Beat

open-weights

93 pieces filed under open-weights, newest first.

AI Storyboard From One Frame: The Free Next-Scene Add-On That Treats Your Prompt as a Cut The Next-Scene add-on for Qwen-Image 2.1 is useful because it was trained on the gap between consecutive film frames, so it reads a direction as a cut to the next shot rather than an edit,…
Briefing · October 10, 2026 · morning Fast drafts now reach a laptop within a day of release, and in the same week anyone gained a free way to check whether Google or OpenAI tools made a picture, so speed is cheap and…
Iris-3B Is a Free Image Model Whose Makers Admit Their Big Idea Didn't Pay Off Iris-3B is worth a creator's attention less for its pictures than for its honesty, because a lab that publishes where its own idea failed hands you a map of which claims to trust, and the…
Briefing · October 9, 2026 · morning Drafts now cost pennies or nothing, so a creator's money and attention move to the single finishing pass, and the tools worth adopting are the ones that hand a draft to a finisher without a…
One Photo, 72 Camera Positions: Building an AI Turnaround Sheet With the Qwen-Image-2.1 Multiple Angles Add-On The Multiple Angles add-on works because it treats camera position as a short fixed vocabulary learned from 3D renders, which makes it a reliable way to rough out front, side and back views…
Briefing · October 8, 2026 · morning Control got cheaper this week on both sides of the fence, with Google halving the price of a reference-heavy image model while open add-ons let a single photo or a 24 GB card do work that…
Kandinsky 6.0 Video: The Open AI Video Model That Makes Its Own Sound, Under a License You Can Actually Sell Kandinsky 6.0 Video matters less for its pictures than for its paperwork, because unlike the other open picture-plus-sound models it carries no revenue cap or territory clause, and the…
Briefing · October 7, 2026 · morning Open releases now cover whole paid jobs end to end, from a talking clip with its own sound to a Premiere-style timeline, so the price a creator pays has shifted from subscriptions to…
Briefing · October 6, 2026 · morning The weekend's open releases were finishing tools rather than generators, so cutting out a subject, finding a shot in a drive full of footage and upscaling a still now run on your own…
Briefing · October 5, 2026 · morning Over the weekend the community moved creative tools onto hardware people already own, so a layered PSD, an Ideogram 4 poster, a sung demo and a cloned voice can now be made on your own…
Real Pixel Art From an AI Video Model: Draw Big, Snap to the Grid, Lock the Palette AI video models fake pixel art because nothing forces them onto a grid or a fixed set of colors, and two new MiniMax H3 add-ons plus a three-command cleanup pass (shrink, one palette for…
LAION's Humaneness Voice Add-ons: 40 Emotions and 500 Voices, Each a One-Megabyte File LAION's 557 voice add-ons make a mood or a voice something you attach like a plug-in, and the 40 emotion files are the useful, low-risk part for creators, while the 500 voice-identity files…
Briefing · October 4, 2026 · morning On a quiet lab weekend the community shipped small add-on files that each teach a big model one job, so a voice, a mood, a pixel grid or an outfit swap now arrives as a download you attach…
ldraw-nova: Design a LEGO Model by Describing It to an AI Agent, Then Open It in Blender Ldraw-nova shows that an AI agent which writes an editable parts file, rather than generating a picture of bricks, gives designers something they can revise, render and price, but its own…
Audio Sliders: Change the Mood of an AI Music Cue Without Rerolling It Audio Sliders turns AI music from a slot machine into an instrument by letting you move one quality of the same clip (mood, energy, ensemble) instead of rewriting the prompt, which matters…
Briefing · October 3, 2026 · morning The community spent a quiet lab week turning generations into things you adjust, so a music cue gets a happier version without a reroll, a LEGO idea comes back as an editable parts file,…
Briefing · September 30, 2026 · morning Voice models now take stage directions line by line, so a writer with a script can cast, direct and re-take a performance without a booth, and the craft moves from recording to directing.
