FervorCreative AI
Live Latest 21.11.26 · morning 422 tools tracked 1040 workflows indexed 651 topics Hot: Claude Code, Gaussian splatting, ComfyUI

A photo of a face and a recorded voice line now make a talking video at every price point, free in a browser, free and local in ComfyUI, or paid through an API, so the scarce skill moves from getting a mouth to move to writing and voicing something worth hearing.

LongCat-Video-Avatar 1.6ComfyUI-LongCat-Avatar-PixelAILabsGoogle VidsGenie 4Suno VisualsYuE2MiniMax-Music3ComfyUI-H3-Image-EasyFixavatar-genvideo-gencomfyuimusic-genopen-weightscreative-workflows

Creative AI Briefing: Thursday, November 19, 2026

Drop a chest-up photo and a nine-second voice line into ComfyUI and you get back a presenter who says the line, lips in sync, in about seven minutes on one big graphics card. That recipe for Meituan's open LongCat-Video-Avatar 1.6 went up yesterday, the same day ComfyUI's own team posted repackaged model files and opened a pull request for built-in support. Google made personal avatars in Vids free for every personal Google account on Tuesday, ComfyUI merged paid HeyGen Avatar 6 nodes the same day, and Suno's new Visuals feature lip-syncs clips to your lyrics. Talking video stopped being a premium product this week. What you write and how you voice it is now the part that costs effort.

New models

LongCat-Video-Avatar 1.6 arrives in ComfyUI, from three directions at once. Meituan published the model on November 13 (Hugging Face, created 2026-11-13) under the MIT license, which lets you use it commercially and modify it. It turns an image, a text description and an audio track into video of a person speaking, and the card claims gains in lip sync, facial and body coordination, and consistency over "videos exceeding 10 minutes." The second-day work is what makes it usable: PixelAI Labs released ComfyUI nodes and a ready workflow on November 18 (MIT), Comfy-Org posted repackaged files that evening, and kijai opened pull request #17831 to add a native LongCatAvatarToVideo node for one to three speakers. That PR is still open. The catch is hardware: PixelAI Labs tested only on a 32 GB RTX 5090 with 64 GB of system RAM, where a 3.8-second clip took 5.6 to 7.7 minutes at 480p, and says it has not tried smaller cards. A community demo runs at victor/LongCat-Video-Avatar-1.6 (running status not confirmed).

Genie 4 puts weather and new characters under your typing fingers inside a playable world. Google announced on November 18 that Genie 4 is available in Project Genie to eligible Google AI Ultra subscribers, with a waitlist for AI Pro. You can type new prompts mid-session to change the weather, add characters or redirect action, and it now generates audio in real time. Genie 3 stays in the picker for longer runs of up to three minutes. Google states no session length, resolution or export option for Genie 4, so treat it as a sketchbook for worlds, not a source of deliverable footage yet.

Video

Google Vids personal avatars are free for personal accounts. Since November 17, anyone with a personal Google account can build an avatar of themselves from a selfie and voice recording (captured by scanning a QR code), then type what it says inside a Vids project. Google restricts each avatar to the account holder's likeness. It does not state length limits or languages. For a creator who needs a talking explainer and owns no GPU, this is now the zero-cost floor.

HeyGen Avatar 6 lands in ComfyUI as paid nodes. ComfyUI merged partner nodes for HeyGen Avatar 6 on November 17, with a pricing update attached, so they bill through Comfy's API credits. Useful if you want HeyGen's avatars inside a graph that also does your backgrounds and grading.

Krea published how it made two short films. The Krea Originals post (November 17) is a case study, not a release, and it is covered under workflows below. Recraft also posted on November 17 that Seedance 2.5 Flash and Seedream 5.5 now run inside Recraft (listing titles only; posts not read).

Image

ComfyUI-H3-Image-EasyFix cleans the speckle out of MiniMax H3 Image. Plain text-to-image with H3 Image often leaves grainy, speckled texture on skin, fur and stone. This node pack, created at 00:16 UTC today under Apache 2.0 (18 stars), adds a Steering node that nudges the first step toward cleaner texture without touching your prompt, a Re-Noise Sampler that makes the model redraw fine detail on middle steps, and a size picker that goes past ComfyUI's 16-megapixel cap. The author says H3 Image is trained for 4 to 20 MP and that 64 MP worked in their tests. It needs a ComfyUI build with H3 Image support, which arrived in v0.45.0 on November 10. The node pack is Apache 2.0, but the model under it is not: the MiniMax-H3-Image Community License excludes the United States, the European Union, the United Kingdom and South Korea, and requires written permission above $1 million in annual revenue or for hosted services.

