LoRA Dataset Studio v2026.08.07

Two ways to install — and they do not update the same way

git clone is the one to prefer. Update & restart then runs git pull --ff-only,
so you get every fix the moment it lands — often days before it is packaged into a
release like this one.

This ZIP is a snapshot of this tag. Update & restart will only ever move you to
the next release, so a fix shipped today reaches you whenever the next one is cut.

git clone https://github.com/perfectgf/lora-dataset-studio.git
cd lora-dataset-studio
start.bat

Already running a clone? You do not need the ZIP below.

🎁 What's new in v2026.08.07

The dataset page uses the whole screen again on a phone

On a phone (and in the installed app), the dataset page drew everything at about three quarters of the screen width — the header bar, the section chips, every card — with a dead strip down the right-hand side. Nothing was cut off and nothing looked broken up close, which is what made it hard to place: the culprit was a one-pixel, invisible label that the horizontal section rail was failing to keep inside itself, which stretched the page wider than the screen and made the browser shrink the whole thing to fit. The rail now holds its contents, so the page measures the width of your screen and uses it. The Settings and Guide chip rails got the same treatment before they could catch it.

Every caption now says who wrote it

The default captioning backend is a chain, not a choice: JoyCaption writes what it can, then the Ollama vision model covers the images it missed — and both halves used to land in the same box looking identical. Captions now carry their author everywhere you read one (bank tile, dataset tile, the expanded editor, the identity-leak list), your own hand-written captions stay marked as yours, and a bank caption run ends with "340 by JoyCaption, 87 by the Ollama vision model" instead of a bare total. On a bank that ✨ Score has found to be mostly NSFW, the caption window now also says what was measured on that kind of image before you start.

Every "What's unlocked" row now takes you to its install

The Krea 2 Edit row told you it was missing and did nothing when clicked — it now opens the install screen where its one-click installer lives (or, when the files are already there, tells you to restart ComfyUI). Clip encoding joins the list too: a machine with no ffmpeg used to be certified complete while it could not export a single clip. And the scraper install warning finally names the packages YOUR machine is missing, instead of reciting three from memory.

Setup now lists every engine you can install

Bank scoring, the SigLIP2 semantic engine, the fast watermark detector and the scraping extras (gallery links & keyless web search) could already be installed with one click, but never showed up on the wizard's "What's unlocked" screen — so finishing Setup could look complete on a machine still missing four of them. All four now get their own row, their own install card, and their own ✓/✗.

Web image search now flags a missing install correctly

If your scraper packages were installed before web image search (DuckDuckGo-style keyword search) shipped, the app used to say the scraper was ready and only fail once you actually searched. The Web images tab now correctly flags it as needing a reinstall, with a one-click "Install scraper extras" button to fix it.

Reddit "Load more" no longer skips posts

On image-heavy subreddits, a batch of posts could hit the per-scan cap mid-way through — and the posts after that point were gone for good, on every page. "Load more" now picks up exactly where the cap cut off, so nothing is left unreachable.

A capped scan says there might be more

Picazor, RedGifs and Instagram scans stop at a built-in limit to stay fast. Hitting that limit used to look exactly like a profile that genuinely had that few posts. Now the picker flags it, the same "some images may be missing" banner used for a scan cut short by a block or a timeout.

A blocked page says so — even on a slow scan

A gallery scan that hit both a login wall and the time budget used to report "nothing here" instead of the real reason. It now shows the actual block. And when a scan is cut short by the time limit, the picker tells you some images may be missing instead of quietly presenting a partial list as the whole thing.

Images scraped into a bank keep their source

Scraping straight into a bank used to drop the Pexels credit or the page a web-search image was found on. It now survives the trip — and shows up again if you later promote that image into a dataset.

See where a web-search image came from

Images you pull in from web search now carry a "Source ↗" link back to the page they were found on, right in the dataset — so you can credit or double-check the origin without hunting for it again.

Paste a gallery from almost any site

Sites without built-in support used to return a single image. They now list the whole page, using the same engine that already powers PornPics and Civitai. A site the engine does not know shows "No images found" instead.

Search the web for images by keyword

Type what you are looking for instead of hunting for a gallery URL first. No account, no API key. Results come from third-party sites, so check the licence before using one.

Pinning a run to the canvas gives each prompt its own grid

Pinning everything a run produced onto the ◉ LoRA Canvas used to drop the pictures of several prompts into one another, so a board carrying two or three runs read as one scatter and you could not tell at a glance which image came from which prompt. Each prompt now keeps its own grid, and a fresh grid is moved aside rather than laid over anything already on the board — including the cards, the checkpoint pills and the pictures you placed by hand, which are never touched. The reason for putting a run on the board at all is to SEE which prompt gave what; that is now what you get.

The duplicates figure counts what you still have to decide

The ≈ Duplicates figure on the bank overview counted every group the last analysis found, including the ones you had already resolved — so it sat at the same number no matter how much of the queue you cleared, and there was no way to tell progress from a stuck pass. It now counts what is left to decide, and says so: “groups remaining to resolve”. A bank whose duplicates are all sorted reads zero, which is the whole point of a headline number.

