FervorCreative AI
Live Latest 18.11.26 · morning 401 tools tracked 992 workflows indexed 626 topics Hot: Gaussian splatting, ComfyUI, ZONOS3-Mini

The community spent the weekend building the editing layer around generative output, so a creator can now fix a scan, rig a character and inspect a file for free instead of regenerating and hoping.

Splat RepairBone 2DLoftShadeNet-5ink-studioZONOS3-MiniaitraceAriadne Mark3d-genai-editingcreative-workflowsopen-weightsimage-genlicensing-provenance

Creative AI Briefing: Tuesday, November 17, 2026

You can open a phone scan of a room today, paint over the tripod leg or the stranger in the doorway, and watch the wall rebuild itself in about a second, in 3D, from every angle. That is Splat Repair, a free editor that treats Gaussian splats (the cloud of tiny soft 3D blobs that phone scans now produce) the way Photoshop treats pixels. It is one of several tools from the last 48 hours that share a direction: they do not generate anything new. They rig, repair, restyle and inspect what you already made. The big labs were quiet. The community built the workbench.

New models

ShadeNet-5 pulls the light out of a photo on an ordinary laptop. singam96/ShadeNet-5-35M (HF createdAt November 14, Apache 2.0 for code and weights) takes one photo and returns three maps: albedo (the surface color with lighting removed), relative depth, and surface normals (which way each pixel faces). Those are the three things you need to relight a product shot, drop a 3D object into a photo with believable shading, or build a quick texture. The quantized ONNX file is 35 MB, and the card reports about 0.26 seconds per image on a CPU for the 448x448 ONNX version, so no graphics card is required. Input must be a multiple of 448 pixels, which means working at low resolution and upscaling the maps. The training labels came from another model (Marigold), so it inherits that model's guesses rather than measured ground truth. There is a free CPU demo Space. Speed figures are the author's own.

Image

A MiniMax H3 Image character swap add-on landed on Hugging Face, with a license problem for most readers. akhaliq/MiniMax-H3-Image-Character-Swap (HF createdAt November 14) puts the character from one picture into the pose and scene of another. The card puts a full local run near 140 GB of graphics memory, so this is a hosted-demo tool for nearly everyone; four community demo Spaces are linked from the card, including hugging-apps. The catch is the MiniMax H3 Image community license summarized on the card: it excludes users in the EU, UK, South Korea and the USA. The card also requires MiniMax's authorization above $1 million in annual revenue or for hosted services. Read it before you use the output for anything.

moss-anypaint adds painted-mask editing to MOSS Image. yijunwang2/moss-anypaint (HF createdAt November 13, Apache 2.0, unofficial) lets you mask any region and fill it, extend the canvas, or change an object, in about 8 steps with a speed add-on. It was tested only on an RTX 5090 with 32 GB, and memory grows with canvas size, so treat smaller cards as untested.

Video

Seedance 2.5's 4K option on Runware, newly noticed, not confirmed new. The Runware model page lists a "4k" preset (3840x2160 at 16:9), priced from $0.921 per second for text or image to video and $1.14 per second for video to video. Our fetch tool summarized the changelog entry as November 15 but could not show it verbatim, and the model page itself carries an August date, so treat the timing as unconfirmed. The price is the story: a 10-second 4K clip runs about $9.21, against about $1.02 at 480p. Draft low, finish high.

Audio and music

ZONOS3-Mini is an open voice cloner small enough for a phone browser. Zyphra's blog post, dated November 12 (byline November 11; the weights repo was created November 9) and missed in our earlier runs, describes a text-to-speech model with zero-shot voice cloning (a short sample, no training) across 43 languages, with weights under Apache 2.0. Zyphra reports first audio in under 570 ms in a browser on a Pixel 9 and under 150 ms on an M3 MacBook Air; those are its own numbers. For dubbing drafts and game prototypes, an open cloner that runs on the device you already own is new. Clone only voices you have permission to use.

Open and local

The local story is tools for the second half of the job. Generators produce raw material; these repositories help you fix, animate, style and label it. Star totals come from the GitHub search API this morning (shields.io was unreachable).