Audio and music

Suno Visuals turns a song into clips, lyric videos and cover art. Launched November 17 on web, iOS and Android, it makes clips up to 15 seconds that follow the song's structure and lip-sync to the lyrics, starting from Suno's own characters or an image you upload. Visual credits are sold separately from song credits, and the pricing page would not render costs for us. The catch for anyone posting elsewhere: Suno says it embeds watermarks in visuals before they leave the platform.

Two open song models now run through llama.cpp. ggml-org posted YuE2-3B-GGUF (created 2026-11-17 16:10 UTC) and MiniMax-Music3-GGUF (created 2026-11-17 12:44 UTC). Both write full songs with vocals and accompaniment from a text file with [Tags] and [Lyrics] sections, run with one llama-tts -hf ... -f song.txt --output out.wav command. YuE2's 8-bit file is 2.31 GB, a small download by music-model standards, but its license is CC-BY-NC-4.0, so no commercial use. MiniMax-Music3 is a 9-billion-parameter model whose 8-bit file is 9.13 GB, under a MiniMax community license we did not read. The YuE2 card is inconsistent about its own size (it says 2B while the name says 3B).

Writing

Two Claude Code skills turn footage and code into finished edits. claude-video-editing-skills (MIT, created November 17, 21 stars) packs 40 named edit styles, from jump-cut talking-head reels to podcast camera switching, that share one engine built on faster-whisper transcription, FFmpeg and a headless browser for graphics. product-walkthrough-video (MIT, created November 18, 10 stars) records a running app and writes a two-to-five-minute explainer from its repository. Both are early and young; expect rough edges.

Open and local

The local story today is avatars. LongCat 1.6 is MIT, runs offline, and got three ComfyUI paths in one day, but every published speed figure comes from a 32 GB card. A separate speed project claims roughly 3x to 3.8x faster rendering on the same card. Until someone posts a 16 or 24 GB result, assume this is a big-GPU or rented-GPU tool.

  • Pixelailabs/ComfyUI-LongCat-Avatar-PixelAILabs: photo plus voice line to a lip-synced talking video, with a drag-in workflow file. Created November 18, 5 stars (repo).
  • ckinpdx/longcat-avatar-fast: a launcher that renders a 3.2-second LongCat segment in 30 seconds at its six-step default on an RTX 5090, author's own figures. Created November 16, 3 stars, Windows-tested only (repo).
  • PedroMarinhoDev/ComfyUI-H3-Image-EasyFix: cleaner skin, fur and stone texture from MiniMax H3 Image, plus sizes above 16 MP. Created today, 18 stars (repo).
  • Viprasol-Tech/claude-video-editing-skills: 40 edit styles you call by name from Claude. Created November 17, 21 stars (repo).
  • toyxyz/toyxyz_RGBMocap: motion capture from ordinary video on Windows, retargeted to humanoid rigs with FBX and VMD export. On today's Trendshift list, about 1,850 stars, created October 26 (repo).
  • apekshik/anymachine: turns a generated 3D mesh into a working mechanical rig with hinges and legs, every control a 0-to-1 slider. On today's Trendshift list, about 380 stars, created November 4 (repo).
  • danemarguglio/sprite32: text to 32x32 transparent sprites, icons and emoji. Created November 16, 2 stars (repo).

Creative workflows

1. A talking presenter from one photo and one voice line, in ComfyUI. Workflow file: example_workflows/LongCat_Avatar_1.6_PixelAILabs.json.

The steps: Install the node pack through ComfyUI Manager by Git URL. Put the umt5_xxl_fp8_e4m3fn_scaled text encoder in models/text_encoders; the loaders fetch the rest (about 25 GB) on first run. Load a chest-up photo framed the way you want the video, load clean speech, write one paragraph naming who it is, where, and "static camera on a tripod," and run. Test with four seconds of audio first.

How it works: Whisper turns the voice into timing features for the mouth, the model generates the video in windows of about 3.7 seconds, and each window continues from the last until the audio ends. The video is exactly as long as the audio.

Why it is good: It is MIT-licensed, runs offline, and needs no face swap or editing for the README's examples. An optional Vocal Extract node pulls a voice out of a song or noisy clip.

Where it breaks: Every figure comes from a 32 GB card with 64 GB RAM, and each extra 3.2 seconds of audio adds 1.5 to 2.5 minutes. Hands wander unless you write "hands relaxed" or "hands out of frame." A bad take needs a new seed, not a new prompt.

2. Clean speckled texture out of H3 Image renders. Workflow file: workflows/H3 Image EasyFix - T2I.json.

The steps: Install the pack, load the workflow, and point its three loaders at your H3 Image model, text encoder and VAE files. It ships at 16 MP with Steering at 0.7 on the first step and the Re-Noise Sampler on its default RNmid plan with s_noise 0.95.