The face-mask preview can be stopped — and picks up where it left off

"Looking for faces… analyzing image 4 of 153" had no way out: the only choice was to wait for the whole pass or leave the page and lose it. There is a Stop button now, and it keeps what the pass already found — start it again and it resumes at image 5 instead of image 1. The button says what stopping costs at the moment you press it, and the cost changes as the pass runs: during the model load nothing has been analyzed yet, so only the load is given up; once images are being analyzed, every face already found is kept. Change your kept images and the saved work is dropped rather than reused, because boxes from photos that left the set would describe a run that no longer exists.

Setup now counts the video pieces — and its repair menu can reach them

The setup wizard could certify "12 of 12 capabilities ready" on a machine whose Video bank could not open a single file: the two video pieces were not counted, not listed in the Install-or-repair menu, and the wizard skipped its own install screen because everything it DID count was green. The summary now counts 14 — reading video files and shot detection included, each ✗ row clickable to where it installs — and 🎬 Video decoding and 🎞️ Shot detection sit in the Install or repair individually menu like every other component.

Shot detection no longer fails every file right after a clean install

The 🎞️ Shot detection install put the model in place but not the decoder it reads files with — so the install reported success, the readiness badge turned green, and then every single file answered "failed shot detection". The install now carries PyAV into the same environment, the badge only turns green when the worker can actually open a file, and a contract test holds the three ends (worker, installer, probe) to the same list. If you hit this: Setup → 🎞️ Shot detection → ↻ Reinstall, then run Find shots again — your files were never the problem.

The Bank page opens in a blink instead of a second

Every time you opened 🗃️ Bank, the app re-listed every image in every bank's source folder before drawing the page — on a real library of 86 000 images that was a full second of disk work, paid even when you were only passing through. It no longer does: the page opens on what it already knows. A folder is still re-checked automatically when you OPEN its bank, so images you dropped in it show up as before, and a new 🔄 Rescan folders button on the list checks them all at once. The page says how fresh its counts are, and a folder that went missing (unplugged drive, renamed folder) is still flagged straight from the list.

Build the SigLIP 2 index on the GPU you already have

✨ Score could already borrow a CUDA Python from ai-toolkit or ComfyUI instead of crawling on the CPU. The SigLIP 2 semantic index could not, so it kept indexing in the app’s own CPU-only environment even on a machine with a 4090 sitting idle. The Semantic engine panel now says which device the index will really use, and offers the same picker. It checks each interpreter dependency by dependency — and the list is the right one: SigLIP 2 never needs OpenCLIP or timm, so a ComfyUI venv Score refuses can be perfectly good here. Nothing is ever installed into an environment the app did not build: the pinned weights live in your data folder, and Setup ▸ Quality tools keeps installing into the app’s own Python even after you borrow one.

The video extras can now be installed where the app said they were

The Video bank's banner told you "Install the video extra from Setup" — and Setup had no such button: both installs existed, but only for the API. Setup's optional-helpers step now carries the two missing cards: 🎬 Video decoding (PyAV + bundled ffmpeg into the app's own Python, no torch) and 🎞️ Shot detection (TransNetV2 into the scoring Python, CPU is fine). One click each, live progress, and the banner clears without a restart. A new test holds Setup to every install the video banners promise, so a pass can no longer point at a button that does not exist.

Spot the watermarked shots before they teach your LoRA a logo

Rushes come off stock sites and other people's uploads, and a logo sitting in the same corner of every frame is the most consistent thing in your dataset — so it is the first thing a LoRA learns to draw. You cannot catch that by scrolling 90-pixel thumbnails. The new 🔖 Watermarks pass runs the same detector the image bank uses over each shot's sharpest frame and flags what it finds. Nothing is deleted: it is an amber flag you can filter on and act on. Needs the watermark detector from Setup; the cut sits in 🎚 Quality cuts at the measured 0.94, and a shot the pass has not judged is never called clean.

Find the takes you already have, without watching them twice

Ten near-identical takes of one gesture do not teach a model ten things — they teach it one thing ten times as loudly, and that is how a LoRA ends up unable to do anything else. The new ✂ Duplicates pass compares your shots to each other and groups the near-identical ones, keeping the sharpest of each pile unflagged so you know which one to keep. It costs no GPU and no waiting: it reuses the frame vectors 🔎 Find scenes already cached, so it is dot products over a file you already have. Flags only — nothing is rejected or deleted for you.

Act on a quality flag instead of just reading it

The amber flags in a video bank could be read one shot at a time and nothing more. There is now a row of chips above the gallery — "Barely moves (14)", "Same as another shot (31)" — and pressing one narrows the grid to exactly those shots, so you can select them and reject the lot in one gesture. The counts cover the shots currently loaded and the row says so when there are more to load.

Half-second flash cuts stop cluttering your triage

Shot detection deliberately keeps very short cuts — a real flash cut is a real shot, and a detector that hides them also hides genuine boundaries. The price was a grid full of half-second shots you scrolled past over and over, and that could never reach a dataset anyway. 🎚 Quality cuts now has a "Minimum length" field: type 1 second and every shorter shot wears an amber flag you can see and sort by. It is the one cut that works straight after detection — it reads the shot bounds, so you do not have to run the measuring pass first — and Preview tells you how many it would flag before you apply it. Nothing is deleted: it is a flag, like every other cut in that panel.

