ausboss-ComfyUI-AusBoss/CHANGELOG.md

Changelog

All notable changes to ComfyUI-AusBoss are documented here.

Unreleased

• After an update, old files in your browser can no longer break AusBoss. In Firefox, a page could keep running files from before an update even after Ctrl+Shift+R, so nodes lost their preview bar, card or editor until the browser's cache was cleared. Every update now gives the pack's files new addresses, so the browser always fetches the new ones. Restart ComfyUI and reload the page once; there is nothing to clear. (The server adds a version tag to the pack's own script imports as it serves them; the files on disk are unchanged.)

2.5.1 - 2026-10-01

• Manager search finds AusBoss by its repo name. Searching for ComfyUI-AusBoss or AusBoss nodes found nothing; only ausboss worked. The listing now uses those names too.
• After an update, AusBoss no longer half-loads. After updating the pack in Manager, some nodes could lose their card, editor or settings until the page was force-reloaded, because the browser kept using old copies of some files. The pack now asks the browser to check those files every time. A page that is already running old files shows "AusBoss was updated" with the one thing to do: press F5.

2.5.0 - 2026-09-30

• See-through pictures now outpaint properly. A product cutout with no background, a photo with see-through round corners, or a picture with a hole in it could come back on flat gray, or with a dark ring round the picture. Image Crop + Rotate + Pad now treats the see-through parts like the padding: they are filled, the mask marks them, the model paints them, and Stitch Inpaint never puts them back. A new last output, prompt_image, shows those parts as white. Wire it to whatever writes your prompt, so the prompt describes a real backdrop instead of the gray fill. The Krea 2 Outpaint and Krea 2 Rotate + Outpaint examples are wired this way. The original output also shows see-through parts as white instead of black, which fixes the Klein 9B Outpaint prompt as well. A part counts as see-through when it is less than 90% solid. Pictures without such parts come out exactly as before.
• Save Image, Select Frame, Mask Refine and LaMa Inpaint keep their size in Nodes 2.0. They grew to fit the picture they showed and covered the node below them, and after a run the picture showed twice. Now it fits inside the node, as in the classic view.
• Nodes 2.0: undo no longer leaves a node showing its old settings. After Ctrl+Z or Ctrl+Y, what you change on an AusBoss node reaches the run.
• Workflow Note grows to fit its text. A note saved a few pixels too short no longer hides its last row behind a scrollbar, and it can't be dragged shorter than its text. It never shrinks by itself.
• Stitch Inpaint: Tone match now works with Seam set to blend in. An outpaint can come back a little lighter, darker or warmer than your picture, and blend in used to keep that colour. With Tone match on, blend in now checks how the model repainted the strip of your picture next to the new area and takes that change back off the new area. It follows any edge, turned or straight, keeps blacks black, and leaves the colours alone when the change doesn't hold up along the edge. Blend in with Tone match at 0 stitches exactly as before, and the classic seam is unchanged.
• Stitch Inpaint says when the new area comes back unpainted. When an outpaint returns with plain bars or corners in the fill colour, a message and a console line say how much of the new area was left and what to try. It also catches a gray fill handed back a few shades darker, as Krea 2 can do. Painted pictures stay quiet, dark ones too.
• Krea 2 outpaints now paint dark and colourful pictures too. With the AnyPaint LoRA, a dark or very colourful picture could come back with its new area still gray: a gray frame around the picture, or a gray corner on a turned one. The model was copying the gray padding instead of painting it, whichever Krea 2 model file was loaded. Krea 2 Encode has a new mask input. Connect the pad node's mask to it, and the model sees the new area in your picture's own colour, the way AnyPaint was trained, and paints it. The Krea 2 Outpaint and Krea 2 Rotate + Outpaint examples are wired this way. They also feather the join 32 px instead of 12, which gives the model room to carry your picture over the edge: a thin strip of new area no longer comes back as a flat band with a hard line. Workflows without the link run exactly as before.
• Pressing Run before loading a picture now says so. When a picture or video loader has nothing picked (core Load Image and Load Video, Load Image + Pad, Image Crop + Rotate + Pad, Load Video and the Video Crop + Rotate + Pad nodes), the run stops before it starts, one message reads "Load a picture first:" with the node's name, and that node gets a red outline. Before, core Load Image started the run and failed with "[Errno 21] Is a directory". Loaders that are bypassed, muted, fed by a link or not needed by any output never stop a run, and runs sent straight to ComfyUI's API are not checked.
