cloud.json; background in Comfy's PiD tutorial.

The steps. Place moss_image_sft_bf16.safetensors, qwen3.5_9b_bf16.safetensors and flux2-vae.safetensors (Comfy-Org/MOSS-Image), the four-step add-on moss_image_sft_4step_v2.0.safetensors, and the decoder pid_1.5_flux2_1024_to_4096_4step_bf16.safetensors plus gemma_2_2b_it_elm_bf16.safetensors (Comfy-Org/PixelDiT). Set width and height for the small render in multiples of 16 (704 by 384 by default). Stage one: 4 steps, CFG 1, euler, simple, shift 5, no negative prompt. Stage two: 4 steps, CFG 1, lcm, simple, latent format flux, degrade_sigma 0. Write long, specific scene descriptions; the README says MOSS prefers them.

How it works. The big model composes a small picture cheaply; the decoder paints the fine detail as it enlarges, instead of a generic upscaler guessing it.

Why it is good. Poster-size drafts without a separate upscaling pass, and the author ships 100 example images with prompts and seeds to compare against.

Where it breaks. The decoder weights are non-commercial, so client work needs a different finish. The listed files total roughly 39 GB on disk. No timings are published, and the Python script is NVIDIA-only (Mac users must use the ComfyUI workflows).

3. Carried forward: build a sound scene in layers with SFX v3. Source: ElevenLabs SFX v3 post (October 29) and docs. Start with the main action, add surroundings in the same prompt, set prompt influence high for literal results, loop for beds, export 48 kHz WAV. Where it breaks: counts drift, 30 seconds per generation, API access limited by region.

Worth testing

Idol Space: one image to a downloadable 3D splat, free. Tradeoff: shared ZeroGPU queue, and you still need splat tools downstream.
Phosphene: local Mac video with sound, free and open source. Tradeoff: 16 GB minimum for the smaller model, 36 GB for H3, around 40 GB of disk to start, and the H3 licence territories still apply.
MOSS × PiD workflows: 4K drafts in eight steps. Tradeoff: non-commercial decoder and a 39 GB download.
HairDeck: swinging avatar hair from one drawing. Tradeoff: the free Space queues, and local use wants Linux and CUDA.

What actually matters from today's signal

The lag between "released" and "runs on my machine" collapsed to about a day this week. That is good news for a solo creator's budget: you no longer have to rent a data-centre card to try something, and a Mac with 32 or 64 GB is now a credible place to test video and 3D. Spend your first hour with any new model on the free Space, your second on the port that fits your hardware.

The cost moves to judgment. Every port today is one person's work, tested on one machine, marked preview or early. Numbers like "25.7 GiB" and "40 seconds" come from the porter's own runs. Expect drift from the original model and budget time to compare against the official demo before you build a pipeline on it.

The licence travels with the port. Pika's Mac build is still banned in the US and EU; the PiD decoder that makes 4K look effortless is research-only; Idol is Apache, but its helpers carry their own terms. The most useful habit this week is reading the bottom of the model card before the top.

Source access notes: Quiet primary-lab window for creators, October 29 to 31: nothing new from Adobe (latest creative AI post October 19), BFL (October 15), Stability (August 25), ElevenLabs blog (October 20; SFX v3 covered yesterday), Suno (October 13), fal (October 20), Replicate (April), Luma (guides only, October 27), Krea (Ad Studio covered yesterday), Recraft (guides only), HeyGen (August), Midjourney via Releasebot (October 8), DeepMind (Genie 4 covered October 29). Failed or empty: OpenAI news (fetch error), Ideogram changelog (fetch error), Runway news (listings did not load), blog.google AI (no listings). Hugging Face blog had no creative launch. PetaPixel, Creative Bloq, No Film School and CDM had no AI tool news in the window. HN Algolia returned stale results. Trendshift showed no creative AI repo besides a design-file collection. Civitai and blog.comfy.org not attempted. All Hugging Face dates are API createdAt values. Star counts from shields.io. Adversarial fact-check ran (Sonnet subagent): it confirmed every Hugging Face date, star count, Idol, Pika Mac, Phosphene, Kandinsky, MOSS-PiD and PiD licence figure, and caught these, all fixed: the HairDeck Space is live (we had called it unconfirmed), the MFlux Qwen bullet had the wrong repo name, size and conversion date, the ComfyUI Idol port does document real mode for photos, the ddalcu Mac build landed Friday night not morning, Idol's weight size is 3 to 4 GB, and the mlx-serve Idol figures come from the Hugging Face card rather than the GitHub repo. It did not check the carried-forward ElevenLabs workflow in full.