AI Image and Video Release Notes

Latest AI Image and Video Updates

• Sep 12, 2026

  • Date parsed from source:
    Sep 12, 2026
  • First seen by Releasebot:
    Sep 12, 2026
  

  Civitai

  Weekly Update: September 5 to September 12, 2026

  Civitai adds major generation upgrades, a redesigned account settings page, smarter model and collection workflows, and a wide round of quality fixes. The changelog also previews closed-beta Training Studio and Apps improvements, plus new controls for paid models, announcements, and bug reporting.

  Civitai Changelog: September 5 to September 12, 2026

  🚀 New Features

  Generation

  • Ideogram 4.0. In the generator now, text to image, and strong at typography.
  • GPT-Image-2.5 Flare and Sunburst. Two new versions on ChatGPT Images.
  • MiniMax H3 is the default video engine. If you have already picked a video ecosystem, your choice sticks and nothing changes for you.
  • ControlNet for MiniMax H3.
  • A resolution tier for Krea 2. Krea 2 generated at a fixed size across every aspect ratio. There is a 1K and 2K choice now, where 2K roughly quadruples the pixels. It defaults to 1K, so your Buzz cost only moves if you opt in. The tier applies to the Raw and Turbo builds.
  • ZImage and Qwen run on comfy now.

  Account

  • The account settings page is redesigned. Two panes instead of one long column, across all eight sections.

  Models

  • Suggested Resources can link both ways. Tick the box while editing and this model is added to the suggested resources of each resource you add in the same edit, for models you own. Resources belonging to someone else are skipped rather than rejected, so a mixed selection still saves everything it can. Removing a link is not symmetric: a back-link belongs to the other model's list and is removed there.
  • Hide paid models. The model feed and model search both have the option now.

  Additional New Features

  • The support button opens a menu, and bug reports go somewhere useful. It used to open a modal whose options all led off-site or into the ticket portal, so a bug report became a support ticket that someone had to chase and usually could not resolve. "Report a bug" now opens the feedback panel instead, which attaches your session, the optional page capture and your images, so the report arrives with the console errors behind it.
  • A pending review counter on collections you own or manage, so a queue waiting on you is visible without opening each one.
  • Creator announcements are open to everyone. Post an announcement of your own from the Creator Studio composer and it shows on your profile. Announcements can be reported now, and a link out of one warns you that you are leaving Civitai before it opens.
  • A "Made by you" filter in the sticker tray.

  ✨ Quality of Life

  • Each file on a version downloads under its own name. A version's files all resolved to one filename, so on a model offering full and pruned builds, taking both left you with one file.
  • Sharper images on post detail and the model showcase, which now serve a 2x variant.
  • The resource picker no longer overflows on a narrow screen, drops searches you have moved on from, and stops retrying ones that timed out.
  • The mention popup stops saying "No results" while you are still typing.
  • The notifications drawer closes on tap, with the close button on the title line.
  • A checked filter chip has a visible border.
  • Placement descriptions show on the Buzz dashboard, so a placement row says what it was for.
  • A published LoRA is named after its model rather than its version.
  • Search results show the real avatar, fetched rather than taken from the search document.

  🐛 Bug Fixes

  • An article being scanned showed a 404. It shows a processing state now, rather than looking deleted to its own author.
  • Images stranded on "Analyzing image" are unstuck, and a scan that gets stuck now raises an alert instead of sitting there.
  • An image whose file never made it to storage is refused at publish rather than becoming a permanently broken row.
  • Uploads relay through our own origin when a direct upload cannot reach the storage host, instead of failing.
  • An image could be credited to the wrong person when another part of the site happened to bundle the same file.
  • Collection owners are notified on every submission path. Some routes in did not notify at all.
  • Search index. Collections re-index when their membership changes and when a post or article in them is deleted, a restored document survives a delete that arrives in the same batch, scheduled posts re-index when they actually go live, and a failed sync reports as failed instead of success.
  • Editing a model was blocked when its metadata was empty in a particular way.
  • Remixing with empty parameters no longer breaks validation, and a non-text metadata hash no longer errors.
  • The model showcase hands off to the collection page instead of paging through it.
  • Anima training no longer adds licensing fees that should not apply.
  • The trainer validates your submission on the server, with one home for the SafeTensor rule rather than several.
  • AIR links no longer render as NaN for base models that are not ours.