• One problem, one error. A missing picture on Image Crop + Rotate + Pad used to show "20 errors", the same message once per setting. Each AusBoss node now reports a problem once, on the setting it is about. The same fix lets ComfyUI check these nodes' values again: a number out of range or a choice that isn't offered is caught before the run instead of partway through it. A crop ratio from someone else's presets still runs, and uploads, subfolders and MaskEditor saves still load.
• "No mask painted" instead of a size error. Krea 2 Inpaint Masked, and any workflow that sends Load Image's mask to Crop For Inpaint, Mask Refine, Color Match or LaMa Inpaint, stopped with "Mask size (64, 64) does not match image size" when no mask was painted or a new picture was picked after painting one. It now says: "No mask painted: right-click your picture, choose Open in MaskEditor, paint the area, then click Save." LaMa used to hand the picture back unchanged in that case.
• Save Video: the NVIDIA formats explain themselves. Picking "mp4 h264 nvenc" or "mp4 h265 nvenc" without an NVIDIA GPU now says those formats need one and to pick "mp4 h264" instead, rather than a bare encoder error.
• Load Video and Video Crop + Rotate + Pad → Clip check the trim window when they run. A start after the end, or Fixed frames longer than the video, is reported by the node as it loads the video, where a trim bound fed by another node has its real value.
• Clip timeline: the bar always scrubs. With Length on, as the LTX 2.3 Video Outpaint example opens, dragging along the bar did nothing and IN would not move, with no word why. Now pressing or dragging anywhere on the bar moves the playhead, Length on or off. To move the whole clip, drag IN or the small grip in the middle of the kept part. When the Length already takes the whole clip, the Length row says "= the whole clip. Shorten it to move IN."
• Preview switch: on again brings the picture back. On Mask Refine, LaMa Inpaint, Select Frame and Save Image, turning the preview off and on again left an empty box and a shorter node until the next run. Now the node goes back to its height with the last result in it. If the node ran while the preview was off, the box says to run the workflow again instead of showing an older picture. Mask Refine's AUTO button now says what it does when you hover it.
• Crop + Rotate + Pad: the padlock you can see and trust. Reset and tapping the lit ratio turn it off. On the untouched picture it does nothing, so trimming there only trims. While it is on, the row says Held, new bands split evenly on both sides, and Reset crop also removes the bands it added. The lit ratio and the lock now look like every other "on" in the pack (solid teal), and hovering a button only outlines it.
• Crop + Rotate + Pad: nothing added when there is nothing to add. A picture or clip within about 1% of a ratio counts as that ratio, so a nearly 9:16 photo no longer gets a 1 px band. When the lit ratio is already the picture's shape, the size line under the picture says so ("already 9:16: pick another ratio or turn it"), instead of letting you pay for a render that paints nothing. A lit ratio saved in a workflow still applies to every new picture, and the size line names it where it acted ("pad to 16:9 1821×1024").
• Crop + Rotate + Pad: the resize warning in plain words, with a fix. It read "2.2% wider from steps". Now it says "2.2% wider: each side rounds to 32 px", its tooltip names a Step that avoids it, and with Fit on pad an Even out button adds a few pixels of padding so the picture is not stretched at all.
• Crop + Rotate + Pad: added space is easy to see. Padding and the corners a turn opens are drawn in the real fill colour with a faint hatch, so black bands show on the black stage, and only the picture the crop cuts away is darkened. The round rotate knob keeps clear of the padding diamonds on wide canvases.
• Crop + Rotate + Pad: controls that say what they need. Fit is dimmed and says "pick a ratio first" until a ratio is lit. Over the picture, the cursor and small arrows show only the ways it can move.
• Crop + Rotate + Pad: the full editor uses the node's controls. Its sidebar has the node's ratio row, padlock and Fit switch instead of a list, a checkbox and two buttons. The rotation is a number box you can drag (Shift for 0.1°) instead of a slider that turned 57° for a 9 px nudge. The size box stays clear of the handles, and the number boxes look like the node's.
• Crop + Rotate + Pad: big nodes stay tidy. On a wide node the controls stop at 460 px, centred, and only the picture grows. A wide picture picked into a node sized for a tall one shrinks the picture area to fit it; the node never grows for a new picture, so it never covers the node below.