EditVoice Fixes a Wrong Word in Your Voiceover Without a Re-Record EditVoice matters for voiceover and podcast work because it lets the corrected line be a different length from the original, which is the thing that makes patched speech sound rushed, and…
AnyBokeh Changes a Photo's Focus and Aperture After You Shot It AnyBokeh treats the blur already in your photograph as information about the lens rather than damage to be removed, which is why it can move the focus point and change the f-number without…
Briefing · September 28, 2026 · morning The weekend's releases all reach backwards into a picture you already have and put the camera decisions back in your hands, focus, aperture, angle and framing, which means fewer reshoots…
A 10 MB File Lets You Watch an AI Image Form, So You Can Kill a Bad One at Step Four Preview decoders are not new, but publishing the fifteen-minute recipe to train one yourself changes who has to wait for the next model to be supported.
Your Local Music Model Is Not Broken. The Wrapper Around It Is. A music model that sang gibberish on a Mac turned out to be receiving a message in the wrong format, and the two-line fix is a pattern worth recognising because it fails silently and sounds…
Briefing · September 27, 2026 · morning Nobody shipped a model this weekend, so the gains came from the two ends of the loop instead, better words going in and faster seeing coming out, and both are free.
The Six-Step Image Model Is Real. The Default Way You'll Install It Throws Away a Third of It. A free hosted demo proves the six-step version of Qwen-Image-2.1 holds up against the forty-step original, but the standard way of adding it to ComfyUI discards a measurable chunk of the…
Briefing · September 26, 2026 · morning The version of last week's model you can actually run this weekend was not made by the lab that trained it, it was made by somebody else three days later, and it is between five and eleven…
How to Get a Speaker-Labelled Transcript From Two Models That Never Talk to Each Other The useful thing about NVIDIA's new speaker model is not its benchmark ranking but its size and its licence, and the pipeline it enables joins two separate models on nothing more than a…
Ming-Image-0.1-Design-Layer Pulls a Flat Design Back Into Editable Layers The interesting half of this release is the model that takes a finished design apart rather than the one that makes a new one, and the MIT licence on it is what makes the output yours to…
Briefing · September 25, 2026 · morning The most useful open releases this week take a finished file apart rather than making a new one, so the thing sitting in your folder is turning back into editable source under licences that…
Briefing · September 24, 2026 · morning Two hosted video models got retirement dates this week and one of them stops generating today, so the durable question for a creator is no longer which model looks best but which of your…
You Can Now Direct an AI Video Model With a Drawing The new control pack for MiniMax H3 adds three input types that let you steer a video model from a drawing instead of from footage you already shot, which moves it from a post tool to a…
Briefing · September 23, 2026 · morning Every useful creative release of the last 48 hours was an add-on to somebody else's model, and the gap between a model landing and its steering wheel arriving has closed to about a day,…
WorldCrafter Turns One Photograph Into a Space You Can Walk Through WorldCrafter replaces prompt adjectives with a written shot list of camera moves and remembers what it already drew, which is the real shift, but the word Fast in its name refers to step…
Briefing · September 22, 2026 · morning Three separate releases this weekend moved a camera through footage that was already shot, and none of them did it with a prompt sentence, which means the control surface for AI video is…
Your Picture Model Is Not What Fills Your Graphics Card. The Prompt Reader Is. Two independent measurements published on the same weekend found that the model reading your prompt is the largest thing resident in a local generation pipeline, and shrinking that file…
Qwen-Image-2.1 Makes Transparent PNGs Without a Cutout Step. Then the Licence Shows Up. Qwen-Image-2.1 generates cutouts with a real transparency channel instead of making you key them out afterwards, which removes a step that has survived every previous generation of image…
Briefing · September 21, 2026 · morning Two people working on different media over the same weekend published the same finding, which is that the text encoder and not the generation model is what fills your graphics card, so…
Someone Finally Tested Six Ways to Train an Image Add-On. Five of Them Load Nothing. Five of the six common formats for saving a trained image add-on are unreadable by the standard loader, which hands you the plain model without an error, so the first thing to check when a…
Chain Three AI Video Shots Into One Take, and the Drift Fix Everyone Gets Backwards Chained AI video shots drift because each shot regenerates from the previous shot's own output, and the per-shot colour and brightness corrections most people reach for cannot fix it,…
Briefing · September 20, 2026 · morning Two independent publications yesterday did the thing the add-on boom skipped, which is measure whether the add-on is doing anything at all, and one of them found that five of six popular…
FLUX.2 Klein Draws in One Pass Now. The Free File That Did It Is 735 MB. The most useful image release of the week is an add-on rather than a model, and the reason it matters to a working creative is not the speed but the licence attached to it.
