Feed
- Image outputs have an Edit button in the full-size viewer. It opens the image in the asset editor, and every version you make is saved to your library as a new image. The cell and the export keep the run’s original. After you commit, a message links to the new image.
- The feed tools in the Orisu MCP server reject arguments they don’t recognize. A misspelled argument now returns an error instead of being ignored.
- Create is a single desk for an image or a clip without building an agent: a prompt, references brought in with @, a model, and a Direction console with four slots. Style holds the curated looks, shelved by studio. Camera, Light and Color are dials: lens, aperture, angle, shot size, key light, time of day, contrast, palette. Every dial has a picture of what it does, rendered from its own prompt words.
- Two studios. Marketing opens on image with product, ad, UGC, brand and illustration looks. Cinema opens on video with film looks, camera moves and period.
- Each dial adds one clause to the prompt, and Shot notes under the composer shows the exact words the model is given. A chosen dial replaces what a look implied instead of arguing with it: Neon Night with Golden hour now asks for neon practicals at golden hour, not night and low sun at once. Looks that change the medium open the prompt, so an ink drawing is an ink drawing.
- The Brain on the desk. Switch on the Brain chip and the prompt is rewritten in your brand before shooting: your palette named in words, your materials and settings, your products where they fit, and none of the words you never use. The rewrite draws on the memories nearest the prompt. The row keeps what you typed and shows the rewrite beneath it, so a reroll starts from your words. Each rewrite costs a few credits, billed with the generation.
- Reach it from the rail’s plus, the dashboard, or Quick create on the agents page.
- Put outputs into named collections. Use the folder button on an output, at the top of a row for the whole step, in the lightbox, or on the selection bar for several at once. The Collections chip next to Kept shows one collection at a time, and only the outputs you filed in it. Deleting a collection never deletes an output.
- The model picker marks a model as Failing when most of its recent runs have failed, and the model chip warns you while one is selected. Failed runs are refunded and the mark clears once runs succeed again.
- In the studio, hovering a model now shows how long it usually takes and how many of its recent runs failed, measured from real runs on Create.
- Draw on an image in the editor (arrows, boxes, pen strokes, labels) and the Commit button turns into Commit annotations. Clicking it bakes your marks into a new image and saves that as its own asset, with a link back to the original. Whatever you opened the editor from (a node output, a review item, the archive) receives the new asset, and the original stays as it was.
- The marks themselves are stored on the new asset too, along with a short description such as “a red arrow pointing at the right”. A later step, an image-to-video node for example, can read what you meant instead of guessing from the pixels.
- On Create, a committed edit now lands in the composer. Edit a reference card and the card switches to the new version, keeping its name so @mentions still work. Edit a finished shot and the new version joins your references.
- Connect an annotated image to Generate Video (first frame, last frame or reference image), Generate Image (reference image) or Edit Image, and the prompt gets one more sentence, such as “Follow the drawn annotations on the first frame: a red arrow pointing at the right.” Draw a path on a still and the clip is asked to follow it. Images without annotations change nothing.
- Create does the same for annotated reference cards. Shot notes shows the sentence before you shoot, and the history row shows it under your prompt.
- On the prompt library, the filter panel no longer follows you down the page. Once you scroll past it, a slim bar keeps the search field and one chip naming your current filters. The chip takes you back to the full panel.
- Reading your website now shows its progress. A tray at the bottom of the Brain map follows the four steps (visit the site, read the brand, write facts, collect images) and counts the facts and images as they arrive. When it ends, it says whether the read finished, finished with a gap, or failed, and offers to try again.
- Reading a website again no longer fills the review list with facts the Brain already has. The reader is shown what the Brain knows, and anything that only repeats it is left out. A fact whose numbers changed still comes through, so updates are recorded.
- The review list has Confirm all and Dismiss all. Each asks once, then acts on every fact waiting for review. Dismissing moves facts to history rather than deleting them.
- The Media tab has Remove all. Search first and it removes only what matches. It asks once before anything goes.
- The brand profile holds more: colour roles, fonts by job and weight, tone words, values, audiences, products with their selling points, pronunciation, and tone by channel. It opens from the brand chip in five tabs, and each part you fill in shows on the map beside its category.
- Voice, palette and fonts edited on the Brain now reach your runs. Before, only the word rules did.
- The palette, fonts and voice read from an uploaded brand book arrive in review. Confirming one adds it to your profile.
- The Brain asks about what it is missing. A questions chip on the map opens them one at a time, and each answer goes straight into your profile. The copilot asks at most two per conversation, and never about something already filled in.
- Memories about channels, campaigns, feedback, products and audiences now expire. An expired one stays listed but stops reaching your agents until someone marks it Still true. What your brand is, and its standing rules, never expire.
- Searching the Brain matches exact names as well as meaning, so a product name finds the memory that names it.
- Seedance 2.5 renders a single take of up to 30 seconds. One shot can follow up to 30 reference images, 10 reference videos and 10 audio clips.
- Hailuo H3 from MiniMax comes in three tiers. H3 goes up to 4K and takes reference images, video and audio. H3 Max and H3 Max Turbo stop at 1080p and cost less per second, Turbo most of all.
- Wan 3.0 makes clips of 2 to 30 seconds with sound, from a prompt, a start and end frame, or reference images. Its Prime tier costs a little more for higher quality.
- FLUX 3 is Black Forest Labs’ first video model. It renders up to 20 seconds with sound from a prompt or from a start and end frame, and its Draft tier renders at 720p for less. Edit Video can also use it to extend a clip shorter than 15 seconds.
- Grok Imagine Video 1.5 goes up to 1080p and can work from up to 7 reference images.
- Kling 3.0 has two new tiers, Turbo Standard and Turbo Pro, for quicker and cheaper clips of 3 to 15 seconds. Turbo runs don’t generate audio and ignore end frames, negative prompts and elements.