• Clearer names on the Crop + Rotate + Pad nodes. Align is Round canvas to, Snap is Frames for with LTX (8n+1), Wan (4n+1) and any, and Local path is Server file. Resize is an off | on switch like the pack's other switches, and To start and Full clip look like the buttons they are. Saved workflows are unchanged: only the words moved.
• Compare: A and B follow the split line. The A | B badge stayed in the middle while the line moved. Now A sits just left of the line and B just right of it.
• Number boxes: Shift slows whole-number boxes. On boxes like Feather, whose smallest step is already 1, Shift changed nothing. It now slows the drag, so Shift always means fine.
• Workflow Note: models kept in a subfolder. The note said "installed" while a loader asked for the bare file name, which ComfyUI could not find and marked red. Such a row now says where the file is ("in LTXV 2.3/") and a Use it button points those loaders at it.
• A run from another workflow tab stays on its own tab. ComfyUI files each node's result under the node's number, so when a run from one tab finished while you had another tab open, its video or picture turned up on the open tab's node with the same number. On a Video Crop + Rotate + Pad node it showed as a second video under the buttons and covered the node below. Now the Crop + Rotate + Pad nodes never show a second preview, AusBoss nodes ignore results from other tabs (Show Text and Seed no longer save another workflow's text or seed), and when you go back to the tab that ran, its Save Video, Show Text, Compare and other AusBoss panels show that run. Core nodes behave as ComfyUI has them.
• Save Video and Compare keep their result when you switch tabs. Go to another workflow tab and back, and the player said "Run to preview the saved video" (Compare: "Run to load the A/B previews") over a result it had just made. They show it again.
• Nodes 2.0: Save Video shows the clip once. It also drew ComfyUI's own copy of the clip under its player.
• Crop + Rotate + Pad: the ratio buttons say what the canvas is. A lit ratio is the shape the canvas has now. Drag a handle to another shape and it goes dark, the row says Custom, and the size line gives the real ratio (1.49:1); before, 16:9 stayed lit on a 1889×1280 canvas. Tap a ratio to pad to it, tap the lit one to go back to the whole picture. Keeping the shape while you drag is its own padlock button at the end of the row, instead of a second tap on the ratio. The button at the start turns the shape between portrait and landscape (16:9 to 9:16) and never throws away your crop; with nothing picked it only turns the ratio labels. Crop or pad is now a small Fit switch, so it no longer looks like a ratio. Rotating keeps a lit ratio, a ratio you tap before loading a picture is used when you load one, and dragging the picture itself moves it inside its padding.
• Crop + Rotate + Pad: the editor uses the node's words. Its Multiple and Steps are Round canvas to and Step, as on the node.
• Crop + Rotate + Pad: the picture on the node stays big enough to see. You can no longer drag the node so short that the preview becomes a thumbnail: the picture keeps at least three fifths of its width in height (200 to 340 px), and a new node opens that way. The "Choose an uploaded video" line under the picker goes away once you pick one, so the picture gets that room too. A workflow saved with a shorter node opens with it a little taller, and the LTX 2.3 Video Outpaint example is laid out for the new size. In Nodes 2.0 the node also stops at the width its buttons need instead of squeezing them.
• Crop + Rotate + Pad: picking the first file in the list starts fresh. In a workflow saved with no file picked, like the examples, picking the first file in the list kept the rotation, crop and padding the workflow was saved with, which were made for a different picture. Any other file already started over. Now the first file does too, and it is padded to the ratio that is lit on the node. This goes for the image node and both video nodes.
• Mask Refine: Fill holes can leave the part you kept alone. Paint everything except a person or a pet, turn on Fill holes, and it filled them in too, so the part you meant to keep was painted over. The new Max hole size sets the biggest hole Fill holes fills, as a percent of the picture. At 2, the small gaps a quick brush leaves still fill, and a person, pet or object you left unpainted stays. It starts at 0, which fills every hole as before, so saved workflows run the same.
• The red "missing file" warning clears when you pick a file. Open a workflow that was saved with a picture you don't have and ComfyUI marks Load Image + Pad red and lists it under Setup required. Picking your own picture in the node used it, but the red box and the warning stayed on, even though the workflow ran. They now clear as soon as you pick a file. Every control on the node cards now tells ComfyUI when you change it, so its own checks stay in step with what you set.
