Creative AI Briefing November 8, 2026 morning edition 11 min read

This week's useful releases work on the last mile of a project, so a creator can take an AI draft to the right size, codec and platform spec without leaving the free tools they already use.

RelayVSR ComfyUI Cosmos3-Edge Underdog Woof GIF Factory ray-sticker-shop PulsoIcon upscaling video-gen creative-workflows local-creative-ai audio-gen

Creative AI Briefing: Sunday, November 8, 2026

You can now take a soft 480p AI clip, upscale it inside ComfyUI with a free model that needs no prompt, and save the result as a ProRes file your editor will open without complaint. Both pieces reached ComfyUI this week, a day apart, and they set the pattern for a quiet weekend: the releases that matter are finishing tools. A sticker kit packs one photo into a platform-ready set, a Mac port runs NVIDIA's small video model on a laptop, and a 40 MB voice model reads scripts on a CPU. Nothing here invents a new kind of image. It gets the images you already have out the door.

Video

RelayVSR, a fast free video upscaler, is now built into ComfyUI. The RelayVSR paper (September 29) and its Apache 2.0 code describe a two-part upscaler: a large Wan 2.2 model redraws one frame in every 15, and a small "Flash" decoder fills in the frames between them from those keyframes and your original footage. The paper's own numbers, at 1080p on one A100 data-center card: 29.29 frames per second and 13.82 GB of memory, against 7.80 frames per second and 24.447 GB for FlashVSR-Tiny. On November 5, Comfy-Org posted repackaged x2 and x4 files (Apache 2.0, created 11:54 UTC, past 32,000 downloads in three days by the API count), and PR #17511 merged two core nodes, RelayVSRConditioning and VSRDecode, at 21:14 UTC. You need no text prompt; VSRDecode has a detail dial. The catch is timing: the merge came after v0.44.0 (November 4), so as of this morning RelayVSR lives in ComfyUI's development version, not a numbered release, and an open bug filed today reports a progress-bar crash in the new code. The LightX2V team's 2x version (created October 29, no license stated) claims 243 frames in 5.348 seconds on an H100 against 352.82 seconds for SeedVR2-3B on the same clip; that is the team's own benchmark. A Renderforest pull request reports a 5-second clip going from 480 to 720 in about 8 seconds instead of about 30 with SeedVR2, but names no graphics card. The original code drops audio ("audio is not copied") and runs on Linux with NVIDIA. A paper mode the authors clocked at 35.89 frames per second is not released yet. A free Space (created November 7, running on Hugging Face's shared free GPU tier) takes clips up to about 10 seconds at 2x or 4x; we could not confirm it renders.

ComfyUI now saves ProRes, with alpha. v0.44.0 (November 4) added ProRes output through PR #17463. In Save Video, pick mov and you get ProRes. In Create Video, a new chroma_subsampling setting decides the flavor: 444 writes ProRes 4444, 422 writes ProRes 422 HQ, per the code. The release's tests write and read back ProRes 4444 with a transparent alpha channel. Ask for 12-bit and you get 10-bit with a console warning. For an editor, this ends the H.264 detour: AI shots can go straight onto a Premiere or Resolve timeline as an intermediate codec. The encoder is FFmpeg's ProRes encoder (prores_ks), not Apple's own, and the files are large. Full recipe in the workflows section.

NVIDIA's small video model runs on a Mac. NetProjects posted an MLX port of Cosmos3-Edge on November 7 (7.6 GB full size, 5.0 GB in the 8-bit build). NVIDIA's Cosmos3-Edge is a 4-billion-parameter model that makes 256p or 480p video, 50 to 150 frames, from text or an image, under the OpenMDW 1.1 license, which allows commercial use. On an M5 Max with 128 GB, the porter reports the 8-bit build running NVIDIA's image-to-video example (832 x 480, 121 frames) in 113 seconds with an 18.8 GB memory peak. It is a world model built for robotics and driving, so expect plausible motion, not cinematic shots; NVIDIA's own card says outputs "should not be treated as physically accurate simulation." The port is unofficial ("Not affiliated with NVIDIA"), has no demo, and its GitHub install page failed to load for us.

Image and design

One photo becomes a consistent animated sticker set, packed to spec. imraywang/ray-sticker-shop (created November 5, MIT) is a Claude Code and Codex skill that turns one photo or character drawing into a full set of animated stickers for WeChat's sticker platform: 240 x 240 GIFs squeezed under 500 KB, plus cover, icon, banner and tip images, a spec check, listing copy and a client delivery pack. The README prices eight stickers plus five extras at about $0.70 at OpenAI's list rates, rerolls extra, and logs real spend to costs.jsonl. No API key? The agent hands you prompts to run in ChatGPT and assembles what you bring back. The README's warning is the useful part: use only likenesses you have rights to, and WeChat decides approval.