Your video sets can now be trained here, without leaving the app

A promoted video set now carries a ▶ Train this dataset button, and it hands the clips straight to the ai-toolkit already installed on your machine — no export, no copy, no config to write by hand. It shares the GPU with everything else honestly: a captioning pass or a ComfyUI render in flight refuses the launch instead of fighting over the card, and an image training already running blocks it exactly as another video run would. MiniMax H3 is wired in alongside Wan, with the quantisation, the noise schedule, the audio flags and the guidance its own trainer actually expects — a mismatch there does not crash, it just trains a slightly wrong model, which is why each value was read in the installed trainer rather than guessed. Two things it refuses to do quietly. H3 needs about 43 GB of weights: if they are not on your disk the button says so, names the repository and the size, and waits for a yes rather than turning into a silent overnight download. And a set re-promoted to a different target is refused rather than resumed, because the run folder still holds the previous model’s LoRA. Wan 2.2 is the one target a finished run has been through here; the card says plainly which of the others are wired but not yet proven.

A promoted video dataset can now be trained on a rented GPU without leaving the app

Open a video training set in your library and it has a Cloud training panel: pick a step count, click Train in the cloud, and watch the run — the GPU it landed on and what it costs per hour are on screen as soon as the pod has them. When it finishes, the checkpoints are listed by step and downloadable from the same panel. Wan 2.2 saves TWO files per step, the high-noise and low-noise experts, and either one alone is a LoRA nothing can load — so a step is offered as a step, with both of its files, never as a loose file list. A run that failed has a Retry, a run that finished has Train further, and continuing a Wan run sends both experts back to the new pod rather than resuming half of it. Before the job starts, the pod is asked to decode one of the clips you just uploaded with the same decoder the trainer uses; a pod that cannot read them fails in the first minute instead of billing you for hours of training on nothing.

Captions can now speak plainly instead of describing around the subject

Next to 🗣 Describe shots there is now a Caption wording choice. Standard is the wording that shipped and stays the default. Plain gives the model explicit permission to name what is actually on screen rather than reaching for vague stand-ins. This came out of a measurement, not a hunch: on real footage, four combinations were compared and the WORDING mattered more than the model did — the stock model asked plainly named things precisely and wrote the best action description of the four, while an uncensored model asked the old way still described around the subject. That matters because a caption that talks around its footage is a dataset defect you cannot see: the text reads perfectly well, the training set looks complete, and the LoRA learns the evasion. Every caption now records which wording produced it, so a bank captioned across a change is still one you can reason about. Pick it per run, or set it once in your config.

Captions can now speak plainly — the model that writes them is yours to choose

The 🗣 Describe shots pass had one model wired in. On a real corpus that turned out to be a dataset problem rather than a matter of taste: a captioner that describes what it sees in evasive terms produces captions that are about something slightly other than your footage — and a LoRA trained on those learns to look away too, with nothing in the output to reveal it. The captions read perfectly well; they are just not about the shot. So the checkpoint is now a setting, video_caption.model. Leave it empty and nothing changes: the same model as before, the same captions. Point it at any checkpoint of the same architecture and the pass uses that instead. Two things come with it. If the model is not on your machine yet, the pass SAYS so in its own progress line before it starts, because the first run downloads it and that should never be a silent twenty-minute wait. And every caption now records which model wrote it, so a bank captioned half before the change and half after is still a bank you can reason about.

Your shots get described — so you can search for what HAPPENS, and so they train on words

There is a new 🗣 Describe shots pass. It watches eight frames spread across each shot and writes what happens in it — “a woman turns and walks away”, not an inventory of objects — and that one line does two jobs. It becomes the clip’s .txt sidecar at promotion, which IS the prompt it trains on: until now every promoted clip shipped with an EMPTY prompt, and the trainer accepts that in silence. And it makes 🔎 Find scenes able to answer a question it structurally could not before. CLIP looks at frames, so it finds what a moment LOOKS like; an action is a fact about time and no single frame carries it. With captions the search reads both, and the panel says which halves are running so an empty result can be read correctly. Captions are drafts: open any shot and edit it, and a bulk re-run will not overwrite what you wrote. The promotion now also tells you how many clips are about to ship with no caption at all, before it encodes anything.

The per-source cap you could already read about now has a knob

The build dialog gains a “Max clips per source” field. The cap itself is not new — it has capped nothing so far because nothing could send it: the setting was implemented and reachable from neither the dialog nor the API. It matters because a 50-clip set that is three videos over-represented looks exactly like a diverse one on disk. Leave it empty for no cap. Each source keeps its EARLIEST clips, so promoting the same bank twice gives the same dataset, and a source with fewer clips than the cap keeps all of them — it trims dominance without punishing scarcity. And when a finished set turns out to lean on one file anyway, the result now says so with the real share instead of leaving you to notice.