- Gemini Omni Flash is now version 1.1. You can pick a resolution between 360p and 4K, animate between a first and last frame, and guide a shot with up to 10 reference images and 3 short reference videos. Graphs built on the old version run unchanged.
- GPT Image 2.5 has two tiers. Flare is the fast one, and Sunburst takes more care over fine detail and lettering. It offers six quality levels and transparent backgrounds, and an edit can take up to 16 reference images.
- Seedream 5.0 Pro generates and edits from up to 10 reference images.
- Grok Imagine Image 2.0 generates and edits at 1K or 2K from up to 3 references.
- Claude Fable 5.1, Claude Opus 5 and GPT-6 Astra (Standard and Pro) are in the model picker of every text node.
- Create’s default model list now has GPT Image 2.5, Seedream 5.0 Pro, Hailuo H3 and Seedance 2.5.
- Veo 3, Sora 2, Imagen and SeedEdit v3 have been retired by the service that hosts them, so they no longer appear in model pickers. Graphs that use them keep their settings, but switch them to a current model: Imagen runs already fail, and the others could stop working at any time. Veo 3.1 is the closest replacement for Veo 3.
- Seedance 2.5 now charges by resolution at the same rates the model is billed at. A 480p clip costs about half what it did, and 1080p costs slightly more.
- Editing with Seedream 5 Lite works. It had been failing because reference images weren’t sent to the model.
- Veo 3.1 Lite at 1080p with sound costs a little more per second, matching what the model is billed at.
- On Veo 3.1 Lite, a clip with both a first and a last frame has to be 8 seconds long. Other lengths now stop with a message saying so, instead of failing at the model.
- Veo 3.1 opens a clip from one image. To guide a clip with several reference images, pick Veo 3.1 Reference.
- Veo 3.1 Fast and Lite no longer show Enhance Prompt. Only the Standard tier ever used it.
- Kling Motion Control on Generate Video sends your reference video as the movement to copy. Runs had been failing because the video never reached the model.
- Kling O3 and O1 on Edit Video use your reference images. They had been ignored.
- Edit Video on Kling O3 Pro and 4K runs on the tier you picked. It had been running, and charging, Standard. Pro and 4K edits cost more per second than Standard.
- Graphs still using the retired Kling 3.0 Elements entry run again, and charge for sound when it’s on, like Kling 3.0.
- LTX 2.5 from Lightricks comes in Fast and Pro. It makes clips with sound from text or a start frame, Fast goes up to 4K, and if you connect an audio track it builds the video around it.
- Luma Ray 3.2 makes 5 or 10 second clips from text at up to 1080p. Luma Ray 3.2 Image to Video animates a first frame, with an optional end frame, into a 5-second clip for less.
- Kling O3 can now start from a first frame, or follow up to 7 reference images or one reference video, on Standard, Pro and 4K. Reference images had been ignored before this.
- Veo 3.1 Reference makes an 8-second clip around up to 3 reference images, so a person or product stays the same from shot to shot. It comes in Standard, Fast and Lite.
- Edit Video can turn a clip into another shape with LTX 2.3 Reframe or Luma Ray 3.2 Reframe. Both fill the new edges instead of cropping. LTX Reframe still asks for an instruction, which it ignores.
- Edit Video can also re-render a clip from a prompt with Luma Ray 3.2, and add 6 seconds to a clip with Grok Imagine Video 1.5.
- Veo 3.1 Extend adds 7 seconds, with sound, to a clip on Edit Video. It comes in Standard and Fast.
- Topaz Video Upscale is a new model for Upscale Video. It enlarges clips up to 5 minutes long by 2x or 4x, in Precision, Creative or Generative mode.
- Sync Lipsync 3 re-times a video’s mouth to new audio, or makes a portrait photo talk.
- HeyGen Lipsync comes in Precision and Speed. Speed costs half as much and suits drafts.
- VEED Lipsync 2 matches a video’s mouth to a new voice track.
- OmniHuman 1.5 turns one photo and a voice track into a talking video where the whole person moves, gestures included, at 720p or 1080p.
- Qwen Image 3 handles text inside images well, in English and Chinese, and edits from up to 3 reference images, at 1K or 2K.
- Recraft 4.1 has four tiers: Standard, Pro for large formats, and two faster Utility tiers. Standard and Pro can output SVG vectors.
- Luma Uni-1 generates from up to 9 reference images with optional web search, in shapes from 3:1 banners to 1:3 tall, and edits an image from an instruction. It comes in Standard and Max.
- Topaz Upscale has four modes. Precision is the default and works as before. Creative and Generative rebuild detail in AI-made or low-resolution images, and Transparent keeps the alpha channel and always enlarges 4x. The mode you pick in the panel is now the one that runs and the one you’re billed for.
- MiniMax Music 3 makes songs with vocals up to 5 minutes long. It sings your description as the lyrics, so write the words you want, or turn on instrumental.
- Lyria 3 from Google comes in Standard and Pro at a flat price per track, and picks the length itself.
- Mirelo SFX 1.6 adds sound effects to clips up to 60 seconds long. Mirelo 1.5 stopped at 10.
- The copilot now runs on Claude Opus 5 and Claude Sonnet 5. They think before answering, so some replies take a little longer.
- The feed’s AI column offers Claude Sonnet 5 as its higher-quality option, and column mapping during import runs on Claude Opus 5.
- Seedream 4.5 image edits were failing and work again. They now take up to 10 reference images. Billing fixes
- Nodes inside a loop now charge for what each run actually used. Lip sync, video upscaling, video editing and other models priced by the length of the input clip could come out close to free when run in a loop.
- Prices that depend on what you connect now apply. Seedance 2.0 gives its lower rate when you use a reference video, and Seedream 5.0 Pro and Luma Uni-1 charge a small amount for each extra reference image.