Briefing · September 19, 2026 · morning The files that moved the most ground this week were not models at all but add-ons that ride on top of models creators already have, which changes how much bandwidth, disk and attention a…
VDN-H3 Makes Fourteen Seconds of Video in Nine. Then the Licence Tells You Where You Can't Run It. Video generation has crossed playback speed, which changes the creative loop from submit-and-wait to look-and-revise, but the fast open version carries a licence that excludes most of the…
Your LTX-2.5 Download Is Only a Third of the Bill The compressed LTX-2.5 video file everyone links to is one of four pieces you need, and the piece nobody mentions is the same size again, which is why sizing the download by your card still…
Briefing · September 18, 2026 · morning Video generation just got faster than video playback, which turns a clip from something you submit and wait for into something you sit and iterate on, and the streaming plumbing to steer it…
Linum Threw Away the Compressor Inside Their Image Model. The Detail Got Worse First. Removing the compression stage from an image model made fine detail worse, and the fix was to split the model into one half that plans the composition and another that renders it.
Ideogram 4 Does Not Want Your Sentence. It Wants a Layout Spec. Ideogram 4 was trained only on structured captions with coordinates and hex colours, so typing a sentence at it is not prompting badly, it is using the wrong interface entirely.
Briefing · September 17, 2026 · morning Nothing new arrived in the models today, but three separate groups deleted a required piece of the machinery around them, and every deletion lands on your card as free memory or on your…
A 28 MB File Separates the Light From the Paint in Any Photograph ShadeNet-2 pulls a relighting stack out of one photo on a CPU for free, and it works because its lighting layer is trained on nothing but the requirement that colour times light has to…
Meridian Re-Shoots Your Footage From a Camera That Was Never There Meridian's real contribution is a quarter-second preview that shows you the holes in your shot before you render it, and the thing standing between most editors and that preview is a…
The MiniMax H3 Three-Step Adapter Everyone Is Downloading, and the Part That Got Left Behind The three-step TaoMate adapter really does make MiniMax H3 fast on one card, but the streaming runtime it was built for is the part worth wanting, and that part did not come along.
Briefing · September 15, 2026 · morning The week's capabilities are being cut out of vendor systems rather than shipped as products, so the thing you could not run yesterday is now a single file that still carries the original…
Paint Six Flat Colours, Get Fire: The Drawing Is the Real Prompt In this kind of tool the painted control frames carry the art direction and the text prompt carries almost nothing, and the team published enough of the recipe that the method transfers to…
LTX 2.5 Can Now Edit Your Footage. Nine Tools Exist, and Seven Went Missing. LTX 2.5's add-on editing tools turned it into a practical way to change footage you already shot, and the tools that did not get carried over are what should decide whether you upgrade.
LTX 2.5 Now Renders Ten Real Seconds on a 24 GB Card, and the File That Does It Advertises Its Own Weakness For local video the number that decides everything is whether the whole model fits on the card, which means a less faithful file that stays resident beats a more faithful one that has to be…
Briefing · September 13, 2026 · morning No lab shipped a creative model this window, and yet the ceiling on what a 24 GB card can make went up roughly threefold, because the work that landed was conversion work on models you…
Change the Prompt Mid-Shot With Evoke, and Watch the One Number That Says When It Stopped Remembering Evoke can repaint a scene mid-rollout because the geometry lives outside the model in a point-cloud bank, and that same architecture prints a per-chunk coverage number that is the only…
AuK Makes the Sentence the API, and Its Own Docs Cannot Keep the Sentences Straight AuK has no task selector, so the instruction string is the entire API surface, which makes its inconsistent documentation a functional defect rather than a cosmetic one.