Trim the dissolve off both ends of every clip you export

A shot boundary is where a cut just happened, so the first and last frames of a detected shot are disproportionately dissolves, fades and leftovers of a transition — and a dataset whose clips all open on half a dissolve teaches the model to open on half a dissolve. The build dialog now has a “Trim each end” field: a number of seconds taken off BOTH bounds of every clip. 0.25 is the common figure; the default is 0, so an existing recipe exports exactly what it exported before. What it will NOT do is hand you a short clip. Frame counts are a property of the target model’s VAE, and ffmpeg happily writes a 32-frame file and exits 0 when asked for 81 — so a clip that no longer supplies the count is dropped rather than exported short. The dialog tells you how many clips the trim will cost before you press the button, and counts them separately from clips that were never long enough: only the first kind is fixed by lowering the trim, and reporting them as one number is how a setting quietly halves a dataset while the material looks to blame.

Your shots are now listened to, not just looked at

For LTX and MiniMax H3 the source’s audio is muxed into every clip you export — and until today nothing had ever listened to it. A shot whose track is a silent stretch, a dropout or a muted camera passed exactly like a shot with sound, because the file on disk is the right length, the right sample rate, and mute. A dataset of silent clips teaches the model to be silent. Measure now reports, per shot, how much of it is silence and its overall level in dBFS, with two new cuts to go with them — Silent share and Loudness floor — raising two different flags on purpose, since a quiet clip can be normalised and a silent one cannot be rescued. Three states are kept strictly apart, because collapsing any two of them makes the bank lie: a file with NO sound track is never flagged (Wan datasets are supposed to look like that), a track that is there and carries nothing is the actual defect, and shots measured before this shipped have no sound reading at all — an audio cut will never flag those, and Measure with re-measure is what fills them in.

Type a word and find the scene, in a folder of rushes with no names

A video bank is a haystack whose needles have no names: quality cuts tell you which shots are sharp and which move, and nothing tells you which one has the red car in it. There is now a 🔎 Find scenes box above the gallery. Run the pass once — it looks at a few frames of every shot — then type “a woman walking on a beach” and the gallery is replaced by the shots that look most like it, best first, instantly. Several frames per shot on purpose: a car that only drives into view in the last second would be invisible to a search that had looked at the opening frame, and you would get no hint it had been missed. So every shot contributes a frame near its start, its sharpest frame and one near its end — and every result tells you WHICH SECOND matched, with the player opening right there. Two things it says out loud rather than hiding. It is a ranking, not a filter: every shot scores something against every phrase, so the results always come back full and the line above the gallery tells you how strong the top and the tail really are, plus how many shots have not been looked at yet and could not be searched at all. And “without” does not work — ask for a street without cars and you get cars, because the model ignores the word rather than honouring it. Type “-cars” instead: that pushes them down the ranking, which is a promise the app can actually keep.

A badly cut shot is no longer a shot you have to throw away

Until today, a shot the detector cut one second too early — or one holding a frozen tail — had exactly one available gesture: ✕ Reject, which threw away the eight good seconds to be rid of the bad one. Open any shot and there is now a ✂ Trim & split panel under the player. Nudge either bound by one second or by one frame (one frame OF YOUR FILE, at its own rate), snap a bound to wherever the playhead sits, split a shot in two at the playhead, or draw a shot the detector missed entirely — scrub anywhere in the rush and press ＋ New shot from here. Splitting keeps the decision you were making: split a kept shot and both halves stay kept, so you do not have to find them again among hundreds. One thing worth knowing before you trim: for image-to-video targets, the trainer conditions on the clip’s FIRST frame. Moving a start is therefore not trimming — it is choosing the exact image the model learns to animate from, and the panel says so where the buttons are. Re-cut shots lose their thumbnail and their quality scores on purpose: a thumbnail of a frame the shot no longer contains is not stale, it is wrong. Run Make thumbnails again when you are done cutting. And re-detecting a file no longer destroys the cuts you made by hand.

Your video bank now measures every shot — and tells you which ones to look at

One pass reads every frame of every shot and scores what quietly ruins a video dataset: shots that barely move, shots that are all blur, black moments, frozen stretches. Nothing is rejected for you — flagged shots get an amber mark in the grid, and the verdict stays yours. The cuts are yours too: there are deliberately NO default thresholds, because the same number that flags 2% of one bank flags 12% of another. Open 🎚 Quality cuts, set a value, and Preview shows exactly how many shots each cut would flag — per rule, before anything is applied. On a real 4.5-hour test bank the most valuable filter turned out to be the frozen-stretch one: 15% of shots carried a freeze the average could never see. Bonus: thumbnails now come from the SHARPEST measured frame instead of the middle guess.

