Creative AI Briefing November 11, 2026 morning edition 12 min read

Generation moved inside the editors creators already own, so the repair happens on the timeline or in the catalog, and the bill now comes from whichever model sits under the button.

Adobe MAX Adobe Premiere Lightroom Open Eyes Avid Ready to Play ReCaVSR ComfyUI-H3-FaceRefine-Plus ai-editing video-gen image-gen upscaling comfyui creative-workflows

Creative AI Briefing: Wednesday, November 11, 2026

You can now mask a boom mic in a Premiere timeline, track it, click Remove, and keep cutting. On the same Tuesday, Avid shipped a Media Composer extension that does the same job with Google's Veo, Lightroom started swapping closed eyes for open ones borrowed from the next frame in your burst, and Epic previewed an Unreal Engine with image and video generators sitting inside the editor. The pattern is consequential for anyone who edits: generation stopped being a separate website you export to and import from, and became a button in the app where the shot already lives. That saves round trips. It also means every repair now carries somebody's model, somebody's credits and somebody's terms, and the open community spent the same day shipping the repairs the hosted tools still miss: steadier faces and fast 4x upscaling on a gaming card.

New models

No major lab shipped a new open image, video or audio model in the window. Hugging Face's newest-model lists for video, image, speech and audio had nothing created after November 7 from a known lab, and the trending list's newest creative entries date to November 5. The model news came bundled inside Adobe's apps, below.

Image

Adobe MAX turned Firefly into a studio with other people's models inside it. Adobe's press release (November 10, Miami Beach) adds a Creative Agent that runs Adobe tools and third-party models from plain-language requests, with a Plan Mode that shows you the steps before it acts and a Model Selector per session. Firefly Image gets Edit Object, Edit Text, Relight and Retouch. The release names "recent image and video model additions" as Dreamina Seedance 2.5, Kling 4.0, GPT Image 2.6 Sunburst, GPT Image 2.6 Flare and Nano Banana 2.1. Availability is the fine print: "Edit Text, Retouch, Relight and Quick Actions are generally available in Firefly today," and "other new tools and capabilities are available in beta." No prices or credit costs appear; the release points to firefly.adobe.com/pricing. Firefly Design (one canvas for illustration, image editing, animation, 3D and print) and Firefly Graph (repeatable multi-model workflows) are beta; Firefly Builder is a private-beta waitlist.

Photoshop gets a slider for a generated style. Adobe's Creative Cloud post describes Generative Restyle: prompt an edit, convert the result to a layer, then use "the generated custom slider to precisely refine the intensity without needing to re-prompt." That turns a reroll into a dial. Relight adds new light sources you can move and recolor, and the Remove Tool now takes a subject's footprints, reflections, shadows and waves with it. The post gives no availability or credit cost for these; Photoshop's own what's-new page had not been updated at run time. Mockups, Camera Angles and Generative Recompose sit in the Photoshop beta app. TechCrunch reports two Topaz Labs models, Wonder 4 (upscale and denoise) and ImageHDR; we found no Adobe page confirming them, so treat that as secondary.

Lightroom fixes blinks with the person's own eyes. Lightroom desktop 10.0 (what's new) adds Open Eyes as an Early Access feature, per Adobe's help page (updated November 10). You pick the faces, then a Reference: "Nearby photos" (the default, frames shot around the same time), "Choose photos," or "None." Adobe's tip is to pick frames "with their eyes open and with a similar face angle and lighting." The same detector powers Assisted Culling, which labels frames likely open, likely closed, or can't tell. Adobe's blog calls it public beta while the help page says Early Access; credits are not stated. Full recipe below.

ArtStation now hides declared AI work by default. Under a policy announced November 9, uploaders must declare "primarily generative AI," "mixed," or "100% non-generative," declared AI work is hidden by default, and suspected AI is filtered at Strict, per Creative Bloq. If you post mixed-process work there, label it, or it may not be seen.