- Topaz Video Upscale charges double for 60 fps clips, as the model does. Before this, every clip was priced as 30 fps.
- Qwen Image 3 edits cost the 2K price every time, because the model can pick a 2K size on its own.
- LTX 2.5 Audio to Video is now its own model. Connect a voiceover or music track, and optionally a first frame, and it makes a 1080p clip as long as the audio, charged for that length.
- Models priced by the length of your clip, such as lip sync, charge for the real length. Some nodes had been charging a fixed 5 seconds whatever the clip ran to.
- The copilot, AI columns, feed mapping and the Brain now charge each Claude model at its own price plus the same margin as every other model. Copilot turns on Opus are about half what they were, Haiku calls cost a little more, and Sonnet 5 is unchanged. The one-time feed mapping fee drops from 14 credits to 7.
- Claude Sonnet 5 in text nodes now costs what Sonnet 5 costs, which is less than the older Sonnet models.
- Sync Lipsync 2.0 has a Quality setting, Standard or Pro. Pro now really runs the Pro model. Before, every run used Standard, whatever the setting said.
- The Upscale Video panel only shows settings the selected model uses. FlashVSR keeps its resolution, speed, audio and format options, and Topaz gets a 2x or 4x upscale factor.
- LTX 2.3 Reframe and Luma Ray 3.2 Reframe no longer ask for instructions in Edit Video or the asset editor. Connect a clip, pick the new shape and run. You can still type instructions to steer what fills the new space.
- MiniMax Music 3 has a Lyrics box for the words the singer sings, with tags like [verse] or [chorus] on their own lines. Leave it empty for an instrumental track. Your style description is no longer sung as lyrics. References in Generate Video
- You can connect several audio clips to Generate Video, up to 10 on Seedance 2.5. Videos and audio wired into Context are now used as references instead of being ignored.
- Generate Video warns you when you connect more reference images, videos or audio than the model takes, when Seedance gets audio without an image or video, and when a first or last frame will be treated as an ordinary reference.
- The panel shows how many references the model takes, and the aspect ratio picker offers the ratios the selected model supports, including Auto on Seedance 2.5.
- On models that don’t support subject references, such as Seedance 2.5, element images are sent as reference images instead of only adding to the prompt.
- Seedance uses every audio clip you connect: up to 10 on 2.5 and 3 on 2.0. An audio clip on its own no longer switches it into reference mode.
- Only an image in First Frame opens a clip. Brand, product and context images are sent as references on Seedance, Wan 3.0, H3, Happy Horse, Gemini Omni, Grok Imagine Video 1.5 and Kling O3. Before, a brand logo could become the first frame.
- Wan 3.0 takes up to 5 reference audio clips, and H3 and H3 Max take up to 3.
- Connecting a second node into an input that takes one connection, such as a first or last frame, now replaces the old connection. Before, the extra wire was silently ignored. Undo brings the old connection back.
- Create takes up to 8 references per shot. When the strip is full, the add buttons turn off and say so, so you find out before you press Generate.
Community apps as feed columns
- The feed’s agent picker now has an Explore apps section listing public community apps — each card shows real outputs the app has produced, what it makes, and who published it. Pick one and it mounts as a run column like any of your own agents — Orisu saves a private copy to your workspace behind the scenes, so runs bill to your credits and the author’s later edits never change what your feed produces. Adding the same app again reuses the copy.
- Adding an agent column is now a three-step wizard: connect inputs (Orisu pre-matches app inputs to your columns by name and type), test on a single row — one real run with that row’s values, output shown right in the dialog before you commit to a batch — then name the column and pick which outputs become their own columns.
- When the author of an app you’ve added ships a new version, the column’s settings show an update button — one click pulls the new version in and re-pins the column. Until you click it, nothing changes.
- The reverse door works too: every app in Explore has a Run over a feed button. Pick a feed (or create one) and you land straight in the column wizard with that app attached.
- Settings → Connections: one place to see every external account your workspace has connected (Shopify, Google Drive), connect new ones, reconnect expired ones, and disconnect — instead of hunting for the right dialog.
- Date cells and the bulk editor now open their calendar on the month of the selected date instead of the current month.
- Right-click and dropdown menus across the app share one design: the same surface, spacing, and destructive styling the studio’s canvas menus use.
- The feed header now works for you: an amber “N to review” counter aggregates every agent column and clicking it opens the first cell waiting on you, a pulsing badge shows how many runs are in flight, and the feed’s name opens its own menu — rename, description, duplicate, delete — without going back to the feeds home. ⌘K opens the copilot from anywhere on the page.
- Click any image in the grid and arrow keys walk you through the whole column, product by product, with each row’s title as the caption. Works for your own product photos and for generated outputs alike.
- Review decisions are no longer final: flip an approval to a rejection after the fact, or revert a cell to needs-review. Same-decision clicks stay no-ops.
- Export knows when it’s done: downloading the CSV now shows the exact row count and filename when it finishes, with pointers to Meta Commerce Manager and TikTok Catalog Manager. Row counts are exact even when product descriptions span multiple lines.
- The feeds home meets you at your size: with a handful of feeds there’s no folder rail or search bar — just your feeds, full width. The chrome appears as the library grows, and creating a folder always brings the rail back. First-time visitors see the loop itself — import, generate, publish — instead of an empty box.
- Finished cells read as calm text instead of a wall of chips; only cells that need you — review, errors — keep their color.
- The copilot on the feed can now use your Brand Brain: it checks your voice and rules before writing product content, and files brand facts it learns along the way. The same Brain switch from the studio sits beside the composer.
- Brain in the sidebar is where your brand knowledge lives. Point it at your website and it reads your palette, fonts, voice and claims, and pulls in your product photography. Drop in a brand book and it reads the document page by page, citing the page each fact came from. Tell the copilot something durable about your brand and it files that too.