Your folder of rushes is now a training set — and a .mp4 is no longer ignored in silence

Drop a video into an image bank and until today it was skipped without a word: no row, no warning, nothing to click. Videos now get their own bank. Point it at a folder of rushes and it cuts every file at its shot boundaries, so you triage SHOTS instead of files — a two-hour rush becomes three hundred things you can judge in an afternoon. Click any shot to watch exactly that moment; the grid stays thumbnails, so a bank of hundreds of shots stays as light as a page of photos. Nothing is copied and nothing is re-encoded while you triage: a bank stores where each shot starts and ends, and only the ones you keep are ever encoded. When you build the set, the length menu offers only the frame counts your target model can actually ingest — 29 frames is legal for Wan and illegal for LTX, and no trainer tells you, they just quietly round it down. Two things are written next to the target you pick, because they are what costs a wasted week: whether a LoRA trainer for it is known to exist at all (exactly one of the four), and MiniMax H3’s licence, which grants no rights in the EU, the UK, South Korea or the USA — outputs included. And if a piece is missing, the app names which one: with no ffmpeg you can still scan, cut, watch and triage everything — only the final encode waits.

Choose which images get cropped or repainted — the two watermark levels no longer fire on one click

🚩 Watermarks offered “✂ Auto-crop (16 052)” and “🧽 Inpaint (16 507)” as single buttons: press one and thousands of images were cleaned, with no way to say which. Both now open the same launch window the other passes use — run them on your kept images only, on the undecided ones, on the bin, on all three, or on exactly the images you ticked in the grid. Each line carries the real number for that pile, a run that would touch nothing is refused instead of reporting success, and the window states up front what is reversible: your own files are never written to, the cleaned pixels live in the bank’s own copy, and ↩ Undo cleaning throws them away and re-flags the images — what it cannot reach is an image you already promoted into a dataset, or one whose source file changed on disk since the clean. A run that only covered part of your bank now names what it left out, and flagged images sitting in the bin are reachable at last instead of leaving a dead button. Leave the window alone and both levels behave exactly as before.

Big Banks open in a second again — and Stop answers while a pass runs

A Bank of 50 000 images took over 12 seconds to load its workspace, and more than 28 while a pass was running — long enough that the progress bar and its Stop button never arrived, so a running scan could not be stopped. The workspace was re-checking the location of every single image on disk, twice each, on every 2-second refresh; it now remembers what it already resolved. The progress bar has also moved to its own tiny request, so it keeps ticking and Stop keeps working no matter how large the Bank or how busy the pass. Measured on a 50 397-image Bank: 18.5 s to 1.5 s, and the progress bar answers in 1 ms.

Each Bank can choose CLIP or SigLIP 2 — switching keeps every result

A Bank now chooses the semantic engine behind Find by text, Similar, diverse and balanced picks, coverage, and same-shot grouping. CLIP remains the zero-change default and still owns Score; SigLIP 2 is an explicit optional install with its own pinned, resumable index. The two caches and the two same-shot partitions are stored separately, restored when you switch back, and carried through Bank → Dataset, Dataset → Bank and Bank → Bank transfers. Live readiness counts now follow the images that are actually in the Bank, so deleting or rejecting the last group cannot leave an old “65 ready” graph behind.

Say what you want in a sentence, and the bank sets its own filters

Type "an amateur photo set, least polished first" above the triage chips and the bank moves ITS controls — medium, flags, resolution, sort — then the counters you already had say, measured, how many images that lands on. The model never looks at your pictures and never picks any: it only reads the sentence, so a wrong reading costs one glance at chips you can edit rather than a silent selection of a few thousand images you would have to trust. It answers over what your bank has actually measured, with the real per-value counts in front of it, so it cannot reach for a bucket that holds nothing. When part of a request has nowhere to land — asking for what is IN the pictures on a bank whose captions cover 1% of it — it says that part is not expressible here instead of inventing a filter that would return three thousand convincing, unrelated images. And it will not turn an exclusion into a search phrase: the ranker in this app returns MORE of a negated thing, not less, so "without a watermark" is reported rather than quietly sent.

Bank passes now say what they were doing, what they grouped, and what they left out

✨ Score used to go quiet for minutes once its counter reached the last image, bar full, Stop still offered — on a 21 000-image bank that silence read as a freeze. It was writing a 70 MB cache and then comparing every embedding against every other; both steps now name themselves while they run, and the ones where Stop costs something different say what it costs before you press it. The style grouping stopped hiding behind a number too: "1 style group of 2+" was printed both when almost nothing grouped and when a single group had swallowed 24 928 of 24 931 images. It now states the size of the biggest group against the total, and names the threshold when the grouping is too loose to separate anything or too tight to join anything. And every pass that skips part of your bank says so at the end: 🎨 Classify medium answering "0 classified" on a 50 000-image bank was exact — 25 464 rejected images had been dropped by the scope before the pass even began, and that figure appeared nowhere. It does now, on 🔎 Scan, 🏷️ Caption, 📐 Framing, 💧 Watermark and 🎨 Medium alike, and stays silent when a run really did reach everything. 🎨 Medium's launch window also counts, per scope, the images it cannot answer for because ✨ Score never reached them — and refuses to start a run where that is all of them.