• Crop For Inpaint: workflows saved before 2.4.0 now open with Stay in picture on. In 2.4.0 most of them opened with the switch off and kept the old crop, the one that runs past the picture's edges, while the same workflow sent through the API ran with it on. Both now get the new crop, as the 2.4.0 notes said they would. Turn the switch off on the node for the old crop. A switch you set yourself keeps its setting.
• Fewer example workflows. The examples are down to the seven that show what these nodes do that core ComfyUI doesn't: outpainting and inpainting that put your original picture back, lining an edit back up with its source, and removing a watermark from a clip. They are Krea 2 Outpaint, Krea 2 Inpaint Masked, Krea 2 Rotate + Outpaint, Klein 9B Outpaint, LTX 2.3 Video Outpaint, Qwen Image 2.1 Edit + Realign and Simple Video Watermark Remover. Krea 2 Inpaint Masked and Krea 2 Rotate + Outpaint are new.
• The example workflows open with their loaders empty. Every image and video loader in them is saved with no file picked, so you load your own picture or clip. The pier photo and the pier clip are still in example_workflows/inputs/ if you want to try one first.
• Krea 2 Outpaint example: the padding follows your picture. The example now pads with Image Crop + Rotate + Pad instead of Load Image + Pad. Pick a ratio or drag the orange handles, and the padding fits whatever image you load; before, it was saved in pixels for the sample picture. The prompt is written from the padded picture.
• Saved workflows keep only the nodes' own settings. Each node card and panel used to save an empty value of its own, and a setting added to a node in a later version could open holding it. That is how Stitch Inpaint's Seam and Crop For Inpaint's Stay in picture went wrong in older workflows. Workflows you save now leave the cards and panels out, and workflows saved before still open exactly as they did.
• Workflows from the 1.x versions open with the right switches. In LaMa Inpaint, Mask Refine and Select Frame the preview switch showed on, but no preview was made. It now opens on and works. LoRA Loader's settings showed Stop on missing LoRA off, even though the node still stopped the run. It now shows on, matching what the node does.
• The other examples got the same smaller layout. Klein 9B Outpaint, LTX 2.3 Video Outpaint, Qwen Image 2.1 Edit + Realign and Simple Video Watermark Remover now have named groups for your input, the models and the result, with the plumbing in two boxes. Node titles use the real node names, so you can find them in the node search.
• Simple Video Watermark Remover needs no other node pack. It finds the mark with ComfyUI's own SAM 3.1 (ComfyUI 0.20 or newer) instead of ComfyUI-RMBG, so there is nothing extra to install and no 3.45 GB download on the first run. On the sample it also stopped painting out a tram's number plate next to the mark.
• Klein 9B Outpaint example: no second face on tight portraits. Its automatic description used to describe the person, and Klein then painted them a second time under the join. It now describes only the setting. The padding follows your picture, as in Krea 2 Outpaint.
• Examples no longer carry the sample's turn and padding. Krea 2 Outpaint and Krea 2 Rotate + Outpaint open with an empty loader but still saved the sample's padding and turn, and the first file in your input folder could pick them up. They now save none. Krea 2 Outpaint opens with no ratio picked, so your picture isn't padded until you choose one.
• Krea 2 Inpaint Masked example: Stay in picture is on. The example had it saved off, so a big painted area was cropped past the picture's edge and painted smaller and softer.
• Krea 2 Inpaint Masked example: the person you keep stays. Paint everything around a person or a pet to give them a new background, and the example used to paint over them too. Its Mask Refine now has Max hole size at 2, so brush gaps still fill and what you left unpainted stays. The note also says to paint right up to what you keep, and to change one thing per run.
• Klein 9B Outpaint example: the seam blends in. With the classic seam, the new area's blacks could come back a hazy gray-green, and a turned picture could show a line along its edge. The example now uses blend in with Tone match. Its note says a big pad can grow a second hand or face, and that the result is about 1.6 MP.
• Qwen Image 2.1 Edit + Realign example: the edit changes only what you ask for. Its prompt writer could widen a request: "remove the photos from the wall" also took the fridge handle and the magnets. It now keeps everything the request doesn't name. A transparent PNG's see-through parts are filled with gray before the edit, instead of coming back black.

