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

A single good AI generation is turning into raw material you keep working on, since creators can now re-film a five-second clip from new angles, finish a render like film stock and train an add-on inside the graph they already use.

ai-video-in-codeZuanfilm RetrolabFastFramesComfyUI-OmniLoRAComfyUI-ZONOS3kokorolakecollage-motionlucidml GTA world model3d-gencomfyuivideo-gencreative-workflowsopen-weightslocal-creative-ai

Creative AI Briefing: Wednesday, November 18, 2026

You can take a five-second AI clip of a table full of books, train it into a 3D scene on a Mac in about fifteen minutes, and render an arc, a push-in and a rise that the original clip never contained. That recipe went up on GitHub yesterday as ai-video-in-code, and it sits in a cluster of tools that all treat one finished generation as the start of the work, not the end. A ComfyUI node finishes renders like 50 film stocks. A training pack builds custom add-ons inside the same ComfyUI graph you generate in. The labs were mostly quiet. The community spent the day making a good take worth keeping.

New models

Apple FastFrames turns a big open video model into a two-step, near-live generator, for research only. apple/FastFrames-L (HF createdAt November 16) is Wan 2.1's 14-billion-parameter text-to-video model squeezed into what the card calls "autoregressive, two-step video diffusion models that generate 480p video at interactive latency." Autoregressive means it builds the clip a chunk at a time, so frames can start appearing while the rest is still being made; two steps means it does in two passes what the base model does in dozens. Default output is 832x480, 81 frames at 16 fps. The card publishes timing only for the small 1.3B sibling (FastFrames-S), not for the 14B one, and its larger examples use 8 GPUs, so treat "interactive" as a lab claim on lab hardware. The license is Apple's research-only model license, which excludes "any commercial exploitation," so you cannot use the model in commercial work or products (the license text does not clearly address generated output). A community-built demo Space is linked from the model's metadata; we did not test it. Why it matters anyway: it shows where open video is heading, toward previews you steer as they render instead of waiting minutes per take.

Image

Zuanfilm Retrolab gives ComfyUI a film lab. zuanfilm/Zuanfilm-Retrolab (created November 17, Apache 2.0, 9 stars) is one node with twelve finishing stages: white balance, an "Anti-AI" slider, exposure, bloom, halation (the red glow film gets around bright highlights), light leaks, a 50-look film emulsion menu (Portra, Velvia, CineStill 800T, instax and more), solarization, soft focus, vignette, paper and dust textures, and grain. Every stage defaults to off, and the node returns your image untouched until you move something. It needs no model download and no extra Python packages, runs on CPU or NVIDIA, and keeps grain repeatable across a batch of frames while holding light-leak shapes still, so it works on image sequences as well as stills. The README is plain that the looks are "original, independently authored creative RGB transforms" named after stocks, not calibrated copies of them, and that Anti-AI is a tonal correction, "not an AI detector."

Microsoft lists MAI-Image-2.7 and a Flash version in preview. SiliconANGLE's November 17 report says MAI-Image-2.7 and MAI-Image-2.7-Flash are "now in preview," with sharper editing and more consistent generation at lower cost. The base MAI-Image-2.7 appears to predate this report, so the Flash tier may be the main new part. A DEV Community post says Microsoft lists Flash at about $10 per 1,000 images (against about $35 for the full model) in Microsoft Foundry and on fal, OpenRouter and others. Secondary sourcing only: we could not load Microsoft's own page, so treat the price as unconfirmed.

Video

Seedance 2.5 Flash, a cheaper Seedance tier, appears in third-party listings. A CSuite roundup puts it at 4 to 30 seconds at 480p or 720p with its own audio, at $0.07 per second at 480p and $0.15 at 720p on BytePlus, which would make a 10-second 720p draft about $1.50. Secondary sourcing only, and the roundup gives two different dates (November 15 and 17); we could not reach a BytePlus page to confirm price or date.

A GTA world model you can drive with WASD. The lucidml blog post (November 15) describes a 630-million-parameter model trained on GTA V footage: hold the movement keys and it draws the game one frame at a time with no engine underneath. The author says weights and inference code are open under Apache 2.0 at lucidml/gta, with a Space to try it. The author's own caveats: not trained to convergence, short memory, one GPU per player, sessions capped at 180 seconds by default. The training footage comes from a commercial game, so treat output as research play, not material for client work.

