Creative AI Briefing: Tuesday, October 20, 2026

You can now hand a finished video clip to a model that watches it, keeps the dialogue, strips out the music you never asked for, and writes a full sound pass of footsteps, doors and room tone, all without touching a single frame. That model is Mirelo SFX 2.0, it launched last Wednesday with almost no coverage, and it is one of several releases this week that do the same kind of job. A community add-on stops MiniMax H3 from burning subtitles into talking scenes. Another teaches it the push-ins and orbits it fumbles. A third pulls moire out of photos of screens. None of them make anything from scratch. They fix the one part a generator got wrong and leave the rest alone, which means the cheapest render is no longer a reroll. It is a repair.

New models

Mirelo SFX 2.0 builds a sound design from the picture itself. Mirelo's launch post is dated October 14, and Runway added it to its agent connector on October 15 (Runway changelog). What you can make: a synced bed of sound effects, foley (the small recorded sounds of hands, cloth and footsteps) and ambience for any clip up to 10 minutes, generated from the video rather than from a text prompt. The post describes it as a model "that lets you change a video's audio without regenerating its original video." It keeps existing speech and removes non-dialogue audio such as music the video model added on its own. You can still steer it with a text prompt, and on WaveSpeed's API page a negative prompt (sounds to avoid) and a seed as well. Where it runs: Mirelo's editor and playground, its API and agent connector, plugins for Premiere Pro and DaVinci Resolve, fal, Runway, and WaveSpeed. Price: WaveSpeed lists $0.016 per requested second, so a 60-second pass is $0.96, and you pay for the full duration you request even if the clip is shorter. Mirelo's own pricing page gives a free tier of 5,000 credits a month (it estimates about 8 minutes), with watermarked exports and projects deleted after a week; subscription prices did not load. The catch: you get one mixed track, not separate stems (stems are "coming soon"). The preference numbers (81% over Sony Woosh, 49% over MiniMax H3 Max's native audio, and so on) are Mirelo's own blind test of 25 crowd-sourced raters on 30 videos. Mirelo's terms page did not render, so check commercial-use rights before client delivery. Free demo: the free tier on Mirelo's plans.

Image

DemoireEdit removes the rainbow ripples from photos of screens and fabrics. umu/DemoireEdit (created October 18) is a restoration add-on for Qwen-Image-Edit-2509 that targets moire, the wavy interference pattern you get photographing a monitor, a printed halftone or a fine weave. It runs through the project's own command-line tool, not a standard Diffusers pipeline. The catch is honest and stated: validation so far is structural checks plus one 1536 by 1536 crop, full 4K inference has not been established, and edits can shift colour and fine detail. The code is Apache 2.0, but the card says that does not automatically license the add-on file, and no separate licence is given. Treat it as personal-use until that is clarified.

MOSS-Image learns to colour manga pages. sorryhyun/moss-image-colorize (created October 19) is an Apache 2.0 add-on that takes a black-and-white page and returns a colour version that, per the card, "keeps the page's line work, faces and lettering." Colour tags in the prompt steer the hues; without them it picks its own. It trained in about 71 minutes on one consumer card from 1,000 synthetic page pairs. The catch: the author calls it a "research probe," it is tested only on MOSS-Image-Base, and the ComfyUI node it needs is linked but "not published yet," so most people cannot run it today.

Research to watch: the MT Forcing paper (published October 18, on today's Hugging Face papers list) targets drift across many rounds of conversational image edits, so the tenth edit still looks like the first. Code and weights are linked from the paper page; nothing is hosted for creators yet.

Video

Stop MiniMax H3 from burning subtitles into talking scenes. Justlora/MinimaxH3-Subtitle-Control-LoRA (created October 19, Apache 2.0 add-on on the MiniMax H3 Community License base) targets "subtitle leakage," where H3 prints the spoken line on screen as captions. The fix pairs a 65.6 MB add-on with a prompt rule: wrap dialogue in <d>[Language] text</d> tags and use no quotation marks inside them, because quotes tend to trigger on-screen text. The author calls it "an experimental exploration release," says it reduces leakage in "most scenarios" and to "expect occasional bugs," and shows one 12-second example with no measurements. The prompt rule costs nothing to try even without the add-on.

A camera-move add-on for H3. robertbschuette/minimax-h3-Camera-Motion-lora (created October 19, Apache 2.0) was trained on newly collected footage of moves the base model handles poorly. Start the prompt with "camera motion." The card rates push-in, pull-back and handheld tracking strongest, orbit, aerial and crane as stable, and pans as weak. Run it at 0.8 to 1.0 strength; above 1.2 the picture can destabilise. A 55-prompt library ships with it. Dataset size is not stated.