2.4.0 - 2026-09-29

• Load Video accepts a wired start or end time. Connecting another node (such as Math Expression) to Load Video's start or end time made the whole graph fail to queue with an error. It now queues and runs. A wired time that turns out to be bad still stops the run with a clear message.
• Stitch Inpaint: a Seam choice for outpaints. On a turned or padded outpaint, the join between your picture and the new area could show a lighter band and a smeared strip. Set the new Seam to blend in to fade the model's picture into yours instead: it hands over a little way inside your picture, where the two already line up, and leaves the new area exactly as the model painted it. On seven Krea 2 outpaints the colour step at the join shrank by 85 to 95%. It is the pick for turned pictures and for outpaints with Tone match off; with straight padding and Tone match on, classic usually looks as good. Seam sits in the gear menu in the card's corner, so the node keeps its size. With blend in on, Tone match and Fix edge halo dim, since blend in does not use them, and a blend in chip shows beside the gear. classic stays the default, so saved workflows render exactly as before, and Crop For Inpaint stitchers still paste the classic way.
• Krea 2 Outpaint no longer promises inpainting. Its Workflow Note and the Krea 2 Outpaint Model Patch help page said you could paint over part of the picture with the mask editor and pad at the same time. That does not work: Load Image + Pad does not take a painted mask. Both now say the workflow only extends the picture outward.
• Krea 2 Outpaint Model Patch help: Crop For Inpaint wants whole canvas. The page said a Crop For Inpaint stitcher spreads the reference over the full frame. On source rectangle the reference is pinned to where the crop sat in the full picture instead, so the page now says to pick whole canvas with one.
• Crop For Inpaint keeps its crop inside the picture. When the painted area was big, or the picture small, the crop could reach past the picture's edges, and the extra space was filled with stretched copies of the edge pixels. On a portrait with the coat painted over, only 42% of what the model saw was the picture, and the new jacket was drawn at about half size, then scaled back up. A new Stay in picture switch, on by default, stops the crop at the edges instead. In 8 test renders with wide masks the new area came out sharper every time (28-80% more fine detail), and a "tall window" request that the padded crop turned into more bookshelves came out as asked. Saved workflows whose crop ran past the picture now render a little differently; switch it off for the old crop.
• Resolution Master is now called Latent Size. That name already belongs to another node pack, so this node gives it back. Only the name changed: saved workflows and API graphs load and run as before, because the node's id (AUSBOSS_NODES_Resolution) is the same. A node you gave your own title keeps it; a title that starts with "Resolution Master 🆎" gets the new name in its place. The example is now Latent Size (AusBoss).json.
• Latent Size can be resized. Drag its corner to make it wider or taller, so it lines up with the nodes around it. The controls stretch with it, and the canvas preview takes any extra height. It still opens at its usual size, and it can be as narrow as 320 pixels.
• LoRA Loader: a choice for long names. Gear menu → Long names: end (the default, as before) cuts a name that does not fit at the end; middle cuts the middle instead, so the checkpoint number at the end (…_000004000, …_epoch_10) stays readable in a narrow node. The picker now opens wide enough for long names, and when one still does not fit it shortens the folder first and keeps the end of the file name.
• LoRA Loader goes narrower. Touching its corner used to make it jump to 424 pixels wide, and it could not be made narrower than that. It now goes down to 320 pixels (392 with separate model and CLIP strengths). The frontend adds 104 pixels to a panel's minimum width when it works out how narrow a node may go; these two nodes now take that back off.