Audio and music

ZONOS3-Mini gets ComfyUI nodes. claudianeodev/ComfyUI-ZONOS3 (created November 16, MIT, 3 stars), from Promptus, puts Zyphra's open voice cloner from November 9 into a ComfyUI graph: clone from a clip (up to 60 seconds is used), pick one of 27 built-in voices, or design a new voice with 25 sliders for age, pitch, pace, mood and room. It runs Zyphra's own local server alongside ComfyUI instead of inside it, which keeps its audio libraries from breaking other node packs. One-click setup is Windows with NVIDIA; elsewhere you start the server yourself. Clone only voices you have permission to use.

kokorolake makes Kokoro speak instantly on Intel laptops. sortedcord/kokorolake (created November 18, Apache 2.0, 7 stars) runs the open Kokoro-82M voice on the built-in graphics of Intel Core Ultra 200H laptops with hand-written code. The author's own table on a Core Ultra 5 225H: 79 ms to first audio against 2,624 ms for the standard PyTorch version, 18x faster than real time, 478 MB of memory. It fits that one chip family only.

Open and local

Today's local story is that the second pass got cheaper. You can re-shoot, re-grade, re-voice or teach your generator a character without leaving the tools on your own machine. Star totals come from the GitHub search API this morning (shields.io was unreachable).

  • sirioberati/ai-video-in-code: turns a five-second AI clip into a 3D scene of soft blobs you can film from new camera paths. Created November 17, 5 stars, MIT (repo).
  • zuanfilm/Zuanfilm-Retrolab: one ComfyUI node for film looks, grain, halation and light leaks, on stills or frame sequences. Created November 17, 9 stars, Apache 2.0 (repo).
  • cyberterm/ComfyUI-OmniLoRA: trains a custom style or character add-on (a LoRA) inside ComfyUI, from 10 to 25 images, using the model files already on your canvas. Created November 17, 37 stars, Apache 2.0; the long list of supported models and GPUs is the author's claim (repo).
  • eydrenn/collage-motion: an agent skill that makes paper-cutout videos with no image generator at all; the agent writes shapes and a timeline, your computer renders. Created November 16, 1 star, MIT (repo).
  • yedp123/ComfyUI-Yedp-Face3D: builds a rough 3D face from 478 tracked landmarks to guide face shape in image generation through depth and normal passes; the author calls it experimental and "not a replacement for LoRA." Created November 17, 5 stars, MIT (repo).
  • siliconjungle/shapeshift-labs-ink-studio: the pixel-art splat sculpting kit, now 368 stars, up from 283 yesterday, after its switch to MIT (repo).
  • NicolasDrum/splat-repair: yesterday's splat repair editor, up from 37 to 76 stars (repo).

Creative workflows

1. Re-film a five-second AI clip from camera moves it never had (sirioberati/ai-video-in-code, run_clip.sh, render.mjs, tools/get_brush.sh)

The steps: On a Mac, brew install ffmpeg colmap node, pip3 install numpy, clone the repo, npm install, then ./tools/get_brush.sh (it fetches the Brush 0.3.0 trainer and checks its checksum). Put a clip at footage/myclip.mp4. Start python3 tools/serve.py 5880 . in one terminal and run ./run_clip.sh myclip in another. Collect export/myclip-new-camera.mp4. For your own move, call render.mjs with push, lift, tilt, u0 and u1, then join the frames with ffmpeg.

How it works: ffmpeg pulls every second frame (about 60 for five seconds). COLMAP works out where the camera was in each frame. Brush trains a cloud of Gaussian splats (soft colored blobs with a position, size and opacity) for 5,000 steps until the blobs, seen from those cameras, match the frames. A small WebGL page then draws the blobs from any new path. On an M3 Pro the README reports 12 to 19 minutes of training and about 20 seconds to render a five-second move.

Why it is good: A generated clip you love usually comes with one camera move you cannot change. This keeps the scene and lets you add coverage: a reverse, a push-in, a slow rise for a title. Only the source clip is AI; everything after it is ordinary code.

Where it breaks: The camera in the clip must move, or COLMAP cannot solve it. Plain surfaces turn to smoke (the README's white mug example). Moves that stray outside what the clip saw go black or blurry at the edges. The renderer draws each blob's base color only, so shine and reflections flatten. Tested only on macOS with an M3 Pro.

2. Finish AI renders like film, inside the same graph (zuanfilm/Zuanfilm-Retrolab, comfy node install zuanfilm-retrolab)

The steps: Install from ComfyUI Manager or the Registry command, restart, put the node after your decode or Load Image node, and queue once so the live preview has a source. Then adjust without re-queueing: pick a stock, set Emulsion Response (0.5x to 3x), add a little Anti-AI, halation and grain, and check the 240-pixel native crop for grain size. Queue again to write the full-resolution image.