GIF Factory covers the rest of the GIF menu. WoYun520/gif-factory (created November 5, MIT, about 714 stars on shields.io) is a Codex skill with nine GIF style entries and 56 examples: pixel stickers, clay mascots, two-image "lenticular" flip cards, comic-print dynamic posters, plus brand films and short storybooks. Docs are in Chinese. Its gallery animates real public figures and famous anime characters, while its own SKILL.md says "Use only imagery the user provided or authorized," so follow the rule, not the gallery.

PulsoIcon animates 9,955 icons, one motion each. CityPulso/PulsoIcon (created November 4, ISC license, free for commercial use) gives every icon its own hover animation ("A bell rings... A trash can opens its lid") for React, Vue, Svelte, Jetpack Compose and SwiftUI, and drops in for lucide-react in one line. Try it in the playground. It is a coded asset library, no AI at runtime, which is exactly why a motion designer can trust it in production.

Audio and music

A 40 MB English voice model that runs on any CPU. ConwayResearch's Underdog Woof (created November 5) is a tiny text-to-speech model with two preset voices, Juno and Mira, that needs only NumPy. On one thread of an M5 Max the vendor measured Juno at 13.8x real time and Mira at 7.2x, with peak memory of 402 MB and 601 MB. That makes it a scratch-narration tool for animatics on a laptop with no graphics card. The license is non-commercial ("Commercial use needs a license"), the card lists no voice cloning, it is English only, and every score on the card is the vendor's.

Research worth one line: Hugging Face's daily papers on November 6 included TailSteer, aimed at the drift that piles up over many rounds of image edits, and WeLM-Image, a unified image model. Neither has a creator-ready release.

Open and local

The local story is the upscaler. Until this week, good AI video upscaling meant SeedVR2 (slow and memory hungry) or a paid partner node. RelayVSR's design, one big redraw every 15 frames and a small decoder for the rest, is why it should be lighter than redrawing every frame, but the paper's A100 figure is the only measured memory number so far. Run it on a short clip before you trust it on a feature.

kopperx/RelayVSR: free x4 video upscaler from one keyframe in 15, Apache 2.0, the source of everything below. 124 stars on shields.io; paper September 29 (repo).
Comfy-Org/RelayVSR: ComfyUI-ready x2 and x4 files. Created November 5, core nodes merged the same day (files).
Comfy-Org/ComfyUI: v0.44.0 adds ProRes 4444 and 422 HQ saving, plus a MiniMax H3 latent upscaler. Released November 4 (releases).
netprojects/Cosmos3-Edge-Omni-mlx-bf16: NVIDIA's small text- and image-to-video model on Apple Silicon. Created November 7 (card).
ConwayResearch/Underdog-Woof: 40.6 MB English voice model for CPU scratch narration, non-commercial. Created November 5 (card).
imraywang/ray-sticker-shop: one photo to a platform-ready animated sticker set for about $0.70. Created November 5; 371 stars (repo).
WoYun520/gif-factory: GIF stickers, flip cards and dynamic posters as a Codex skill. Created November 5; about 714 stars (repo).
CityPulso/PulsoIcon: 9,955 hand-animated icons for five frameworks. Created November 4 (repo).

Creative workflows

1. Upscale an AI clip with RelayVSR in ComfyUI, from PR #17511 and the relayvsr_x2_video.json linked in its description

The steps. Update ComfyUI to the development version (the nodes are not in v0.44.0). Put relayvsr_x2_lightx2v_int8_convrot.safetensors or relayvsr_x4_int8_convrot.safetensors in models/diffusion_models/ and the matching relayvsr_flash_vae_x2_fp32.safetensors or relayvsr_flash_vae_x4_fp32.safetensors in models/vae/ (files). Load the workflow file linked in the PR description, feed your low-resolution clip into RelayVSRConditioning, and decode with VSRDecode. Start with a 5-second clip at 2x. Raise detail only after you have seen the default.

How it works. RelayVSRConditioning reads your clip and builds the starting point, no text needed. The big model redraws a keyframe every 15 frames; VSRDecode uses those keyframes plus your original frames to finish every frame, which is why it is fast, and why the LightX2V card says it can "substantially alleviate" the flicker of frame-by-frame upscaling.

Why it is good. Free, Apache 2.0, inside the tool you already use, and fast enough to treat as a step rather than an overnight job.

Where it breaks. Development-version only, a fresh progress-bar bug, original code drops audio, no published memory figure for consumer cards, and the speed comparisons all come from the authors or one studio's untested PR. Fine textures (hair, foliage) are where any upscaler invents detail, so check them at 100 percent.

2. Save AI video from ComfyUI as a ProRes master, from PR #17463 in v0.44.0

The steps. Update to v0.44.0. For a finished clip, set Save Video's format to mov (ProRes is the only codec offered there). For frames you are assembling, use Create Video, set codec to prores, and set chroma_subsampling to 422 for a ProRes 422 HQ delivery file or 444 for ProRes 4444. Leave bit depth at 10. Import the .mov into your editor and check that it plays without conversion.