How it works: Steering pushes the first step toward what the model draws when handed an empty reference picture, which reads as cleaner texture. The re-noise steps take the model's guess at the finished image, add fresh noise back, and make it redraw fine detail instead of keeping the grain it settled on early.

Why it is good: It fixes the problem at render time, so you skip a denoise pass that would smear real detail. The author's side-by-side crops are at full resolution.

Where it breaks: H3 Image's license does not cover use in the US, EU, UK or South Korea, so most English-speaking readers cannot use the underlying model at all. The pack is one day old with one author's tests. It leaves edit mode (reference pictures) alone, and sizes above 64 MP are untested.

3. Approve stills, then let Claude write the shot prompts. Source: Krea Originals: ROCKET and RECESS.

The steps: Build one Nodes canvas per film. Make character sheets and location stills with Krea 2 Turbo and finish them first. Have Claude write shot-by-shot prompts and send them to the canvas through Krea's MCP connection, then generate each scene with Seedance 2.5, which runs up to 30 seconds and can include several cuts.

How it works: Locked stills fix look and character before you pay for motion; the language model handles the repetitive prompt writing per shot.

Why it is good: Video is the expensive step, and Krea says so plainly; this order spends video credits only on approved frames.

Where it breaks: It is a vendor's own showcase. Some shots worked only after the team wrote part of the prompt in Chinese, which tells you how much trial is hidden in the finished films.

Worth testing

  • Google Vids personal avatar: free with a personal Google account. Tradeoff: only your own likeness, and Google states no length or language limits.
  • victor/LongCat-Video-Avatar-1.6 Space: try the open model before you download 25 GB. Tradeoff: a community Space with an unknown queue and status.
  • YuE2-3B-GGUF: a full song from a lyrics file, on your own machine. Tradeoff: non-commercial license.
  • Suno Visuals: a lip-synced clip for a track you already made. Tradeoff: separate credits at prices we could not read, and watermarks off-platform.
  • H3 Image EasyFix: rerun a grainy H3 render with the shipped workflow. Tradeoff: one day old, and H3 Image's license excludes the US, EU, UK and South Korea.

What actually matters from today's signal

Lip-synced talking video now exists at three price tiers in the same week: free in Google Vids, free and local with LongCat in ComfyUI, and per-clip with HeyGen inside the same graph. When the same capability shows up free at the bottom and as an API at the top, it has stopped being a differentiator. A talking head that says something boring in perfect sync is still boring. The time you used to spend chasing a working lip-sync tool now belongs to the script and the voice take.

The open route is real but not cheap in hardware. Every LongCat number in circulation comes from a 32 GB card. If you own a 12 or 16 GB card, the honest choice today is Vids, a community Space or an hour of rented GPU, and waiting for kijai's native node to land, since in our reading built-in support tends to bring the memory-saving loaders the first community nodes lack.

The risk is the one nobody's release notes discuss: these tools make a convincing version of a real person from one photo. Google restricts avatars to your own face. LongCat's MIT license restricts nothing. If you make avatars for clients, get written consent for the likeness and the voice before you render a frame.


Source access notes: Hugging Face pages and API, Google, Suno, Krea and Recraft were read through a fetch agent; Hugging Face dates are from API createdAt. GitHub stars and creation dates come from the GitHub search API, READMEs from raw.githubusercontent.com with cache busting, and pull request details from the GitHub search API (the per-repository API was blocked this run). shields.io returned matching star counts for three repos via the fetch agent. OpenAI news and blog.adobe.com failed to load. Runway's news page carries no dates and its changelog was not read. BFL (newest Oct 15), Stability, Replicate (Nov 3), fal (Oct 20), ElevenLabs (Nov 16, an acquisition), Luma (Nov 16, engineering post), HeyGen's blog and Diffusers (v0.42.1, Oct 27) had nothing creative dated November 17 to 19. Midjourney searches returned nothing dated in the window. Adversarial fact-check pass ran (Sonnet subagent): no falsifications; it caught two star totals that had drifted by one (now rounded), an unsourced "ordinary laptop" claim for YuE2 (removed) and an opinion about native ComfyUI loaders stated as fact (reworded). It could not confirm PR #17831's one-to-three-speaker node or #17793's pricing update from search titles; both were confirmed earlier this run from the full PR bodies via the GitHub search API. Recraft's model-availability titles and the LongCat Space status remain unverified and are labeled as such. Suno's visual credit pricing page rendered no prices. Trendshift showed no daily deltas, so the two Trendshift entries are listed by presence on today's page, with GitHub star totals. apple/FastFrames-L was yesterday's lead and is not repeated. bothaggsigym1988/frameglide (64 stars, created Nov 18, no license, unverifiable author) was skipped. Civitai and blog.comfy.org were not attempted.

Articles from this briefing