  🧪 In Testing

  Available to a limited group while we finish testing. If you can't see these yet, that's expected.

  • Training Studio. A rebuilt training experience in closed beta: continue training from an existing run, train on a custom base by pasting a model AIR, reuse a dataset you have already uploaded, audio playback in the caption tiles, live progress, and a Buzz picker that spends Blue first.
  • Paid model loading. Loading a checkpoint on demand when it is not already resident, so far behind a moderator gate while the groundwork is verified.
  • Apps. An Apps pill in the sub nav, the build funnel consolidated into one state-aware page, /apps/installed renamed to /apps/activity, an app's permissions disclosed on the store page before you launch it, and per-user app storage quotas.

  Original source

  Show more
• Sep 10, 2026

  • Date parsed from source:
    Sep 10, 2026
  • First seen by Releasebot:
    Sep 12, 2026
  

  Firefly by Adobe

  Record and manage voiceovers in timeline

  Firefly adds beta voiceover recording in the video editor, letting users record narration directly onto the timeline, manage clip placement, adjust volume and speed, enhance audio, and rename or delete voiceover clips in the project.

  Learn how to record and add voiceovers directly onto the timeline and manage them.

  Firefly video editor is a beta feature in Adobe Firefly, now available for testing and feedback.

  Record voiceovers

  Add to the timeline the video or image for which you want to record the voiceover. Then, position the playhead at the point from which you want to start the voiceover.

  Note

  You can't add a voiceover to an empty timeline.

  Select Record voiceover in the timeline panel. The first time you use it, or after a browser settings reset, your browser may ask for permission to let Adobe Firefly access your microphone.

  A. Input device selection menu
  B. Record/Stop button
  C. Record voiceover icon
  D. Sound meter
  E. Mute/Unmute timeline playback while recording

  In the Record voiceover menu, if you have more than one voice input device, select the one you want to use. Then, test the sound level by speaking and checking the sound meter.

  To unmute or mute an existing audio in the timeline for the duration of your voiceover, select Unmute timeline playback while recording or Mute timeline playback while recording. Existing audio is muted by default.

  When ready, select Record. Wait for the countdown of three seconds to finish, and then start the narration. As you record, the video or image will play automatically, starting from the playhead. You can narrate for up to five minutes, and beyond the video's or image's duration if necessary.

  After you complete your narration, select the stop button in the Recording menu. The voiceover clip is automatically added to the first available audio track in the timeline.

  To ensure that your voiceover is accurate and in sync with the video or image, move the playhead to the starting point of the voiceover, and then select Play. Trim or move the voiceover, if necessary.

  To record additional voiceovers on top of the one you already recorded, repeat the steps. By default, the latest voiceover clip will be in the audio track closest to the video or image in the timeline.

  To remove a voiceover clip from the timeline, select it and then press Delete.

  Edit voiceover properties

  Right-click the voiceover clip in the timeline, and then select Edit properties.

  In the Properties panel, set the following settings as needed:

  • Move the Volume slider to adjust the volume. To mute the voiceover, select Mute next to the slider.
  • Move the Speed slider to increase or decrease the voiceover speed without trimming.
  • Turn on the Enhance audio toggle to improve audio quality.

  Rename or delete voiceover clips in the project

  Right-click the voiceover clip in the Project media panel, and then select Delete from project or Rename, as required.

  Note

  To delete a voiceover clip from the project, remove it from the timeline first.

  Original source

  Show more
• All of your release notes in one feed

  Join Releasebot and get updates from Civitai and hundreds of other software products.

  Create account

  Get updates with: RSS Email CLI MCP API Slack n8n
• Sep 9, 2026

  • Date parsed from source:
    Sep 9, 2026
  • First seen by Releasebot:
    Sep 10, 2026
  

  Moondream

  Photon 2.2: Five New NVIDIA GPUs, Ampere to Blackwell

  Moondream releases Photon 2.2 with support for five more NVIDIA GPUs, expanding coverage to seven GPUs across Ampere, Ada, Hopper, and Blackwell. It also brings faster speech, lower latency, and higher throughput for Whisper, Qwen3-ASR, Parakeet, and Moondream models.