How it works. The new code picks the ProRes profile from your color setting: 4:2:2 gets 422 HQ, 4:4:4 gets 4444, and 4444 can carry transparency. Requests the encoder cannot honor (12-bit, for one) fall back to the nearest setting with a warning in the console.

Why it is good. Editors stop transcoding H.264 before they can cut, and 10-bit color survives grading better than an 8-bit MP4.

Where it breaks. It is FFmpeg's encoder rather than Apple's, so check any delivery spec that names an encoder. Files run far larger than MP4. ProRes cannot add quality the model did not make. How alpha reaches Create Video from a graph is not documented outside the tests.

3. One photo to a platform-ready sticker set, with ray-sticker-shop

The steps. git clone https://github.com/imraywang/ray-sticker-shop ~/.claude/skills/ray-sticker-shop, then pip install pillow and set OPENAI_API_KEY. Tell Claude Code or Codex to turn your photo into a sticker set. Review the output against the platform checks the skill runs, then hand over the delivery pack.

How it works. One reference image locks the character for the whole set, and the skill handles the chores that make ChatGPT's own GIF tool painful at scale: sizing, compression under 500 KB, extra assets, spec checks.

Why it is good. About $0.70 for a set, a cost log you can show a client, and a no-key fallback.

Where it breaks. Built around WeChat's rules and documented in Chinese. Character drift still happens; rerolls cost extra. Approval is the platform's call, and fan art gets rejected.

Worth testing

RelayVSR Space: free 2x or 4x upscale of a clip up to about 10 seconds. Tradeoff: shared free GPU queue, unverified by us, and a 10-second cap.
PulsoIcon playground: try animated icons before you install. Tradeoff: one house motion style across 9,955 icons, so brand-specific motion still needs a designer.
Underdog Woof: scratch narration on a CPU. Tradeoff: two voices, English only, non-commercial.
ComfyUI v0.44.0 ProRes save: a one-setting change for editors. Tradeoff: big files and a non-Apple encoder.

What actually matters from today's signal

The finish is getting cheaper faster than the generation. A year of model launches made it easy to get a decent 480p clip and hard to deliver it: upscaling meant waiting minutes per second of footage, and delivery meant an H.264 file your editor had to convert. This week the free stack closed both gaps, upscale and master, inside one program. For a working editor or motion designer, that changes the budget question from "which model" to "how much time does my last mile take," and that number just dropped.

The same thing is happening on the small end. The sticker and GIF skills are not better image models; they are packaging, compression and checking, the dull part that used to eat an afternoon. Expect more of this: agents that do delivery, not drawing.

The counter-signal is maturity. RelayVSR is three days old in ComfyUI, outside any numbered release, with a bug filed this morning and speed claims nobody independent has checked. Adobe MAX opens Tuesday (November 10 to 12) and will likely answer with paid, integrated versions of the same finishing steps. Test the free path now on throwaway footage, so you know what Adobe's version is worth to you when the prices appear.

Source access notes: Blocked or unreadable this run: openai.com/news (client error), blog.google (no dated items returned), deepmind.google blog (undated list), runway.com/news (JavaScript only), ideogram.ai/blog (undated). Midjourney, Kling and ByteDance Seed surfaced nothing dated in range via search. No posts in range from Stability (latest August 25), BFL (October 15), ElevenLabs (October 20), Replicate (November 3), fal (October 20), Pika (October 29) or HeyGen (August 11). Luma's November 5 and 6 posts were buyer's guides; Recraft's November 7 post was an explainer. Hugging Face's November 5 Phonon TTS post was covered in earlier briefings. Hugging Face and shields.io were read through a fetch tool; GitHub data through its API and raw files. Hacker News Algolia was blocked by the network proxy, so Show HN coverage is missing. The Cosmos MLX GitHub install page failed to load. All Hugging Face dates are API createdAt.

An adversarial fact-check pass ran on this draft and caught five issues, all fixed: the Cosmos3-Edge Mac timing belongs to the 8-bit build, not the full-size one; "no voice cloning" for Underdog Woof was unsupported (now "the card lists no voice cloning"); the RelayVSR workflow file is linked in the PR description, not committed; the ProRes PR merged November 1 and shipped in v0.44.0 on November 4, so the two ComfyUI changes arrived a day apart, not four; and an unsourced broadcast-rejection claim about FFmpeg's ProRes encoder was cut. Confirmed: v0.44.0 is still the newest numbered release as of this morning (GitHub releases API). Unverified after the pass: the 8-bit Cosmos repo size (5.0 GB, from an earlier card read), the TailSteer and WeLM-Image paper listings, Adobe MAX dates (from Adobe's November 5 post and an earlier briefing), and whether the RelayVSR Space renders. New models section omitted (no new model launches with creator availability in the window); Writing omitted. Civitai and blog.comfy.org skipped.