H3 on one Blackwell card. FastVideo/FastH3-QAD-v2 (created October 17) is an 8-step, 4-bit build of MiniMax H3 with sound, aimed at a single RTX 5090, RTX PRO 6000 or DGX Spark. It publishes no speed figures, measures only how closely it matches the full model, and lists its licence as "other," so read the repository before using it commercially. Intel's AutoRound build (created October 19) still needs two H200s and is, by its own card, "not faster than bf16 on Hopper." Data-centre only.

Audio and music

Runway's agent connector picked up sound and cutouts. On October 15 Runway added Mirelo SFX 2.0 and BiRefNet v2 background removal to its MCP connector (the plug that lets an AI assistant drive Runway) on paid plans (changelog). An assistant can now take a clip plus a sound description and return a separate audio track, or cut out a product shot, inside the same session that generated the video. No per-use credit cost is listed in the changelog.

Open and local

The local story in plain terms: nothing new and large landed for your own machine this weekend. What did land is small, cheap fixes for the open models you already have, and two agent toolkits that turn a coding assistant into a production assistant. Star counts are GitHub totals read from the GitHub API today; shields.io returned nothing this run.

• vercel-labs/fx0: describe a page and an agent writes it as one HTML file, with ten starter design styles and version history per turn. Why now: created October 19, 209 stars, free trial at fx0.vercel.dev (repo).
• UnderleveledBuilder/awesome-short-drama-skills: a CC0 index of 83 assistant skills for AI short-drama scripts, storyboards and video prompts, sorted by production stage. Why now: on today's Trendshift list, 385 stars; it says its combos are "not yet tested" (repo).
• HeyChengC/wallstreet-explainer: a Claude Code skill for finance-and-tech explainer videos (the sample runs 77 seconds), from footage pulls and news screenshots to voiceover, cutouts and mix. Why now: created October 18, 267 stars, MIT code and docs (third-party footage excluded); Chinese docs (repo).
• Justlora/MinimaxH3-Subtitle-Control-LoRA: keeps spoken lines out of the picture in H3 talking scenes. Why now: posted October 19 (model).
• robertbschuette/minimax-h3-Camera-Motion-lora: push-ins, orbits and handheld moves on cue. Why now: posted October 19 (model).
• umu/DemoireEdit: moire removal for photos of screens and fabrics. Why now: posted October 18 (model).
• sorryhyun/moss-image-colorize: manga page colouring on MOSS-Image. Why now: posted October 19, node still unpublished (model).

Creative workflows

1. Replace a generated soundtrack without re-rendering: Mirelo SFX 2.0 launch post and WaveSpeed API page

The steps. Export your clip with its existing audio. Upload it to Mirelo's playground (free tier) or call the API. Leave speech preservation on (the API default). In the prompt, describe the sound world in plain terms ("wet street at night, distant traffic, heavy boots on gravel"); in the negative prompt, name what to keep out ("music, crowd chatter"). Set the requested duration to the clip length exactly, because you pay for what you request. Download the audio track (the API offers WAV, MP3, AAC or FLAC), drop it under the picture in your editor, and set a seed when you find a version you like.

How it works. The model reads the frames and places sounds where the action is, then keeps any dialogue it detects and drops other audio.

Why it is good. Video models now ship sound, but it is often wrong in small ways: music nobody asked for, footsteps that miss. This fixes the soundtrack for about a cent and a half per second without risking a picture you already approved.

Where it breaks. One mixed track, no stems yet, so you cannot pull the rain down without regenerating. The preference scores are Mirelo's own. Free exports carry a watermark. Commercial terms were not readable this run.

2. Talking scenes without burned-in captions in MiniMax H3: Subtitle-Control card

The steps. Put contextir-hybird-v0.1.safetensors in your ComfyUI LoRA folder and load it at 1.0. Write each spoken line as <d>[English] I told you not to open it</d>, with no quotation marks anywhere inside the tags. Render at 8 steps with the card's suggested setup (Singularity plus vdn-dmd fused inference, or the lavarryv4turbo scheduler).

How it works. Quotation marks read to the model like printed text. Removing them and marking dialogue with tags tells it the words are sound, and the add-on reinforces that.

Why it is good. Burned-in subtitles ruin otherwise usable dialogue takes, and you cannot mask them out cleanly afterward.

Where it breaks. "Most scenarios" is the author's own claim, backed by one example. The card's layer description and its file name do not match its repo name, a sign of a rushed release. The base model's community licence still applies.