  Photon 2.2 adds five NVIDIA GPUs: A10/A10G, A100, RTX 3090, L4, and RTX PRO 6000 Blackwell. With H100 and B200, Photon now runs on seven GPUs across four architectures: Ampere, Ada, Hopper, and Blackwell.

  We benchmarked speech on L4, A10/A10G, and A100 at concurrency 1, 2, 4, and 8. Photon beats the reference engine in 35 of 36 configurations and ties in one.

  • Parakeet TDT 0.6B v3: 3.1x to 4.8x the throughput of NVIDIA NeMo
  • Qwen3-ASR 0.6B: 1.1x to 2.2x the throughput of qwen-asr on vLLM
  • Qwen3-ASR 1.7B: 1.0x to 1.2x the throughput of qwen-asr on vLLM

  Vision, language, and Whisper numbers on the new GPUs will follow.

  Also in 2.2

  • Whisper transcription now runs on L4 and RTX 3090.
  • Lower transcription latency and less host overhead for Qwen3-ASR and Parakeet. The largest gains are at concurrency 1.
  • Lower latency and higher throughput for Moondream and Qwen models.

  Speech benchmarks

  Parakeet TDT 0.6B v3: Photon throughput vs. NVIDIA NeMo

  Higher is better. 1.0x = NeMo.

  Photon throughput divided by reference-engine throughput. Higher is better. C is concurrent requests.

  Reference engines: NVIDIA NeMo for Parakeet, qwen-asr on vLLM for Qwen3-ASR.

  Parakeet gains are large on all three GPUs. Qwen3-ASR gains are smaller. On smaller GPUs, some workloads already saturate compute or memory bandwidth, which leaves less launch and scheduling overhead for Photon to remove.

  The compiler keeps compounding

  Photon compiles each model into one GPU program, a megakernel. All model execution happens inside it.

  When we started, getting one megakernel to compile at all was hard. Making it beat other engines was harder. But each compiler improvement applies to every model and every GPU it targets, so coverage compounds:

  • Photon 2.0, August 3: Qwen, Gemma, and Moondream on H100. That took months.
  • Photon 2.1, August 26: B200, plus Whisper, Qwen3-ASR, and Parakeet. Three weeks.
  • Photon 2.2, September 9: five more GPUs. Two weeks.

  Try Photon 2.2

  pip install --upgrade moondream
  

  Photon is free. Local inference needs no API key.

  Have a model, GPU, or realtime workload that needs to be faster? Email [email protected] with the model, the GPU, and the latency or throughput you need. We will compile it on Photon and send you numbers.

  Original source

  Show more
• Sep 8, 2026

  • Date parsed from source:
    Sep 8, 2026
  • First seen by Releasebot:
    Sep 12, 2026
  

  Firefly by Adobe

  Auto reframe videos to fit any aspect ratio

  Firefly adds Auto reframe in the Firefly video editor beta, automatically cropping and reframing videos for different aspect ratios on social platforms like Instagram and YouTube. It helps reduce manual cropping and supports motion tracking options for smoother exports.

  Learn how to automatically adjust your video frames based on the aspect ratios for different social platforms in Adobe Firefly video editor.

  Firefly video editor is a beta feature in Adobe Firefly, now available for testing and feedback.

  Auto reframe lets you automatically crop and reframe videos for different aspect ratios, reducing the need for manual cropping and making it easier to export videos for the web and social media platforms such as Instagram and YouTube.

  In the Firefly video editor, select the video clip in the timeline that you want to auto reframe.

  Select the required aspect ratio for the video clip using the Aspect ratio icon on the upper-right side of the timeline.

  Note

  The Auto reframe option does not appear if the aspect ratio of the original video and the one set in the timeline is the same.

  Select the Properties panel, and then turn on the Auto reframe option. You can also right-click the video clip in the timeline and then select Auto reframe.

  Turn on the Auto reframe option to automatically adjust the video's frames based on the aspect ratio.

  Tip

  When applying Auto reframe, use it with Detect cuts to avoid unnecessary camera panning.

  Once the auto-reframing is complete, select any of the following options from the Motion tracking dropdown menu:

  • Slower motion: Use it when there is very little or no camera motion in the video. The generated video is almost static and requires very few keyframes.
  • Default: Use it for most videos unless the motion is fast.
  • Faster motion: Use it when the video has a lot of motion, so the reframing follows it. It ensures the moving object stays in the frame and adds many keyframes to your video.