  • NicolasDrum/splat-repair: a Windows editor for Gaussian splat scans with brush, lasso, clone stamp and a repair fill that rebuilds the surface you paint over. Created November 16, 37 stars, Apache 2.0 (repo).
  • iamgatz/Bone-2D: a free browser editor for 2D skeletal animation that exports the Spine 4.3 format, with meshes, weights, IK and physics. Created November 16, 3 stars, MIT (repo, editor).
  • uxmatthew/Loft-3D-SVG: 3D shapes and small hover animations drawn as plain SVG, in six looks from clay to halftone, recolored from CSS. Created November 16, 7 stars, MIT (repo, demo).
  • siliconjungle/shapeshift-labs-ink-studio: the pixel-art splat toolkit from yesterday's briefing changed from UNLICENSED to MIT and climbed from 30 to 283 stars, so you can now build on it. Created November 16 (repo).
  • ariadne-mark/ComfyUI-Ariadne-Mark: ComfyUI nodes that export images and video with invisible watermarks and signed Content Credentials (C2PA, the "made with" label standard). Created November 16, 2 stars, Apache 2.0, qualified only on Linux CPU and needs a separate SDK kit run in Docker (repo).
  • Ghongha007/aitrace: reads what an image or video file says about itself (Content Credentials, AI source labels, ComfyUI or AUTOMATIC1111 prompts and seeds, camera data) and states the limit of each signal. Offline, Python only, explicitly not a deepfake detector. Created November 16, 1 star, MIT (repo).
  • n4ze3m/tiny-voca: the 18 MB browser narrator from Saturday, now on day three. 651 stars, up from 471 yesterday, Apache 2.0 (repo).

Creative workflows

1. Clean a phone scan in 3D before you use it (NicolasDrum/splat-repair, install.cmd, app\start.cmd)

The steps: On Windows 10 or 11 with Node 22 or newer and Python 3.10 to 3.13, clone the repo and run install.cmd, then app\start.cmd. Open a .ply, .splat or .spz scan. Run Clean up to find floaters (stray blobs hanging in the air) and crop with a box. Paint the unwanted part with the Brush (B) or Lasso (L), then choose Fast Fill. For relief that shows a seam, Alt-click a clean source with the Clone Stamp (S) and paint over it. Save as PLY, which round-trips exactly.

How it works: Fast Fill fits the surface around the hole and copies whole splats from nearby, so their size, rotation and view-dependent color match. The README puts it at 0.3 to 1.5 seconds on a 550,000-splat scan, on the CPU. Deep Fill instead renders a view, fills that picture with the LaMa inpainting model and lifts the result back into splats; it needs an NVIDIA GPU and takes about 20 seconds on an RTX 4090. An optional route sends the rendered view to your own Photoshop for Content-Aware or Generative Fill.

Why it is good: Until now, fixing a splat meant deleting blobs and living with the hole, or rescanning. This works where the damage is: in the 3D file.

Where it breaks: Strong repeating relief, like a shoe sole, often shows the fill and needs hand patching. Windows only, untested elsewhere. Deep Fill adds about 3 GB of downloads and a compile step, and LaMa was trained on Places2, which has its own terms, so check them for commercial work. One developer, early release: save copies.

2. Rig and animate a 2D character for free, in Spine's format (iamgatz/Bone-2D, hosted editor)

The steps: Open the editor in a browser. Drop in your character's PNG parts and set draw order. Drag out bones (B), turn parts that bend into meshes (M, auto-traced from the transparency), bind them with auto-weights and fix with the weight brush (W). Add IK (I) for arms and legs and a physics constraint for hair or cloth. Press Tab for Animate mode, key poses, refine curves in the graph editor. Export a ZIP with Spine 4.3 JSON or binary .skel plus atlas, or render a GIF, PNG sequence or WebM.

How it works: Everything runs locally in the browser and autosaves. The export bundles a preview page that plays your file in the official Spine player. The test suite loads exports in Esoteric's own spine-core 4.3 runtime and compares every bone and mesh point against its own math.

Why it is good: Spine's Essential tier cannot save meshes or IK. This free editor does both and writes files the official Spine runtimes read, which matters if you animate AI-painted or hand-painted characters for games.

Where it breaks: Three stars and days old. Sequences and some imported features are skipped and listed in the console. Design notes are in Russian. Spine's runtimes carry their own license, which Bone 2D cannot change, so check it before you ship a game.

3. Check what a file says before you post it (Ghongha007/aitrace)

The steps: git clone, pip install ., then aitrace yourfile.png. Run it on your exports before posting and on files clients send you.