Briefing · September 12, 2026 · morning Three releases this window published the map of where they fail alongside the map of where they work, so a creator can now decide whether a model suits the job before spending an hour and…
YuE2 Writes the Sheet Music Before It Writes the Song YuE2's editable ABC score turns revision into an edit instead of a reroll, and that matters more than the benchmark headline it will be covered for.
ID-V2V Lets Your Keyframe Be Wrong ID-V2V treats identity preservation as a relighting problem, which is why a mismatched stylized keyframe still works and why the workflow is reproducible from one video.
Briefing · September 10, 2026 · morning The releases worth having this window hand back an editable intermediate instead of a finished render, so revising a generated result stops meaning rolling the dice again.
MiniMax H3 Runs on One AMD Radeon Now. Read the Benchmark File, Not the Headline. The AMD MiniMax H3 package is worth your attention less for its 16.66 percent speedup than for a benchmark file that prints its own missed target, its failed experiments, and a measured…
Briefing · September 9, 2026 · morning Four separate releases in 48 hours sold latency rather than quality, so what changed for a creator this week is how many attempts fit in an hour and not how good the best attempt can be.
MiniMax H3 Runs on Two 16 GB Consumer Cards. Your System RAM Is the Part That Breaks. The two-GPU MiniMax H3 build is a host-memory document and a component-substitution document, and both of those decide whether it works for you long before VRAM does.
MiniMax H3 on Apple Silicon: Every Shortcut in h3.c Ships With the Number It Cost Publishing the exchange rate for every speed shortcut, rather than only the win, is what turns an optimization document into a tool you can actually tune to your own quality floor.
Briefing · September 8, 2026 · morning A 33B video-plus-audio model reached three different consumer machines in one window through three incompatible hand-written runtimes, so the hardware question for creators is no longer…
This Kokoro Core ML Conversion Hands You a Duration for Every Phoneme. That Is the Whole Subtitle Problem, Solved for Free. Splitting a TTS model at its duration predictor turns narration timing from a second forced-alignment pass into a free byproduct of synthesis, and the conversion that does it also publishes…
Foundry Turns Stable Audio Into a Drum Machine. Its Two-Second Pads Sit Outside the Model's Own Minimum Length. Foundry wraps a diffusion model in an actual instrument, and the fixed operating point that makes it playable sits an order of magnitude outside the range the checkpoint's base config…
Briefing · September 7, 2026 · morning Generative media moved onto the machine you already own this window, and the people doing the porting are publishing the settings that break it in ways the original model cards never had to.
There Is No Safe Setting for 4-Bit Quantization. There Is One Per Model. A converted checkpoint can load cleanly, dispatch natively, pass every structural check and still render a smear, so the only setting you can trust is the one you measured on your own model.
AbsorbQuant Shipped 4-Bit FLUX and SDXL Weights. The Code to Run Them Does Not Exist. A four-bit checkpoint is a promise about a runtime you have to source separately, and AbsorbQuant is what happens when the weights ship and the runtime does not.