Every bank pass now asks where to run — kept, undecided, unkept, or your selection

Pressing 🔎 Scan, 🏷️ Caption, 📐 Classify framing or any other pass used to fire immediately on whatever the app had decided was its pool. Each of them now opens a window first, and that window says three separate things. Where the run applies: kept only, undecided only, unkept (the bin), all three, or the images you have selected — each line quoting the number of images IT would actually touch, so a scope with nothing to do is refused before the click rather than reported as a success afterwards. What the pass reads: only the settings its calculation consumes, with where they live, so you stop guessing whether nudging a threshold costs you a rescan. And what is NOT decided there: the knobs that only re-sort and re-flag the grid, retunable any time with no pass at all. "Rescan all" and "Rescore all" were never separate actions and no longer pretend to be — they are a tick box in their pass's window, next to the pool they re-run and with their price written next to them. Three passes refuse a partial scope, and now say so instead of quietly not offering it: ✨ Score, 👥 Group by person and ✂ Find crops & variants each produce one numbering of the whole bank, and handed a slice they would land fresh group ids on top of unrelated groups already saved. Aiming a pass at the bin is offered everywhere it makes sense, always with what that run costs, and never as a default.

The Launch-all report calls every step by the name on its button

Two of the eight steps in the "Last Launch-all run" summary printed their internal identifier instead of their name — you would read "semantic_dedup" and "framing" in your own report and have no way to tell which button that was. They are now "✂ Find crops & variants" and "📐 Classify framing", the exact words on the buttons you pressed. The fallback that made the omission invisible is now guarded by a test that reads the server's own step list, so a ninth step cannot ship nameless.

Select images and their tags are right there, with how often each is cited

Once a bank is captioned, selecting an image now shows its tags immediately — no extra click, no hunting for a badge. Select several and you get the tags across the whole selection, each with how many of them mention it: "red dress 7 / 12" tells you at a glance that over half of what you picked is that, which is the judgement you were making anyway. Ticking a chip filters the bank exactly as before, and the row holds still while it does. It is honest about its own reach: images with no caption yet are counted apart and named, so are captions whose every word was a grammar word, and a selection too large to read in one go says how many it left out instead of quietly shrinking the denominator. The 🏷️ button on a tile has also moved out of the badge cluster in the top-left corner — that corner is a state readout where nothing is clickable — down to the bottom-right next to ▶ and ⛶, where the tile's other actions live. When an image has no caption, or a caption with no word worth filtering on, the button stays visible and says which of the two it is, instead of disappearing and reading as a feature that does not exist.

The numbers on the bank filter chips now match what they open

Every chip in the bank — 🌫 Blurry, 📺 Noisy, the resolution tiers, 🔎 Origin, 📐 Framing, 🎨 Medium, ⤢ Angle — used to print a total for the whole bank whatever else you had filtered. Pick ✕ Rejected, read "📺 Noisy 4 043", click it, get twelve images. The counts now describe the bank you are actually looking at: each one is measured with your other filters applied. The value you already picked is deliberately left out of its own count, so choosing 🌫 Blurry never zeroes 📺 Noisy next to it and you can always switch to a neighbour without clearing everything first. Chips stay on screen even when they hold nothing under the current filter, for the same reason. Two smaller things went with it: clicking a ✨ Score or ≈ Duplicates chip no longer silently drops the person or style group you were in, and the "showing N of M" line now notices 🔎 Origin, the 🚫 exclude box and the 🏷️ tag chips, which it used to ignore.

A ✨ Score pass that could not load its models tells you why

A scoring pass could finish, announce "done", and leave every image without a single number — with sorting by aesthetics, find crops and variants all greyed out because there was nothing to sort on. The closing line named the head that had gone quiet and stopped there: "(aesthetic + NSFW head unavailable)". Both models are fetched over the internet the first time they are used — the aesthetic predictor from GitHub, the NSFW classifier from Hugging Face — so on a machine that cannot reach them, typically a container without outbound network, both go down together and the whole pass comes back empty for one reason nobody was told. That reason was known all along, in the scoring log, and never made it to the screen. It does now: the sentence carries the actual error, so "no scores" reads as a network problem you can fix rather than a feature that looks broken. Reported by @_nofaceman on Discord.

✨ Upscale & improve stops re-downloading a LoRA you already have

On a Docker or Linux install, raising the enhancement-LoRA strength made every ✨ Upscale & improve answer “Klein needs klein_enhancement_lora — I’ve started downloading it”, then download the file it already had, then say it again. The file was never missing: the improve workflow was exported from a Windows ComfyUI and names that LoRA klein\realistic.safetensors, and on Linux a backslash is part of a filename rather than a folder separator — so the app looked for one file with a strange name instead of realistic.safetensors inside klein/. Setup, which spelled it correctly, kept showing it installed, which is why the two screens disagreed and only that one file looked broken while every other model loaded. The name is now respelled for whichever system opens it, on the way in and on the way out to ComfyUI. Nothing to reinstall — the LoRA already on your disk is picked up after an Update & restart. Reported by @_nofaceman on Discord.