- Anything Orisu works out on its own waits for a person before it counts as settled. The review queue puts the facts it is most confident about first, so the clearest calls are the quickest to make.
- Every memory shows who taught it and who signed off, so a rule you disagree with has someone you can ask.
- Edit a memory to fix its wording or move it to a better category. Most extracted facts are close but not exact, and correcting one now takes a single action instead of dismissing it and typing it again. Saving a correction confirms it.
- Three ways to read it: a map of your brand as a graph, a searchable list of every memory, and a gallery of every image the Brain holds.
- If you have explained your brand to ChatGPT or Claude before, Import memory hands you a prompt to paste into it. Paste the reply back and Orisu reads the brand out of it. You see what is in the reply before anything is saved, and a profile field you have already written is never overwritten.
- With the MCP server connected, an outside assistant can read your brand knowledge, search it, and add to it, instead of asking you to describe your brand again. Anything it proposes lands in your review queue like everything else.
- The Brain switch beside the copilot composer turns brand knowledge off for a conversation where you would rather work without it.
- A workspace holds one brand. Encoding a second website into a workspace that already has one now asks first, and offers to re-encode the existing brand instead, which keeps every fact, image and agent already attached to it. If you work across several brands, give each one its own workspace.
The Feed: run agents across your whole catalog
- A new grid surface at Feed: rows are products, columns are data, agent runs, or run outputs. Built for bulk work — 5,000 rows scroll smoothly, cells edit inline, and run columns mount any published agent with its inputs bound to your data columns.
- Import from CSV or Shopify with a field mapping you confirm once. Re-imports only touch rows that actually changed, so nothing re-runs or re-bills needlessly.
- Normalize with AI: an agent inspects your source data and writes small per-column normalization scripts (prices, URLs, option values) that run deterministically over every row — with values validated against each column’s type and rejects reported instead of silently stored.
- Run, review, approve: trigger a cell, a column, or a selection. Big batches show a cost estimate and ask first; unchanged cells re-run for free. The review drawer walks you through approve / reject / re-run and jumps to the next cell needing review. One click turns any run output into its own column.
- Export to Meta and TikTok as a hosted CSV at a stable URL — paste it once into Commerce Manager or TikTok Catalog Manager and the channel pulls updates on its own. Blocked rows (unreviewed, images under 500×500) are listed with reasons. Plain CSV download included.
- Feeds everywhere: the same operations are available on the API and as MCP tools, so external agents and the copilot can build and manage feeds too. See the Feed guide.
- Ask the feed Copilot for a bulk edit in plain language (“strip the ‘SALE —’ prefix from every title”) and it writes a small sandboxed script, previews real before → after values for a sample of your rows, and only applies it to the whole feed (up to 10,000 rows) after you approve — always an all-or-nothing write, with rejected cells left unchanged and reported.
- Normalize with AI now shows the actual executed before → after values in the import dialog, not just a description — and once you confirm, the saved script replays automatically on every future re-import.
- Channel-compliant export: Meta and TikTok exports are now checked against each channel’s real catalog spec — required fields, formats, and enums — with blocked and warning reasons per row (prices must read like
"9.99 USD"; newly-required fields like Meta’sbrandwill now name exactly which rows and fields need mapping). ✨ Map with AI proposes a full field mapping plus a compliance report (fixable / blocking / warnings) in one call. - AI-assisted normalize and Map with AI (including the copilot’s mapping suggestions) each cost a flat one-time credit fee per call; the script or mapping they produce then runs free on every row, every time.
- Every data column now has a settings popover (Edit column in the header menu, or double-click the header) for its name, description, default value, and an advisory required flag, plus per-type options like currency, precision, and image size minimums. Date columns can default to the row’s creation time instead of a fixed date.
- Select and multi-select columns get colored options you can create inline right from the cell editor while typing, no trip to column settings needed.
- Image, video, and file cells support drag-and-drop upload or pasting a URL directly in the cell, and date cells edit through a calendar popup instead of typing a date string.
- Change a data column’s type from the same settings popover: existing values convert where possible (converting to select even builds the options list from your data), incompatible values are cleared, and a toast reports how many of each.
- CSV import now infers a type for every new column from the values in your file — numbers, prices, dates, URLs, and more — with a per-column override right on the mapping screen, so catalogs land typed instead of all-text.
- Selecting rows now surfaces a floating action bar: Run a column across the selection (with an up-front “Run M cells · ~E credits” preview before anything actually runs), Edit cells to set a value / copy the first row down / clear across every selected row in one step, Duplicate rows, and Delete rows.
- A live progress strip appears above the grid while a batch runs — Running / Queued / Done / Failed counts and a progress bar — with a Cancel button and a completion summary (succeeded / failed / cancelled, credits used) when it finishes.
- Batches can now be cancelled: queued cells stop instantly, running cells stop shortly after. Cancelled cells show a neutral “Cancelled” chip (not an error) and can be re-run like any other pending cell.
- Hover any row (or press Enter on a focused cell) to expand it into a record panel — every column shown as a labeled field in column order, instead of squeezed into a 42px cell.
- Data fields edit in place with the same type-aware editors as bulk edit: text/number/price save on Enter or blur, toggles and single-select save instantly, and dates/multi-select/JSON fields save when you leave the field. Required fields show an asterisk and an advisory tint when empty.
- Run fields show a live status chip with Run/Re-run and View run; output fields are read-only with a click-through to the run that produced them.
- Prev/next navigation walks rows in the grid’s current sort/filter order, with a “Row N of M” position indicator.
- A Copilot button on the feed detail page opens a chat panel over the grid, scoped to the feed you have open — no more switching to the studio to ask about a catalog.