  Keyframes are currently uneditable; you can view them in the video timeline.

  To further edit your video's properties, turn off the Auto reframe option and modify the transform controls and motion presets.

  After completing your video edits, export the video and use it in your projects or share it across your preferred platforms.

  Original source

  Show more
• Sep 8, 2026

  • Date parsed from source:
    Sep 8, 2026
  • First seen by Releasebot:
    Sep 10, 2026
  

  Runway AI

  Latest updates

  Runway AI adds Runway Plugins for Adobe, bringing image and video generation plus timeline restyling inside Premiere Pro and After Effects.

  Paid Plans

  Introducing Runway Plugins for Adobe

  Runway now runs in a panel inside Adobe Premiere Pro and After Effects. Generate images and video with Runway's full model lineup, restyle the footage already on your timeline with Aleph 2 and place every result directly into your comp or sequence, without leaving your edit. Free download for macOS and Windows, works with your existing paid plan and credits.

  Original source
• Sep 8, 2026

  • Date parsed from source:
    Sep 8, 2026
  • First seen by Releasebot:
    Sep 10, 2026
  

  Runway API by Runway AI

  GPT Image 2.5 on Runway Dev

  Runway API adds OpenAI’s GPT Image 2.5 Flare and Sunburst to Runway Dev, bringing prompt-to-image generation with long prompts, up to 16 reference images, flexible aspect ratios and resolutions, output counts from 1 to 10, and transparent backgrounds in Sunburst.

  September 8th, 2026 - OpenAI’s GPT Image 2.5 Flare (gpt_image_2_5_flare) and GPT Image 2.5 Sunburst (gpt_image_2_5_sunburst) are now available on Runway Dev. Both generate images from a prompt of up to 32,000 characters, accept up to 16 reference images (with optional prompt-referenceable tags), support 30 width:height ratios across 1K, 2K, and 4K tiers plus an auto resolution, and take outputCount from 1–10. quality is low, medium, high (the default), xhigh, or max. Sunburst can return transparent backgrounds with background: "transparent"; Flare supports opaque and auto only. Billed at 1–76 credits per image by quality and resolution, plus 1 credit per reference image per generated image. Use the text to image endpoint to get started.

  Original source

  Show more
• Sep 4, 2026

  • Date parsed from source:
    Sep 4, 2026
  • First seen by Releasebot:
    Sep 5, 2026
  

  Civitai

  Weekly Update: August 30 to September 4, 2026

  Civitai ships a major August 30 to September 4, 2026 changelog with Muse Image in the generator, sharper comment controls, customizable navigation, better collection sorting, and a wide set of quality-of-life improvements and bug fixes across feeds, deposits, messaging, and creator tools.

  Civitai Changelog: August 30 to September 4, 2026

  🚀 New Features

  Generation

  • Muse Image. Meta's image model, in the generator now, with text to image and image editing.
  • MiniMax Music 3 output shows up in the queue rather than generating into nothing you could see.

  Comments

  • Mute a single comment thread. Notification settings used to be all or nothing: turning off replies to threads you are in silenced every thread on the site. You can now mute one thread and leave the rest alone. The mute applies to the whole chain above a comment, not just its immediate parent, so muting partway down a long thread works the way you would expect. Direct mentions still come through, because being named is not the same as somebody replying in a thread you happen to be in. It covers comments on images, posts, articles and bounties. Model page discussions run on an older comment system and are not included yet.

  Models

  • The model type picker is grouped, and some types have been renamed. Types are now sorted into fine-tunes, adapters, component replacements, workflow additives and other. Eleven types that were no longer worth picking have been dropped from the list, and a model already set to one of them keeps it. Clicking the type you already had selected no longer clears the field.
  • Your last pricing is remembered per model type, so setting up a second LoRA does not start from scratch.
  • A retired template type says so when it gets replaced, rather than changing quietly.

  Additional New Features

  • Customise your sub nav. Reorder items, hide the ones you do not use, and collapse the rest into More. Placement comes from your own saved config now rather than from how wide your window happens to be.
  • Collections sort by when things were added. Newest sorted on the item's own age, which for images is upload order, so adding a freshly published image to a collection could drop it far down the list if it had been sitting scheduled for weeks. There is a Recently Added sort now, and it is the default for image, model, post and article collections. Newest still does what it always did.
  • Featured collections rotate on the home page instead of being redrawn each time.
  • A Prime toggle for Wan 3.0 in the generation form.