Video

Premiere 27 shipped, with Color Mode out of beta and generation in the timeline. ProVideo Coalition's Scott Simmons confirms the release (November 10) and warns that projects opened in 27 "will not open in Premiere 26," so install it beside your current version mid-job. Adobe's post says Generative Media brings Generate Video, Sound Effects, Soundscape and Music into the timeline using "Adobe Firefly and models from partners including Google, Kling, Runway and Luma." Generative Remove (beta) takes out "boom mics and logos or background distractions": mask, track, Remove. Adobe's help page for it exists (dated November 10) but its body did not render for us, so limits and credit costs are unconfirmed. Simmons also lists Generative Extend creating missing frames for transitions.

Avid put Veo inside Media Composer, and you pay Google, not Avid. Avid's Ready to Play, per ProVideo Coalition (November 10), uses "Google's Gemini and Veo AI models" for three jobs: Create Video from text or a still into a bin, Modify Video (a text-prompted change, such as removing a boom mic, that "generates a separate output and preserves the original media"), and Analyze Video, which writes descriptive metadata. It is "available now for Media Composer Ultimate and Enterprise users at no additional charge from Avid," but the models need "an eligible Google Cloud billing account." Budget it as a cloud bill with no published per-clip price in the coverage.

Unreal Engine 5.9 Preview puts generators in the editor. Epic calls it "the first Unreal Engine release built primarily with agentic workflows," per Creative Bloq. Unreal Canvas is a shared workspace with text, image, video and mesh generators driven by the generation service you choose, and Canvas Bridge exposes Movie Render Graph as agent tools. It is a preview built for Epic's AI MegaJam (November 12 to 15); Epic says some parts are prototypes.

Audio and music

Premiere's Generate Music and Generate Soundscape now live in the timeline, and on iPhone, Premiere mobile adds Universal and Warner licensed music, "available starting today in Premiere on iPhone for use in personal, non-commercial videos," per Adobe's mobile post. Read "non-commercial" before you score a client reel with it. Ableton Live 13 entered public beta the same day with, in MusicRadar's words, "no AI anywhere to be seen," a useful counterpoint for musicians weighing what to upgrade.

Open and local

The local story is repair, not generation. Two packs posted November 10 and 11 fix what generated video gets wrong after the fact: faces that wobble when re-drawn, and footage too soft to deliver. ReCaVSR is second-day work: the USTC weights went up on Hugging Face on October 3 (API createdAt), and a ComfyUI port made them usable on a gaming card on November 10. Its README reports 149 frames upscaled from 960 by 480 to 3840 by 1920 in 36 seconds on an RTX 4090 and 67 seconds on an RTX 3090 with the 6.27 GB int8 file, numbers measured by the port's author. Star counts below are GitHub API counts.

PedroMarinhoDev/ComfyUI-ReCaVSR: 4x video upscaling in one pass per block, streaming to disk so long clips cost time rather than memory. Created November 10, 16 stars (repo).
AnniCodex/ComfyUI-H3-FaceRefine-Plus: stops refined faces in MiniMax H3 video from sliding on the head, and pulls their color back to the source. Created November 11, 5 stars (repo).
Anionex/h3-image-mlx: MiniMax H3 Image generation and editing on an Apple Silicon Mac; the author reports 1024 by 1024 at a 2.67 GB peak in streaming mode and 4096 by 4096 at 6.86 GB. Created November 10, 5 stars (repo).
win520-1206/ComfyUI-H3-LowVRAM-Workflows: MiniMax H3 text-to-video, first-frame, first-and-last-frame and reference workflows measured on an 8 GB RTX 5070 laptop, at about 18 to 21 minutes per five-second clip with about 32 GB of system memory. Created November 10 (repo).
Moh4696/cinematic-skill: narrated archive documentaries drawn in code, no generated video. Created November 9, now 167 stars, up from 63 yesterday (repo).
timeqsz/motionflux: an agent skill that breaks a reference video down into composable motion pieces. Created November 10, 6 stars (repo).
vamsireddy4/claude-watch: lets a coding agent watch a video from a URL through sampled keyframes, useful for reference breakdowns. Created November 9, 6 stars (repo).