- It can read the feed, add or update columns (including settings and type changes, with a report of what converted and what was cleared), set cells, add rows, estimate what a batch would cost, and run, cancel, or review runs.
- Writes show a card you approve before anything changes, and cost-gated runs show the same estimate-first approval card as the grid; approved changes appear in the grid immediately.
- Conversations aren’t saved between page loads yet — a known v1 limitation.
- Turning sharing on for an agent that was already shared used to mint a new URL and quietly kill the old one, so a link you’d sent a client could stop working because you opened the share panel again in another tab, or asked Claude to share the same agent twice. Sharing an already-shared agent now hands back the same URL.
- Only the Regenerate link button changes the URL, which is still the way to cut off anyone holding the old one.
- Nothing about this changes what viewers see: a share link has always rendered your agent as it is now, so later edits and runs show up on the same link without re-sharing.
- Captions came out invisible in rendered video and images. The renderer drew text using whatever fonts the server had installed, and the server had none, so the text was drawn with nothing while the pictures and clips composited normally. Fonts now ship with the renderer, so text in a render matches what you see on the canvas, for all seven font choices and every weight.
- Renders could cut your footage short and finish on blank background. When the composer couldn’t read a clip’s length it quietly assumed five seconds, so a longer video was trimmed to five and the rest of the composition came out as background colour, with the run still reporting success. Clip length now comes from the file’s own metadata, and a length that genuinely can’t be read stops the run with an error instead of guessing.
- Relatedly, the timeline used to measure at least ten seconds even over a shorter clip, so a caption dragged across the visible ruler could be stored ending later than the video did. The ruler now measures your real content, and a caption sitting on top of a clip no longer stretches the render past it. Set Duration s when you do want a video longer than your footage.
- Every layer is now a frame, with one set of controls whether it holds a caption or a picture. Frame size decides how the container sizes itself; Content fit decides how the content sits inside it — Fill, Fit or Stretch for media, Shrink text to fit for captions, which steps the font down until it fits rather than running off the canvas. Text that still doesn’t fit is clipped to its frame instead of spilling across the render.
- An anchor grid picks where content sits in its frame: which part of a picture survives a Fill crop, or where a caption sits vertically. Squeezing a landscape clip into a vertical ad used to always crop to the centre, which tends to cut off the subject.
- Lock frame freezes a frame’s position and size while its content keeps reflowing, so a placeholder you’ve placed stays put. Captions that still don’t fit at the smallest size get a red outline in the editor instead of being quietly clipped in the render.
- When a connection feeds the composer several values with Loop on, the editor showed only the first one, so you styled against one caption and found out how the rest looked after the renders finished. A variant control above the canvas now pages through all of them.
- Turning on Loop for a single connection now fans out only that connection’s items. Every other input on the port comes along inside each run. Two product shots looping under one fixed logo gives you two runs that each contain the logo, instead of three runs where the logo sits by itself in the third.
- The port-level Loop switch still iterates the connections themselves. Two looping connections multiply their items against each other, and a connection with Loop off adds no dimension of its own.
- The Loop switch on a connection works without turning on the port switch first, and what it shows now matches what the run does. Some composers that had been fanning out quietly started layering instead, so two images on one upload connection rendered a single composition rather than one per image. That is fixed.
- Cost estimates and the iteration preview follow the same rule, so the number of runs you see before pressing Run is the number you get.
- An upload holding two images used to create a hidden layer for each one and composite them into the same render, so the second painted over the first and you got a single image out. Each connection is now one layer, and turning on Loop for it repeats the whole composition once per image. Compositions you already saved keep the position, size, and styling you set.
- PNGs exported by Google carry Content Credentials, and the renderer misread that data as a broken image and dropped the layer. The composition came out missing exactly those pictures while the run still reported success, so there was nothing to tell you what had happened. Those images render now, and a source that genuinely cannot be read gets logged instead of vanishing quietly.
- List Selector accepts image, video, audio, and file lists, not only text. Items show as previews you can actually look at, both on the card and in the focus dialog, instead of reading “[object Object]”.
- A node that ran many times can drop its finished results onto the canvas as separate input nodes, grouped together. Pick through them, delete the ones you don’t want, and wire the rest into the next step. The original node and its connections stay as they were, and one undo reverses the whole thing.
- An upload node holding a single image, video, or audio file renders as a full media card like the generate nodes do, instead of a padded thumbnail row.
- Starting a run and then editing the canvas could put the studio into “Updated elsewhere” and freeze saving, even with nobody else on the account. Your own autosave was racing your own run trigger. Both now take turns, so that state only appears when something really did change your canvas somewhere else.
- A run always executes the canvas as it stood when you pressed Run, so editing during one is fine. The toolbar now says “edits apply to the next run” instead of leaving you guessing.
- Undo, redo, paste, duplicate, deleting with the keyboard, renaming a node, removing an edge, reordering ports, every group action, and converting results to input nodes did not mark the canvas as unsaved. Each one usually got carried along by whatever you changed next, but on its own it could survive right up until you reloaded, and then vanish. All of them save now, as does a workflow the copilot builds for you.
- A save that finished while you kept typing used to report “Saved”. Anything typed during it wasn’t in that save, so closing the tab in the next few seconds lost it. The status stays honest until your latest edit is actually stored.
- A failed save retries by itself instead of sitting behind a red label, and offers “Retry now” if you don’t want to wait. Losing your connection says so plainly, keeps your work on the device, and sends it up the moment you’re back online.
- If another tab or a connected app saves the canvas while you’re editing, Orisu now asks which version to keep instead of telling you to reload. Keeping yours files theirs in version history first, so it’s still recoverable.
- You also find out before it becomes a problem: a “Changed elsewhere” note appears with an option to catch up, rather than waiting for your next save to be refused.
- A rejected generation used to show up as “Unprocessable Entity” and nothing else. The node now tells you what the model objected to, which input caused it, and what to change. If a model blocks your source video for showing a real person’s likeness, that is what the node says, and it points at the video input.