How it works: It reads embedded Content Credentials, the IPTC "digital source type" label, generator settings that ComfyUI and other Stable Diffusion tools write into files, and camera metadata, and reports each with a stated caveat.

Why it is good: Your ComfyUI PNGs can carry your full prompt and node graph. This shows you, offline, what a stranger could read.

Where it breaks: It never looks at pixels. Most files from social platforms return "no provenance signals" because uploads strip metadata, and that result means nothing either way.

Worth testing

  • ShadeNet-5 demo Space (link). Tradeoff: 448-pixel working size, so the maps are soft.
  • Bone 2D in your browser (link). Tradeoff: a days-old project; keep your source PNGs and export often.
  • Loft demo page (link). Tradeoff: not on npm, and the looks are stylized; it will not do realistic rendering.
  • Splat Repair Fast Fill on a scan you already have. Tradeoff: Windows only, and repeating textures show seams.
  • MiniMax H3 character swap Space (link). Tradeoff: the license excludes the US, EU, UK and South Korea.

What actually matters from today's signal

Generation has stopped being the bottleneck. A scan, a character painting or a video clip is cheap now. Getting it to the state you want, without starting over, is the expensive part, and that is what the weekend's tools attack. Splat Repair turns a ruined scan into a fixable one. Bone 2D turns a painting into a rig without Spine Professional ($379 on sale, $449 list). Loft turns a 3D idea into something a web designer can recolor with CSS. Each one saves a regeneration, which saves money and keeps the parts of the take you liked.

The provenance tools point at a second cost. Your files say more than you think, and clients will start asking what they say. A free reader and free signing nodes in the same weekend move "prove how you made this" from a shrug toward a routine check. The signer still needs a separately set-up SDK kit, and without production credentials it produces development-signed files, so it is a start, not a finished pipeline.

The counter-signal is the same as yesterday's: most of the new repos have one maintainer and fewer than 40 stars, and the two that took off (ink-studio and tiny-voca) are days old too. Splat Repair is tested only on Windows and Ariadne's nodes are qualified only on Linux. Use them on copies, and keep the tools you trust for delivery until these prove themselves on a real project.


Source access notes: Adversarial fact-check pass ran (Sonnet subagent). It caught: an overgeneralized star claim in the closing (ink-studio and tiny-voca are well above 40), an overstated "free signer" claim for Ariadne Mark (needs a separate SDK kit; development-signed output without production credentials), "Windows only" and "Linux only" reworded to tested or qualified only, the ShadeNet speed figure scoped to its 448x448 ONNX version, the Seedance 4K headline softened because no November date could be confirmed, the ZONOS3-Mini byline and weights dates added, and the Spine price marked as a sale price. All other checked claims held. Hugging Face (site and API), Hacker News Algolia, shields.io, OpenAI, Runware and most vendor sites were unreachable from both of our shells; Hugging Face API data (ShadeNet-5 createdAt 2026-11-14T22:05Z; MiniMax-H3-Image-Character-Swap 2026-11-14T09:55Z; moss-anypaint 2026-11-13T02:40Z) and vendor pages were read through a fetch agent. GitHub repo metadata and star totals come from the GitHub search API (per-repo API calls were refused this run); READMEs were read from raw.githubusercontent.com with cache busting. OpenAI news and the Adobe blog failed to load; the Midjourney updates page returned an empty shell; the Ideogram blog failed. BFL, Stability, Krea, ElevenLabs, Suno, fal, Replicate, Pika, HeyGen, Recraft, Runway, ComfyUI releases (latest v0.45.0, November 10) and the Hugging Face blog had nothing creative dated November 14 to 17. Luma's November 14 board organizer was covered yesterday. Comfy-Org/PAN2 (created November 15) was skipped: it repackages IFM/PAN2, a 29-billion-parameter world model created October 7. HeyGen Video Translate 1.0 appears on Runware with an April page date, so it is not new. Excluded as likely lures: two "VoiceMod desktop build" repos, twin "stable-diffusion-atelier" repos from new accounts, and an "AI video editing system" repo pushing a PowerShell installer with a fake installs badge. The Splat Repair README links its video as "I built Photoshop for Gaussian Splats in 8 days"; the YouTube title is "I Fixed the Biggest Issue with Gaussian Splatting (in 8 days)", published November 13 Pacific.

Articles from this briefing