Every MiniMax H3 file behind the face, image and low-memory packs sits under MiniMax's community license, which excludes the US, EU, UK and Korea, as covered in earlier briefings.

Creative workflows

1. Stop a re-drawn face from wobbling (ComfyUI-H3-FaceRefine-Plus, example workflows in the repo, assets in assets/mara/)

The steps: Install Carasibana's ComfyUI-H3-FaceRefine first, then clone the Plus pack into custom_nodes. Copy assets/ into ComfyUI/input/h3plus_example/. Load the source shot with VHS Load Video FFmpeg, run Face Track Crop and the H3 refine at the standard 0.15 strength preset, replace the stock stitch with H3 Face Stitch Plus, wire H3 Face Flow Lock right after it (base frames plus stitched frames), and finish with H3 Face Tone Match against the FFmpeg-loaded source.

How it works: The author's diagnosis is that the wobble is not a tracking error. At a typical refine strength the model re-draws eyes and mouth corners one or two pixels off, differently on every frame, while hair and ears keep their original pixels. Flow Lock measures that slip with optical flow, only where the refine changed the frame, and pushes the new features back into place.

Why it is good: Flow Lock alone cut wobble from 0.73 to 0.19 percent of face height on the author's 206-pixel-face test shot. On a small-face shot, the pack measured 0.42 percent wobble against 0.70 for the official refine, kept roughly twice its sharpness, and barely shifted color. It also documents a color trap: ComfyUI's common video loaders decode BT.709 footage with the wrong matrix, which darkens skin and turns lips purple before any AI touches it.

Where it breaks: The author's tests are one clip per case on one RTX 5090, with the author's own metrics. It takes about 2.3 to 2.9 times as long as the official refine. The standard preset scores lower on likeness to the character references than the official refine. It only works on MiniMax H3 footage, under H3's license.

2. Upscale a draft 4x on a 24 GB card (ComfyUI-ReCaVSR, example_workflows/recavsr_video_4x.json)

The steps: Install "ReCaVSR video super-resolution" from ComfyUI Manager (or comfy node install comfyui-recavsr). Download prompt.safetensors and flashdecoder.safetensors from kopper/ReCaVSR plus either the 10.6 GB bf16 transformer or the 6.27 GB int8 file, into ComfyUI/models/recavsr. Open Templates, ComfyUI-ReCaVSR, recavsr_video_4x, load the clip and queue. The node keeps the input's audio and frame rate.

How it works: It upscales 21 frames, then eight at a time, each block in a single pass, reusing what it learned from the previous block, and decodes each block to frames as soon as it finishes.

Why it is good: It turns 4K finishing into minutes. The README's 1440 by 1440 portrait test ran 180 frames in 15 to 40 seconds depending on card and settings, and the badge claims 8 GB of memory is enough at 1440p.

Where it breaks: The fast decoder is softer; the Wan 2.2 decoder option draws finer hair and skin but takes 25 to 40 times as long. The model was trained for 4x; the reference documents 2x and 1.5x, but those are not what it learned. Weights carry their own "recavsr-weights" license from the upstream project, which we did not read, so check it before client work. No hosted demo exists.

3. Fix a blink from your own burst in Lightroom (Adobe help)

The steps: Open the photo in Detail view, choose Remove, then the Eye Correction icon, then Open Eyes. Tick each person under "Apply to these faces," leave Reference on "Nearby photos" or pick frames yourself, and select Generate.

Why it is good: The open eyes come from photos of that person you shot, not an invented face, which is the right default for weddings and events.

Where it breaks: Desktop and Early Access only, with no stated credit cost. A burst with a turned head or different light gives the reference nothing to match, and "None" falls back to invented eyes.