  ✨ Quality of Life

  • The articles feed has its category bar back. It was removed from every feed at once in a change aimed at images. Models got theirs back a while ago and articles were left out, so this restores them.
  • Cash-outs are out of the Buzz dashboard chart. A cash-out is a debit like any other, so it was landing in the same series as your generation spend. At creator volume one bar was thousands of times the next, the axis scaled to that bar, and every other day rendered under a pixel.
  • A tag-filtered feed has a way back out. Opening posts or articles from a link that carries a tag narrowed the feed to it, with nothing on the page you could use to clear it again.
  • The submit bar on a review clears the home indicator on iPhones with a notch. The page was not asking for the full screen, so the bar's bottom spacing always fell back to its default and sat too low.
  • The crypto deposit screen always shows the minimum now. It printed the amount only when the fiat conversion came back, and that value is often missing, so the line could read "Minimum BTC deposit:" with nothing after it. A blank reads as no minimum rather than as a number we could not fetch, and a deposit under the minimum is not recoverable. It falls back to the crypto amount, and says so plainly when it has neither.

  🐛 Bug Fixes

  • A Buzz purchase could credit zero Buzz despite real payment. A multiplier that came through as zero, negative or not a number made the calculation return nothing, and the purchase path then wrote its idempotency marker, so retrying could not repair it. Multipliers are floored at 1 now, on both the recording and the spending side, so a purchase always credits at least what was bought.
  • Magic link sign-in was broken for anyone with a capital letter before the @ in their email. The send path started preserving the case of your address on August 26 while the verify path was still lowercasing it, so the link was looked up under a different address than it was stored under, missed, and bounced to a verification error. It failed on every attempt no matter how many links you requested. 898 addresses hit it, and most of them were new signups with no account yet, so this was blocking people from joining as much as from signing in.
  • A model type change left the old licence fee on its versions. Per image fees belong to fine-tunes. Changing a model to a LoRA afterwards was not checked, so the version kept charging the old fee to everyone who generated with it. A save that omits the fee is now judged against the stored stamp too.
  • Profile pictures were being hidden on other people's content. Avatars were rated against the page they appeared on rather than against themselves, so on a PG image every commenter with a PG-13 avatar had theirs suppressed, even for viewers who accept PG-13.
  • An unpublished article was always called a Terms of Service violation. The take-down banner said that whatever the actual reason was, and the reason list itself was written about models, resources and example images. An article pulled for being thin read as an accusation, and authors escalated it as one. Reasons are now tagged as either policy or quality, the banner follows that rather than assuming the worst, and the wording is written for an article's author.
  • A notification link with no post id crashed the image page.
  • Blocked people's messages are suppressed in group chats.
  • Video controls on the post page. The link wrapper was swallowing gestures aimed at the video controls, and the metadata button sat over the fullscreen control.
  • The guided tours stop failing silently, with the failures now measurable rather than invisible. This is the follow-up to last week's round.
  • Remix provenance is recorded when you click remix, rather than read back off a URL the form may have rewritten.
  • A reaction could come back empty on refresh, and the reaction milestone was counting suppressed accounts.
  • A tooltip on a menu button stopped the menu opening.
  • Sticker artwork stopped rotating under the draft editor's chrome.
  • Restoring a deleted model updates its timestamp, so it reappears where it should, and a sweep that flips a model to Draft does the same.
  • Image deletes dropped by a degraded cache are recovered into the search index.
  • Manual features count toward the per creator cap.
  • Creator Program stage notifications are sent at the actual boundaries.
  • A Creator Shop pack submission fee can be paid in any currency the server accepts.
  • Training label preference was inverted for base model types.

  🧪 In Testing

  Available to a limited group while we finish testing. If you can't see these yet, that's expected.

  • A rebuilt generator form. The generation form is being ported onto a new foundation, staged behind flags, with the old and new lanes running side by side while we check they agree.
  • App Blocks. Store card and layout work, review entry points in the host chrome, and the page resource picker widened to the LoRA family.