Stopping ✨ Score keeps the scores it had already computed

Pressing Stop during a scoring pass used to leave your images exactly as they were: the work reached a cache on disk, but not a single image got its aesthetic or NSFW number. Now everything it computed is written to your images before the pass ends, and the closing line tells you how many were scored, how many are left, and how many were reused from the cache instead of recomputed — so a relaunch is visibly paying only for the rest. The 🎨 style groups are the one thing that still needs a full pass to land, because those ids are a single numbering of the whole bank and half of one would mix two unrelated styles under the same chip; a stopped pass now leaves the previous grouping untouched and says so, rather than writing a contradiction. Two more things that were quietly wrong: a run where the aesthetic or NSFW model could not be downloaded no longer wipes the scores your bank already had, and the images scored during such a window are picked up again once the missing piece is back, instead of staying half-scored forever. Finally, "Rescore all" is available for when you do want everything recomputed from scratch — it is the last line of ✨ Score's launch window, unticked, with its price next to it.

Drop every watermarked image at once — and get them back if you change your mind

Reviewing flagged images one by one is still the careful way through a watermark scan, but it is no longer the only way: a dataset's curation row now offers ✕ Reject all flagged next to 🔍 Review flagged. The number on the button is what it will really reject, not how many are flagged — small-image rescue pairs and failed rows are left out, because including even one rescue row made the server refuse the whole batch and reject nothing at all. It asks before it acts, tells you exactly what it is about to do, and then tells you the way back: rejected images stay on disk, and Show ▸ Rejected in the grid followed by ✓ Keep brings any of them home. The one thing it does destroy is said out loud beforehand — rejecting clears the watermark flags, so 🔍 Review flagged will be empty afterwards and only a new scan can flag them again.

Datasets and banks now find watermarks the same way — and you pick which way

The two screens had quietly drifted apart: a bank used the fast watermark detector whenever its extra was installed, while a dataset always asked the vision model, one slow question per image, with no way to change it. Both now read one setting — Settings ▸ Captioning & quality ▸ Watermark detection — with three values: Auto, Watermark detector, or Vision model. Auto is the default and behaves exactly as before, so an untouched install changes nothing. Pin the detector without its extra installed and the scan still runs on the vision model rather than failing, and the app says so, with the link to install the extra. Where the two engines genuinely differ, the screen says that too: only the detector can flag an image without knowing where the mark is, and those are now counted apart and left for 🔍 Review instead of being marked failed. A ⟲ Rescan incl. dismissed button appears when you have ruled some flags false positives — the only way to have them judged again by a different engine.

A watermark scan can be stopped, and keeps everything it already found

Starting 🧽 Find watermarks on a large dataset used to be a commitment: there was no Stop button anywhere, and the only way out was closing the tab and hoping. The progress banner now carries the same ⏹ Stop the captioning pass has. It stops after the image in flight — never mid-inference — and every watermark already found is kept, so running 🧽 Find watermarks again simply picks up where you left off. The live "Scanning… 12/340" counter now actually moves in the tab that started the scan, too; it only ever updated after a page reload before.

The 🔎 quality scan leaves the rest of the bank usable

On a big bank, pressing 🔎 Scan quality could take the whole app away for a couple of minutes: the bar reached 100 %, said “grouping duplicates”, and nothing else answered — sorting, renaming, promoting, all of it waited. Three things were behind that, and all three are fixed. The duplicate grouping no longer re-runs over your whole bank at the end of a scan that had nothing to scan — it runs when the hashes actually changed, and when you ask for it with “↻ Re-group duplicates”, which still applies a new duplicate distance without decoding a single file. When it does run it now shows its own progress and stops when you press Stop, instead of leaving you in front of a frozen-looking bar with no way out. And it no longer holds the database while it works, so the rest of the bank keeps answering: on a 50 000-image bank the grouping went from up to two minutes of dead application to a phase you can watch and interrupt, and the worst wait a click had to sit through fell from over five seconds to a few dozen milliseconds.

Full-model training now accepts Krea 2 Turbo — and a checkpoint of your own

Full-model training used to accept exactly one thing: the official Krea 2 Raw base. Choosing Turbo did not just refuse — it silently switched the dataset back to LoRA, which is why the option looked like it did not exist. Both refusals are gone, and the reason they were there has been fixed rather than waived: the configuration sent to the rented GPU used to name Raw no matter what you had selected, so simply lifting the ban would have produced a run labelled Turbo that trained Raw for hours. The base is now resolved from your selection everywhere it appears — the job sent to the pod, the run’s provenance, the model card, the recipe summary — and the full-model panel finally has its own Raw/Turbo switch and base picker. You can also fine-tune a Krea 2 checkpoint from your own disk: it rides to the GPU through a private repository on your Hugging Face account, exactly as it already did for LoRA runs, and the launch checks the pod’s credential can really read it before renting anything. What has not changed: a ComfyUI scaled-fp8 export still cannot be trained at all — the loader refuses it — and the app says so when you pick it, while a plain fp8 cast stays allowed. Turbo comes with a plain warning shown before you spend anything: Krea recommends training a LoRA on Raw and applying it to Turbo, nobody has measured full-model training on a distilled base, and it may cost the model its few-step behaviour. It is a warning, not a wall.