- The message also tells you whether retrying is worth it. A busy model or a timeout means run the step again. A content-policy block, or a setting the model won’t accept, means change something first.
- The full message is still under “View details”, and the copy button still copies it if you need to send it to us.
- App mode, the asset editor, and runs started from the API or a connected assistant all show the same wording.
- ChatGPT can now connect to Orisu over MCP. Connecting used to fail right after sign-in because our OAuth discovery metadata strayed from the spec in ways ChatGPT rejects; the metadata is now spec-correct for strict clients. If an earlier attempt failed, remove the connector in ChatGPT and add it again so it refetches the metadata.
- Runs priced per second of video (dubbing, for example) reserve credits based on the input’s length. When the length wasn’t known yet, which was common for a run started right after an upload, the reserve fell back to the one-hour ceiling: roughly 4,500 credits held for a 5-second clip. Orisu now reads the duration from the video file itself when it reserves, so the hold matches the clip. That 5-second dub now reserves about 6 credits. The final charge was always based on actual length, but the oversized hold could block the run entirely on a smaller balance.
- Refresh the studio while a run is in flight and the canvas now catches up: nodes that are running show as running instead of sitting idle, and progress keeps streaming from where the run actually is.
- Pressing Run responds immediately. The run shows as queued right away, the toolbar reads “Starting…” while it waits for a worker, and a notice appears if it stays queued for more than 30 seconds.
- Scrolling pans the canvas, pinch or Cmd+scroll zooms, and dragging with the middle mouse button pans. Press F to fit the whole graph in view. The keyboard shortcuts dialog lists the new bindings.
- A bug left the Product from Website node outside the billing gateway, so extractions ran without drawing credits. Extractions are now billed like every other AI step, and their cost appears in run estimates.
- Runs triggered through the API or a connected assistant accept an optional
max_creditscap. When a run reaches it, execution stops before the next billable step: everything generated so far is kept and billed, and nothing else runs. Assistants set the cap from the cost estimate, so a fan-out that turns out bigger than planned can’t drain your balance. - Assistants can also seed all of an agent’s input values in one call (
set_input_values) and mark an edge withloop: trueto run a step once per item of an incoming list. Both save several round-trips when building a workflow from chat.
- Merge Videos now accepts a single connection when it carries a list of clips (the output of a fan-out), with a reminder to check the item count instead of a blocking error.
- Text added to the Composer starts out presentable: centered, white with a dark outline, and sized to the canvas, instead of tiny black text in the corner. The Composer also warns when a video is connected while the canvas is still set to a square preset.
- Paste a product URL and the node reads the page and pulls out the product: name, tagline, description, features, price, audience, product images, and the logo. See Product from Website.
- Generate Text, Generate Image, and Generate Video have a new Product input. Connect the extracted product and its facts flow into the prompt automatically; product images and the logo double as reference images for image and video generation.
- An optional instructions field steers the extraction (“Focus on the enterprise plan”, “Extract the copy in German”). Connect a list of URLs to process many products in one run.
- We re-verified every model’s credit price against the provider’s current list price. Kling 3.0 and Kling O3 generations now cost roughly half of what they did; Kling 2.1, 2.5 Turbo, Elements, O1, and motion control were also corrected, along with FLUX 2 tiers, iC-Light, Flux Fill, camera angle changes, vectorization, and auto-captions.
- Cost estimates in the studio, API, and connected assistants all use the corrected rates.
- The Duration field on Generate Video is a slider that snaps to the clip lengths your selected model supports: 4s / 6s / 8s on Veo 3.1, 5s or 10s on Kling 2.x, 3–15s on Kling 3.0, and so on. You can no longer type a length the model would reject.
- Switching models carries your duration over to the nearest supported value. Models with a fixed clip length show it as a read-only value, and models that take their length from the source video hide the field.
- Runs from the API or a connected assistant get the same treatment: an unsupported duration is adjusted to the nearest supported value before the model is called, instead of failing at the provider.
- Copilot conversations now survive a page refresh. The panel remembers the last conversation for each agent and picks it back up automatically when you return.
- A history view in the chat panel lists your past conversations for the agent, with titles and timestamps. Open one to continue it, or start a new chat right from the history screen.
- Errors in the chat now read like sentences instead of raw error codes, and running out of credits opens the upgrade dialog directly.
- Connected assistants (via MCP) can now run an agent up to a chosen node: the target node and everything it depends on execute, unchanged upstream steps are reused from cache, and nothing downstream runs. Use it to verify a mid-graph step cheaply before committing to the full pipeline. See the
trigger_node_runtool in the tools reference. - A new
preview_nodetool runs a single logic node (Prompt Concatenator, Split Text, If/Else, and friends) in isolation — zero credits, no side effects — so an assistant can check what a node produces before wiring it in. - Assistant edits are now safer under concurrency: node edits are checked against the latest saved revision, so a simultaneous edit from the studio fails cleanly instead of being overwritten.
- Cost estimates now account for fan-out: a node that runs once per item of an upstream list is priced per item instead of being undercounted as a single run.
POST /v1/assets/uploadaccepts amultipart/form-datafile (up to 100 MB) with your API key and returns an asset reference you can pass straight into an app run’s upload inputs — the raw-bytes alternative to hosting the file at a URL first.
- Drag a node into a group and it becomes part of it — from then on it moves with the group. Drag it out to detach it. Moving the group itself still carries everything inside, and the frame adjusts to fit its members.
- Groups and sticky notes no longer count as nodes: workflow node lists, share pages, template details, and cost estimates count only the nodes that actually run.
- Pasting multi-line text into prompt and instruction editors now keeps your line breaks.
- Failed nodes show a clearer, more compact error card on the canvas.