3. Direct the camera in H3 with a fixed caption order: Camera-Motion card

The steps. Load the 3000-step file at 0.8 to 1.0. Write prompts in the training order: "camera motion," subject, scene, camera movement, lighting and style, quality. Add speed and direction words ("slow," "aggressive," "from left"). For pans, combine with another move.

How it works. The add-on learned moves from captions written in that exact structure, so matching the structure matches what it learned.

Why it is good. Rerolling for a camera move wastes renders; a vocabulary that works the first time saves them.

Where it breaks. Pans are weak, strength above 1.2 destabilises the image, and the dataset size is undisclosed.

Worth testing

• Mirelo SFX 2.0 free tier: run one of your own AI clips through it and compare with the video model's native audio. Tradeoff: watermarked free exports and one mixed track.
• The <d> dialogue rule in H3: try it in your existing H3 workflow before downloading anything. Tradeoff: without the add-on it may not be enough, and the add-on's evidence is one clip.
• fx0: five free prompts for a landing-page draft as a single HTML file. Tradeoff: designs cap at 512 KB with no external assets, and real use bills your own Vercel team.
• Camera-Motion add-on: one push-in test against the base model at the same seed. Tradeoff: pans stay weak.

What actually matters from today's signal

The interesting releases this week are not generators. They are repair tools, each aimed at one failure that generators keep shipping. That changes how you should budget. A bad soundtrack used to mean rerolling the whole clip and losing the picture you liked. Now it means a dollar's worth of sound pass. A subtitle leak used to mean another render; now it is a prompt rule. Plan your pipeline as generate once, repair in layers, and the money goes to the frames that matter.

It also moves the finish back toward the edit suite. Mirelo ships Premiere and Resolve plugins, so the sound pass happens inside the editor. Runway is wiring repair tools into the same agent session that makes the clip. The editor, not the prompt box, becomes where the work is judged.

The risk is evidence. Every repair tool today rests on its maker's own test: Mirelo's 25 raters, one 12-second subtitle example, one demoire crop, a colouring probe with an unpublished node. Repairs can also change the parts you meant to keep: DemoireEdit's own card says "Editing can alter color and fine detail." Check every repaired file against the original before it goes to a client.

---

Source access notes: Bucket A checked with no new creative launches in the window: OpenAI news (newest October 19, a coding post; October 15 education items), blog.google AI page (served an undated, apparently stale listing), DeepMind news (October items are Gemini 4 Argon and science, no media models), Adobe blog (newest October 14, fraud policy), BFL blog (newest October 15, Klein, covered), Stability news (newest August 25), Luma blog (October 14 comparison guides; October 8 Claude Motion, covered), ElevenLabs blog (newest September 30), Suno blog (newest October 13). Runway: runwayml.com/news rendered empty; runway.com/changelog read directly (newest October 15). Kling 4.0 (October 15 to 16) was covered October 18 and is not repeated. HeyGen Voice (October 9) is outside the window. Midjourney: nothing new via search. Bucket B: fal blog newest September 17; fal and Replicate model pages are undated and were used for discovery only. Hugging Face listing endpoints returned inconsistent results (some appear cached at August dates), so discovery leaned on the text-to-video and image-to-image createdAt-sorted lists; every Hugging Face date above is API createdAt. MAGI-2-preview's FP8 conversion (October 19) was omitted because the base model dates to August 3 and needs data-centre GPUs. Bucket C: ComfyUI GitHub releases page served stale August data; blog.comfy.org and civitai skipped. Bucket D: PetaPixel and Creative Bloq had no creative-AI launches (PetaPixel's October 19 false-memory study item failed to load); cdm.link blocked by robots. HN Algolia served August data and was not used. Trendshift used for discovery; its numbers are unlabeled, so star figures here are GitHub API totals (WebFetch of GitHub HTML pages showed stale counts of 1, 5 and 21). Mirelo's terms page did not render; fal's sfx-v2 page lists no price. Mirelo's launch tweet on X was not readable (robots); the blog post is the primary source; WaveSpeed's page is the source for API parameters and the per-second price. Adversarial fact-check pass ran (Sonnet subagent, all load-bearing claims checked against primary sources). It caught: MT Forcing does link code and weights (corrected); the short-drama index counts 83 skills, not 82; wallstreet-explainer's 77 seconds is one sample, not a spec, and stars moved to 267; the subtitle add-on's wording is "most scenarios" (misquote fixed); an invented "untouched" quote and a banned word; unsupported claims that Mirelo works on any video and exports WAV from the blog (reworded and sourced to WaveSpeed); the free-tier link now points to the pricing page; Intel quote extended to "on Hopper".