Worth testing

Firefly Quick Actions and Edit Text, generally available today. Tradeoff: partner models and credits are priced on a page the release only links to, so check what each action costs before batching.
Premiere 27 Generative Remove (beta) on a boom mic in a locked-off shot. Tradeoff: limits and credits are unconfirmed, and projects saved in 27 will not reopen in 26.
Lightroom Open Eyes on a group photo with a good burst. Tradeoff: Early Access, desktop only, and the faces it fixes still need a careful look at 100 percent.
ReCaVSR int8 on a soft AI draft you already like. Tradeoff: needs a recent ComfyUI (v0.44 or later) and an NVIDIA card, and the fast decoder trades texture for speed.
Avid Ready to Play if you cut in Media Composer Ultimate. Tradeoff: a separate Google Cloud bill with no per-clip price in the coverage.

What actually matters from today's signal

The editor won. Adobe, Avid and Epic all made the same bet on the same day: the creator should never leave the timeline, the catalog or the level to generate something. For working editors that is a real time saving, because the slowest part of AI work was the export, upload, download and relink loop, not the render. Avid's Modify Video producing a separate output and preserving the original media is the right design, and it should be the bar for every in-app repair.

The money moved with it. Each button now hides a model choice and a meter: Adobe's credits on a pricing page the release does not quote, Avid's Google Cloud bill, Unreal's chosen generation service. Before you rebuild a workflow around any of these, run a week of real jobs and read the invoice.

The counter-signal comes from the open side. The most specific, measured work of the day was two hobbyist ComfyUI packs fixing faces and resolution, with tables of numbers, failure cases and decode bugs written down. Adobe published none of that for Generative Remove. If a tool will not tell you where it breaks, assume you will find out on deadline.

Source access notes: Pages were read through a separate web reader; GitHub data came from the GitHub search API and raw.githubusercontent.com (shields.io was unreachable, so stars are GitHub API counts). Hugging Face dates come from the API createdAt (kopper/ReCaVSR 2026-10-03T09:20:45Z; PedroMarinhoDev/ReCaVSR-ComfyUI 2026-11-10T12:24:53Z). Hugging Face's createdAt-sorted lists by pipeline (text-to-video, image-to-video, text-to-image, text-to-speech, text-to-audio) appeared stale, with newest entries between August 25 and October 27, so discovery leaned on the trending list and GitHub. news.adobe.com's index served a cached September snapshot; the Firefly release was read directly. Premiere's Generative Remove help page and Photoshop's what's-new page did not show November content. Topaz Wonder 4 and ImageHDR are TechCrunch-only. OpenAI news failed to load; Midjourney returned no content; Runway, blog.google's AI category and blog.comfy.org showed no dated items; Ideogram and PetaPixel carry no dates. BFL (October 15), Stability (August 25), ElevenLabs (October 20), Suno (November 4), Replicate (November 3), fal (October 20), Krea (November 5), Pika (October 29), HeyGen (August 11) and ComfyUI releases (v0.43.0, October 27) had nothing in the window. Luma's nine November 8 posts and Recraft's six November 10 posts are comparison and guide articles, not launches; Recraft's Gemini Omni 1.2 Flash integration repeats a model covered November 6. Microsoft's MAI-Image-2.6 dates to September 4. New models section reduced to a note: nothing qualifying shipped in the window. Fact-check pass: an adversarial pass checked the draft against primary sources and caught seven problems, all fixed: the Photoshop Remove Tool list misquoted ("waves," not "wake"); the FaceRefine Plus 0.73 to 0.19 wobble figure framed as a comparison with the official refine when it measures Flow Lock alone; the ReCaVSR portrait time range understated (15 to 40 seconds, not 15 to 33); "trained for 4x only" overstated; a paraphrase of Creative Bloq's Unreal coverage set in quote marks; ArtStation's default filter described too narrowly and its date stated as an effective date rather than an announcement; and Ableton Live 13's beta wrongly called "limited." The Lightroom 10.0 version number comes from Adobe's what's-new page, not the Open Eyes help page.