- Share links created by connected assistants now use the app domain and open correctly.
- Graphs built by connected assistants no longer overlap nodes; auto-layout accounts for each node’s real size, including groups.
- Brand kit extraction no longer times out on sites that keep network connections open (chat widgets, analytics).
- Provider and integration logos are readable in dark mode.
- Every save is now checked against the latest version of your workflow on the server. If the same agent was edited somewhere else in the meantime — a second browser tab, a teammate, or the copilot — the studio shows an “Updated elsewhere” notice with a one-click reload instead of overwriting that work. Your local edits are kept safe for recovery after the reload.
- Run always executes exactly what you see. Pressing Run now saves the canvas and starts the run in a single step, so there’s no window where a just-typed prompt could miss the run.
- Claude Fable 5, Claude Opus 4.8, and Claude Sonnet 5 are available on all text nodes. Opus 4.8 comes with a Fast tier — the same model at higher output speed for a higher per-token price — selectable right on the node. Opus 4.5/4.6 per-token pricing was also corrected downward to Anthropic’s current rates.
- GPT-5.5 (with a Pro tier), GPT-5.4, and GPT-5.4 Mini join the OpenAI family, and Gemini 3.5 Flash and Gemini 3.1 Flash Lite join Gemini.
- Gemini Omni Flash (Google) is a new video family: text-, image-, and reference-to-video with synchronized native audio — plus conversational video editing on the Edit Video node, so you can replace objects or restyle footage with a plain-language instruction instead of regenerating the clip.
- Seedance 2.0 gains a Mini tier: roughly half the per-second price of Fast, capped at 720p — ideal for drafts and high-volume iterations.
- Nano Banana 2 Lite joins the image lineup: sub-2-second text-to-image at a fraction of Nano Banana 2’s price.
- Quality tiers (Standard / Fast / Pro / Mini / 4K) now reliably route to the matching provider endpoint and price, everywhere a model offers them.
- Model flavors that used to be separate picker entries are now quality tiers on one model: Veo3 and Veo 3.1 (Standard / Fast / Lite), Kling 2.1 (Standard / Pro / Master), and Kling Elements (Standard / Pro). “No Audio” duplicates are gone — audio is the generate-audio toggle. Existing workflows keep running on whatever they had selected.
- New defaults: fresh nodes now start on the current generation — Nano Banana 2 for image generation and editing, Kling 3.0 for video generation, Gemini Omni Flash for video editing, and Gemini 3.5 Flash for text.
- GPT Image repair: GPT Image 1/1.5/2 requests were failing at the provider; they now run reliably, image edits receive your source images, and aspect ratios map correctly.
Creative assistants for text nodes
- Generate Text now offers a library of creative assistants: pre-filled starting points like Nano Banana prompter, Kling cinematic prompter, ElevenLabs dialogue prompter, Extract brand guidelines, and Ad copy & hooks. Each sets an expert system prompt, a suggested model, and sensible defaults so you can skip the prompt engineering.
- Pick one from the node picker (hover Generate Text to browse them by category, then click or drag) or from the Creative assistant dropdown in a node’s config panel. The node is named after the assistant, and your edits to the instructions always win.
- Text Assistant is retired. Its specialized modes now live as creative assistants on Generate Text, in one place with a fuller library.
- Connect multiple lists to a node and it now runs across every combination automatically: 5 prompts × 3 images produces 15 distinct results, all in one click. The node shows the multiplication (
5 × 3 = 15) before you run so there are no surprises. - Partial success: a single failed result no longer throws away the rest. The node completes with however many results succeeded; failed items are flagged individually in the iteration strip.
- Per-result inspection: each result in the strip now shows which input combination produced it, along with the model used, generation time, and credit cost for that item.
- Smarter re-runs: results that haven’t changed are reused rather than regenerated. Edit one input and only the affected combinations re-run; unchanged results are never re-billed.
- Interactive crop: select an image or video layer, click ✂ Crop, and drag a marquee over the full source. Trimmed edges dim as you go, and the result matches the render exactly.
- Snapping + guides: layers snap to the canvas edges/center and to other layers while dragging, with blue alignment guides.
- Align buttons in the Layout panel: one-click align to canvas left/center/right and top/middle/bottom.
- Rich text: text layers gained alignment, line height, an outline (color + width), and a drop shadow (color + blur).
- Resizable panels: drag the dividers beside the Layers and Properties panels (or double-click to collapse/restore) to give the canvas and timeline more room.
- Spacebar toggles timeline playback when the composition has video or audio.
- Illustrated effect pickers: Fit, Ken Burns motion, and entry/exit animations are now visual tiles that preview the effect on hover, and the layer panel is reorganized into Layout / Effects / Timing with slider controls for opacity and volume.
- Media pulled in by the Instagram, TikTok, LinkedIn, and Meta Ads scrapers (and frames captured by Extract Video Frame) is now classified as internal reference material instead of a workflow deliverable. It stays in your asset library but no longer shows up in dashboard output feeds.
- Files brought in via Google Drive Import are now classified as source material, alongside uploads.
- The More filters popover now lists your agents as searchable cards (thumbnail, name, and when they were last edited) and your workspace tags as clickable chips. No more pasting ids or comma-separated text.
- Filtering down to zero results keeps the header and filter bar on screen (they used to disappear), so you can always adjust or clear filters. Active-filter chips now show the agent’s name.
- Instagram, TikTok, LinkedIn, and Meta Ads scrapers now emit a single Results output instead of separate Posts, Images, Videos, and Summary ports. Connect it to any AI node’s Context input in one edge.
- Generate Text, Text Assistant, Generate Image, and Generate Video all gained a Context input that accepts anything, including scraper Results.