  Original source

  Show more
• Sep 4, 2026

  • Date parsed from source:
    Sep 4, 2026
  • First seen by Releasebot:
    Sep 4, 2026
  

  Synthesia

  Adjust speaker attribution for dubbed videos

  Synthesia adds one-click speaker editing for dubbed videos on Enterprise plans, letting creators reassign, create, and rename speakers to fix attribution errors and clean up dubbing workflows.

  Reassign, create, and rename speakers in a dubbed video with one click—available on Enterprise plans.

  You can now edit the speaker attribution, create new speakers, and rename then in your dubbed videos. Available for creators on Enterprise plans.

  How it works and why it matters

  Dubbing relies on automatically detecting who's speaking. That detection doesn't always get it right—a line can end up attributed to the wrong person, or a distinct voice can go undetected entirely. Until now, fixing that meant either living with an unusable dub, or correcting it word-by-word and still ending up with awkward pauses in the result.

  Now you can fix it directly:

  • Reassign a speaker on any segment with one click, when a line's been attributed to the wrong person
  • Create a new speaker if someone's voice wasn't picked up as distinct—this builds a unique voice clone from the audio you assign to them
  • Rename speakers with real names instead of "Speaker 1, 2, 3," so your team can review edits without guessing who's who

  For more on how it works, check out our documentation.

  Original source

  Show more
• Sep 4, 2026

  • Date parsed from source:
    Sep 4, 2026
  • First seen by Releasebot:
    Sep 4, 2026
  

  Synthesia

  Build better videos with Layouts

  Synthesia adds Layouts, a new Editor library of 100+ ready-made scene designs with instant theme previews and one-click swaps in Storyboard, helping creators design polished scenes faster across manual builds and Assistant-made videos.

  Design any scene with Layouts: a library of 100+ ready-made designs right in the Editor, with instant theme previews and one-click swaps in Storyboard

  There's a new way to design a scene in Synthesia: Layouts. Add a new scene to your video and you'll find a Layouts tab right next to Template—a library of 100+ ready-made scene designs you can browse and apply with a click. Live today for all Synthesia creators.

  How it works

  Add a new scene to open up the Layouts menu (next to Template).

  From there you can:

  • Start with Suggestions: our picks for what fits the content already in your scene
  • Or browse by type: Intro, Speaker, Media, Text, Numbers, and more
  • Preview different themes before committing: see the same layout in a few different looks. If you're not on Enterprise, this is where you'll find our four built-in styles (Ember, Bloom, Dusk, and Crimson) so your scenes look polished from the start
  • Apply with one click: your script and content stay exactly as they are, only the design changes

  This isn't limited to scenes you build by hand: Assistant draws from the same layout library, so videos it puts together for you get the benefits too.

  What you can do

  • Browse and apply Layouts that match what you're trying to say
  • Preview a Layout with different themes before you commit
  • Swap Layouts in Storyboard to try a new design quickly, without retyping anything
  • Undo a Layout change if the previous version worked better

  Original source

  Show more
• Sep 4, 2026

  • Date parsed from source:
    Sep 4, 2026
  • First seen by Releasebot:
    Sep 4, 2026
  

  Synthesia

  Find the right voice faster with the refreshed voice picker

  Synthesia introduces a refreshed voice picker that makes browsing and previewing voices easier, with a bigger full-page layout, featured voices at the top, filtering by language or accent, instant previews, and quick switching across Synthesia, personal, and shared voices.

  Browse Featured voices, filter by language or accent, and preview instantly, all in a bigger, easier-to-scan voice picker.

  Today, we're introducing a refreshed voice picker—bigger, easier to scan, and built for browsing when you don't already know exactly which voice you want.

  What’s new

  The old picker was a compact dropdown—fine once you knew the voice you wanted, but cramped for actually browsing a large catalog.

  The new picker opens as a full page instead. You'll find Featured voices (our picks) right at the top, or you can explore everything across three tabs: From Synthesia, My voices, and Shared with me, each showing how many voices are available.

  What you can do

  • Search by name, filter by language or accent, or sort the list to narrow things down fast
  • Press play on any voice row, or use the sticky preview bar at the bottom (with waveform) to audition a voice without losing your place
  • Swap voices directly from the list with the "+" button on each row, instead of opening one first
  • Switch between Synthesia's voices, your own, and ones shared with you

  For more on how it works, check out our documentation.

  Original source

  Show more