Re-caption no longer destroys the captions you wrote yourself

🔄 Re-caption used to warn you, honestly, that it could not tell a caption you had typed from one a model had written — one column, one string, no author. It can now: every caption records who wrote it, whether that is JoyCaption, Ollama, or you (typing one in a dataset, running a find/replace, or bringing .txt files back from another tool), and that record travels with the text through dataset → bank, bank → bank, promotion and backups. So a forced run keeps your own words and says so, the way the face pass already skips a person you declared yourself. The button now tells you three separate things instead of two: how many images it rewrites, how many captions it keeps because you wrote them, and how many it overwrites whose author was never recorded — captions written before this existed, which cannot be recovered and are re-captioned, so you learn that before the click and not after. If you actually want your own captions redone by a better model there is a tick box for it, offered only when you have something to lose.

🎨 Medium and 🔎 text search stop failing over a line of chatter

On some installs 🎨 Medium and 🔎 text search died with “the text encoder produced no result — check the ✨ Score interpreter”, on machines whose ✨ Score interpreter was fine — it had just produced the embeddings both features read. The cause was one line: the first thing the text encoder printed had to be its answer, so a first-run banner, a weights download or any greeting from the ML environment ahead of it was read as a failure. Both features now step over anything that is not an answer, the way the ✨ Score pass already did. And when the encoder really does fail, the message quotes what it actually printed — including the error output, which used to be discarded — with home-folder paths stripped, so it is safe to paste into a help thread and it no longer sends you to check a component it never looked at.

SeedVR2 settings tell you whether your target will be tiled

Tiling starts strictly above the crossover, and the crossover is 1.5× your tile size — so it lands exactly on the round numbers people type. Ask for 1536 px with the default 1024 px tile (or 768 px with a 512 px tile) and the upscale ran whole, with nothing anywhere saying why: no tiles, no warning, no line in the panel. The SeedVR2 card now names the lane your configured target will actually take, and when it sits on the crossover it says so and gives you the three ways to change it.

A busy fp8 conversion says so, instead of blaming your disk

Starting a second fp8 conversion while one was already running could answer “not enough disk space” — true, but not the reason it refused, and it sent you deleting gigabytes for a job the disk was never blocking. The refusal now names what is actually in the way: something is already being prepared, wait for it or stop it. It also answers instantly, because it no longer reads the Hugging Face repository just to say no.

Setup stops calling a damaged optional file a broken engine

A corrupted Klein consistency LoRA lit up Setup in red with “⚠ On disk, unreadable”, exactly like a dead model file — while generation carried on working perfectly, because that LoRA has never been required. The alarm said the engine was down when the engine was fine. It now says what is true: an unreadable file that nothing waits on is flagged in amber as optional, with a line telling you Klein still generates without it and a button to download it again; only a file the engine really needs keeps the red badge. The same file also stopped hiding on the other screen — the download buttons used to print “✓ Installed” over it, because it was not on the required list, so the one place offering to fix it was the one place claiming nothing was wrong.

See which engine actually wrote your captions

The captioning engine is set to Auto by default, and Auto is a chain, not a choice: JoyCaption writes what it can, the Ollama vision model writes the rest, and on a Concept dataset Ollama rewrites JoyCaption’s drafts. Those engines do not write alike — so captions could come back in two different voices, or in a different voice than last week, with nothing anywhere to explain it. Every pass now reports who wrote what, in the toast and on a line under the caption buttons: “Written by JoyCaption.”, “Drafted by JoyCaption, rewritten by the Ollama vision model.”, or “8 by JoyCaption · 4 by Ollama” when the batch was shared. If you want one consistent voice, ⚙️ Options lets you name a single engine — and now you can tell whether it obeyed.

Choose who writes your bank captions, and which pile gets them

The 🏷️ Caption pass used to take whatever engine and vision model your Settings held, and always ran over everything you had not rejected. It now has its own row: pick the engine and the Ollama vision model for THIS run without touching your Settings, and aim the pass at the kept images only, the undecided only, or both. Which model writes a caption is not a matter of taste — one that describes things in evasive terms produces captions that are about something slightly other than your images, and a LoRA trained on them learns to look away too. The button also stops saying “Caption all” and starts saying how many images it is really about to write, because already-captioned ones are skipped and the count you could see was never the count it acted on. Rejected images stay out of reach, whatever you pick.

Redo a bank’s captions with a better model — and see what that costs first

Once every image in a bank had a caption, 🏷️ Caption reached zero and went grey — taking the engine and model selects beside it out of reach, on exactly the bank whose captions you wanted to redo. A 🔄 Re-caption button now sits at the end of the Caption options row and runs the same pass over the pile you chose, with the engine, model, register and length you picked. It overwrites, so it tells you the numbers before you click: the button quotes how many images it will rewrite, an amber line quotes how many of those already carry a caption, and the confirmation repeats both. It also says the part we cannot fix: this app stores one caption per image and records nothing about who wrote it, so a caption you corrected by hand looks exactly like a generated one and is overwritten too — and no undo covers captions. With images selected the button stays inert and says why, because a selection can span pages that were never loaded and the count would be a guess.

Full changelog: v2026.08.04...v2026.08.07