- Items in the bundle get stable reference names (
[post_1],[image_2],[text_1]). Use them in your prompt to target specific pieces (e.g. “Rewrite the caption ofpost_2for our brand voice”). - Images in the bundle attach natively for vision-capable text models; existing graphs migrate automatically.
- New users now get an interactive demo in the studio: Orisu assembles a three-node workflow before your eyes (your name, an AI prompt writer, and an image generator), then runs it to create a personalized origami avatar.
- Skip anytime; the example workflow stays on your canvas either way.
- Cover/Contain fit ends accidental stretching: choose how media fills its box, plus per-side source cropping.
- Ken Burns motion: slow zoom in/out across a layer’s visible window, adding motion to image slideshows.
- Crossfades in Sequence mode: set an overlap and clips dissolve into each other instead of hard-cutting.
- The timeline zooms (buttons or Ctrl+wheel), video bars show real filmstrip frames, audio bars show waveforms, and audio tracks can fade in/out.
- Playback respects the export range, and a repeatable full-pipeline render smoke now guards the whole ffmpeg graph (full render in ~0.5s, 1080p/10s benchmark ~7s).
- Trim any clip by dragging the edges of its bright timeline bar (the darker region shows what’s cut), and trim the whole exported video with bracket handles on the ruler (or Video Range in the canvas panel).
- Layers animate in and out: Fade or Slide (left/right/up/down) with per-animation durations, previewed live.
- Reorder layers by dragging directly on the timeline, and a
?next to the Overlay/Sequence toggle now explains the two modes in place. - New guide: Composing video covers the whole workflow, including the JSON config for API and copilot use.
- A play button on the timeline previews the whole composition in the editor: real video frames with sound, audio tracks entering on cue, fades, and timed layers appearing and disappearing.
- Scrubbing now seeks video layers to the exact frame instead of showing their first frame.
- A new Sequence arrangement plays video clips back-to-back in layer order, with each clip’s length resolved from the actual video at render time. It works even before upstream videos have been generated, and regenerated clips can never drift out of sync.
- Timed clips now play from their own first frame at their start time (previously they joined mid-stream), Trim skips into a source, and a clip with no end time plays out its full length.
- Timeline bars snap to other layers’ edges, show where each source ends, and display estimated lengths (with a note) until the workflow has run once.
- Connect audio nodes to add music or voice-over to a composition. Tracks get their own timeline rows with timing and per-track volume, and everything mixes into the output video.
- Video layers’ own sound is now included (compositions used to render silent) and follows the layer’s timing window.
- Visual layers can fade in and out: pick a duration in the layer’s Layout panel, with fades anchored to the layer’s enter/exit times and visible in the scrub preview.
- A timeline now appears below the canvas when a composition includes video: one row per layer, with draggable bars that set when each layer enters and leaves the video.
- Scrub the ruler to preview exactly which layers are visible at any second before rendering.
- The output video automatically extends to cover the latest-ending layer, and exact start/end seconds can be typed in the layer’s Layout panel.
- Layers wired from AI generator nodes now show their actual outputs in the visual editor instead of gray placeholders.
- Rendered output now matches the editor faithfully: rotation pivots, text rotation and word-wrap, font weights, text backgrounds, and aspect-ratio-preserving defaults (no more stretched images).
- Video compositions no longer scramble layers when an input is disconnected, and text layers land exactly where you placed them.
- Smoother editing: cursor-anchored wheel zoom, arrow-key nudging, a regrouped properties panel, and far less lag while typing values.
- A new utility node saves a run’s images into a reusable Element as part of the flow: feed it a photo plus generated variations and it writes them to your library, then passes the element on for immediate use downstream.
- Create a new element each run, or update an existing one (append to or replace its gallery).
- Works in published apps too: build elements automatically from your users’ uploads.
- The empty Element node now shows a fanned stack of recent library covers instead of a dropdown.
- Picking an element opens the full library browser (sidebar, filters, search) in select mode.
- Polished element cards, a friendlier empty state, and a redesigned create dialog with a visual subject-type picker.
- The node now returns real ad creatives (images and videos) directly from the public Meta Ad Library.
- New Country and Media Type filters let you narrow results to a specific market or creative format.
- Searches are consistent and repeatable: the node maps your settings directly to an Ad Library search instead of improvising one.
- All three nodes now return real public posts with actual creatives: post images, video covers, and video files where available.
- Post records include engagement data: likes, comments, plays, and shares (where the platform provides them), plus author identity (name, handle, avatar) and timestamps.
- Redesigned result cards show author identity, compact engagement counts, and a relative timestamp alongside the post creative.
- Searches are consistent and repeatable: each node maps your query and period settings directly to a public search instead of improvising one.
Element Library
- Build a library of reusable subjects (characters, products, and other recurring elements) and drop them into your agents.
- Drag an element from the Library straight onto the canvas, or pick one in the new Element node.
- Create elements from uploaded images, or right-click a selection on the canvas to turn it into an element.
- Edit an element’s images, attributes, and cover image from its detail view; reorder images by dragging.
- Supported image and video generation nodes accept elements as inputs, so a subject stays consistent across everything you generate.
- Publish your own elements to the public library for others to use.
- The Assets page is rebuilt as the Specimen Archive: your generated images, video, and audio organized into shelves you can browse and filter.
- Bookmark favorites and filter by date.
- A Trash drawer lets you restore deleted assets or remove them permanently.
- A new
/exploregallery surfaces assets with a full-detail specimen view.
- Generation nodes show a live estimated credit cost that tracks the selected model and its settings.
- A tier selector sits under the model picker so you can choose a quality tier per node.
- Nodes with multiple outputs render results in a carousel.
- Failed nodes get a clearer error state with copy, expand, and retry actions; skipped nodes explain why they were skipped.
- A new origami-themed onboarding wizard walks you through your first steps.
- Active users see persistent discovery sections on the dashboard.
- Fixed a label that showed free-plan users as Pro.