How it works: Every stage is a plain math transform in linear color, applied in a fixed order with grain last. Seeds keep texture and grain repeatable, so a batch of frames gets matching but non-identical grain.

Why it is good: The plastic look of many AI renders comes from tone and color, not detail. Fixing it in the graph saves a round trip through Lightroom or Resolve for every variation.

Where it breaks: The looks are named after film stocks but are not measured copies of them. A heavy preset hides a bad render rather than fixing it. Nine stars and one day old.

3. Train a character add-on without leaving ComfyUI (cyberterm/ComfyUI-OmniLoRA, example_workflows/omni_lora_basic_workflow.json)

The steps: Install from ComfyUI Manager, drag in the basic workflow, pick your model, text encoder and VAE in the usual loaders, point the dataset node at a folder of 10 to 25 images, set a trigger word, and queue. The trained file lands in ComfyUI/models/loras/ and the patched model can go straight into a sampler to test.

How it works: It pre-computes your images and captions once, trains the add-on on the model already loaded, and shows a live loss graph with a save button. Captions are optional; missing ones fall back to a default caption plus your trigger word.

Why it is good: No second Python setup and no copying model files into a separate trainer.

Where it breaks: Training memory is still the limit; the README suggests settings for 8 to 12 GB cards but no measured results per model. The model support list is long and the project is a day old, so test on a small run first.

Worth testing

  • ai-video-in-code on a clip you already have, with ./run_clip.sh name 1000 for a quick pass. Tradeoff: locked-off shots fail outright.
  • Retrolab on your last ten renders. Tradeoff: easy to overdo; compare Before and After at full size.
  • lucidml GTA Space (link). Tradeoff: short memory and a 180-second session cap.
  • FastFrames-L community Space (link). Tradeoff: research-only license, 480p.
  • collage-motion with any coding agent. Tradeoff: one look only, paper cutout, by design.

What actually matters from today's signal

The most expensive thing in AI video is a reroll. You get a shot you like and then need a second angle, a different grade, or the same character in another scene, and the usual answer is to generate again and hope. Today's tools attack that cost from three sides. ai-video-in-code gives you new coverage from the take you have. Retrolab gives you a grade without a trip to another app. OmniLoRA makes a character add-on a ComfyUI job instead of a separate project. Each one turns a lucky generation into an asset.

FastFrames points the other way, at generation fast enough to steer while it renders. If that arrives under a usable license, the reroll gets cheap instead of avoided. For now it is research-only and measured on lab hardware, so the keep-and-rework path is the one you can use this week.

The counter-signal is maturity. Every repo above is one to three days old, most have under 40 stars, and each is tested on one platform: ai-video-in-code on one Mac, kokorolake on one Intel chip family, ZONOS3's one-click setup on Windows with NVIDIA. Run them on copies, and do not put them on a deadline yet.


Source access notes: Adversarial fact-check pass ran (Sonnet subagent). It caught: the FastFrames license sentence overstated (it restricts model use and does not clearly address outputs), an unsupported "late October" date for MAI-Image-2.7 plus a missed secondary price source, an unproven "two days" timing for ink-studio's relicense, an unsupported pose-and-lighting claim for Face3D (README says shape guidance via depth and normal passes), "last week" loosely dating ZONOS3-Mini (createdAt November 9), and an unconfirmed "one start frame" detail for the GTA model. All fixed; every star count, creation date, license, quote and README figure otherwise held. Hugging Face (site and API), Hacker News Algolia, shields.io and every vendor site were unreachable from both shells; Hugging Face API data (FastFrames-L createdAt 2026-11-16T17:21Z; LongCat-Video-Avatar-1.6 createdAt 2026-11-13, outside the window and skipped; ZONOS3 family dates) and vendor pages were read through a fetch agent. GitHub star totals and creation dates come from the GitHub search API; READMEs were read from raw.githubusercontent.com with cache busting. OpenAI news failed to load and Runway redirected to runway.com and was not followed. Google, DeepMind (month-only dates), Adobe (newest Nov 11), BFL, Stability, Replicate, fal and Suno (newest Nov 13) had nothing creative dated November 15 to 18. ElevenLabs' November 16 post is an acquisition (Kotoba), not a creative release; Luma's November 16 post is an engineering write-up. The FastFrames arXiv page was rate-limited, so its paper claim is unread. Microsoft MAI-Image-2.7-Flash and Seedance 2.5 Flash rest on secondary sources only. pareekshit28/4danyone-workbench was skipped as a fork of ant-research/4DAnyone. Midjourney, Civitai and blog.comfy.org were not attempted given the egress block.

Articles from this briefing