Briefing · September 6, 2026 · morning The week's shipping was compression, and the Hadamard rotation that makes four-bit weights look good is the same thing that silently ruins every LoRA you own, with no error message and no…
VDN-H3 Renders Video Faster Than It Plays. The Row That Matters Runs on One GPU. VDN-H3's headline is an eight-GPU datacenter number, but the single-GPU row underneath it cuts an eight-step render from four and a half minutes to ninety seconds, and that is the change…
Briefing · September 5, 2026 · morning Nobody shipped a new generator this week, so the work went into output formats instead, and teaching a 33-billion-parameter video model an entirely new frame layout now costs twelve dollars…
Provcheck Signs Your ComfyUI Renders From Inside the Graph. Check Two Settings First. A provenance node that runs inside ComfyUI can attest the generation graph itself, which no downstream signer can do, but two of its defaults will hand a first-time user an unmarked file or…
H3-World Turns a Keyboard Into a Prompt Compiler, and the Whole Thing Is 131 MB H3-World shows that action control in video models is now a text scheduling problem rather than an architecture problem, but the 131 MB adapter needs a 135 GB base model and a patched…
Briefing · September 4, 2026 · morning The prompt box stopped being the only steering wheel this week, but every new set of controls is a thin layer bolted onto a model that shipped weeks ago, so the license you inherit is the…
OpenShot 4.0 Runs Subject Masking on Your Laptop With No AI Account. The Model Licenses Are the Part Nobody Read. OpenShot 4.0's local subject masking is genuinely free of subscriptions and cloud uploads, but it is assembled from three upstream models under three different licenses, and only the…
Matrix-Game 3.5 Is Apache 2.0. The Depth Model It Cannot Run Without Is Not. Matrix-Game 3.5 is a genuinely useful open camera-controllable world model whose Apache 2.0 badge does not survive contact with its own quickstart, because the depth checkpoint both…
Briefing · September 1, 2026 · morning Three separate releases this week attacked the same wall, which is that generated video forgets what it just made, and the practical result is that the unit of AI video stops being the shot…
Briefing · August 31, 2026 · morning One video model turned into a parts economy this week, and the parts are the interesting part, but the license underneath all of them shuts out the four markets where most of the people…
Text-to-MIDI Generation: How to Get a Score You Can Edit Instead of a Song You Cannot The commercial music generators still have no score underneath, so any MIDI they hand you is transcribed back out of audio they already rendered; a small set of open models emit the notes…
RenderFormer Renders Global Illumination With No Ray Tracing. The Weights You Can Download Are Not the Model in the Paper. RenderFormer is the only scene-generalizing neural renderer with weights you can download, but its released large checkpoint is not the architecture its own paper analyzes, and the fast…
Briefing · August 30, 2026 · morning The hardware requirement is the thing collapsing this week, not the models, and the machine that benefits is the Mac already on your desk.
SANA-Video 2.0 Renders 720p in Four Steps. The Apache 2.0 Badge Covers Less Than It Looks Like. SANA-Video 2.0's four-step preview is a real speed win under a genuinely permissive license, but that license covers only NVIDIA's transformer, and NVIDIA's own mirrors of the two files you…
ControlFoley Adds Sound to Silent Video. The Mode Worth Your Time Is the One Nobody Demos. ControlFoley's reference-audio mode, which copies a sound's character without copying its rhythm, is the one capability that justifies the pipeline, and a free browser demo has existed…
Briefing · August 29, 2026 · morning The two numbers that decide what a creator can make this week are the step count and the license, and both moved, while the models themselves barely changed.
FastMetal-QAD: Local Video Generation on a Mac, and Why the Biggest Model Is the Wrong Pick FastMetal-QAD makes five-second video generation practical on a 16GB Mac, and the workflow that works inverts two instincts: take the 5B over the 14B, and treat prompt caching as the main…
Bonsai Image 4B Runs a FLUX-Class Model on an iPhone. The 95% Number Hides Where It Costs You. Bonsai Image 4B's ternary compression is real and honestly documented, but the headline 95% quality figure is an average across three benchmarks that conceals a 12% drop on the exact axis…
Briefing · August 28, 2026 · morning The only thing that shipped this week is speed you have to rent, while the tools that run free on hardware you already own have been sitting finished and largely unnoticed for weeks.
Shrink Your LoRA Adapters by 80 Percent With an SVD Rank Resize Rank resizing turns oversized LoRA adapters into small ones with a documented, reproducible loss, and the fidelity numbers printed on model cards measure weight reconstruction rather than…
ACE-Step 1.5 Runs a Music Model on Your Mac. Nobody Will Quantize the Audio Decoder, and It Turns Out to Be the Cheapest Part. Every quantization of ACE-Step 1.5 independently refuses to compress the audio decoder, and that decision is correct and nearly free, which means the multi-gigabyte local bundles are large…
Briefing · August 27, 2026 · morning Creative AI models are shrinking fast enough to live on a single machine, which turns per-generation billing into a one-time download and changes what a solo creator can afford to…

Trend Board · All topics A–Z