2.3.0 - 2026-09-27

• Run Timer no longer slows the whole canvas (#77). Once the readout held a time, its glow ran a blur filter on the graph canvas three times a frame, and Chrome pays for each as a pass over the entire canvas: a test graph dropped from 60 fps to about 17 while panning. The glow is now painted into a small layer only when the digits change and copied every frame, so the readout looks the same and the canvas stays at 60 fps.
• LoRA Loader accepts a LoRA file that is a link. A LoRA kept on another drive and linked into models/loras was refused as "escapes the loras folders" and then reported missing, which stops the run in a workflow set to stop on a missing LoRA. The check now reads the path as written, so a linked file loads as it does in ComfyUI's own loaders, and a name that climbs out of the folder with .. is still refused.
• Qwen Image 2.1 Edit example: the consistency LoRA. Qwen Image 2.1 restyles often come back about 4-5% taller and shifted off the source's frame. The example now loads the Qwen Image 2.1 Consistency LoRA (ausboss/Qwen-Image-2.1-Consistency-LoRA on Hugging Face) in a LoRA Loader between the diffusion model and the reference cache: step 1500 is on, and the tighter step 2000 file is one switch away. Over 36 held-out restyles the worst corner moved 24.3 px (median) without it and 1.6 px with it. The Loader stops the run when the file is missing instead of quietly rendering plain Qwen 2.1; switch its row off to run without it. The Workflow Note lists both files, and the picture beside the example is a new render with the LoRA.
• New node (experimental): Realign to Source 🆎. Lines an edited picture back up with the original. Qwen Image 2.1 often draws edits, especially style changes, slightly zoomed in or shifted, by a different amount every seed. The node measures how far the edit moved and moves it back. For the best result, pad the picture with Load Image + Pad's new mirror fill before editing and plug its stitcher into the node: the edit gets room to move, and you get your picture back with real picture all the way to the edges. report says how far the edit had moved, and empty_mask marks any strip that slid off the edge. Tested on 166 Qwen 2.1 edits: style edits went from 35.5 px off at the worst corner to 4.6 px (median). With a 64 px mirror margin, 1 of 16 edits kept an empty strip, against 13 of 16 without one. New example: Qwen Image 2.1 Edit + Realign. Type a short edit like "make it a soft watercolor painting", and Qwen3-VL 8B, the text encoder the edit already loads, writes the full instruction before Qwen edits a padded copy and Realign lines it back up.
• Load Image + Pad: a mirror fill. Flips the picture outward at every edge, so the padding looks like more of the scene. It is the best margin before an edit you realign: Qwen Image 2.1 zoomed in to fill a mirror margin on 2 of 16 edits, against 7 with edge-pixel padding and 5 with flat gray. Saved workflows keep their fill.
• Settings menus no longer overlap in short windows. When a node's gear menu was taller than the window allowed, its rows were squeezed and the descriptions ran into each other. The menu now scrolls instead. This affects every node's gear menu.
• LTX 2.3 Video Outpaint example: LoRA Loader 🆎. The distilled LoRA and the outpaint IC-LoRA now load in one LoRA Loader 🆎 instead of two core loaders, with the same files and strengths.
• Load Image + Pad takes a wired image. A new optional source_image socket pads an image from another node instead of the chosen file, so a step before it (putting a transparent picture on white, say) no longer has to be saved and loaded back. The Source row dims while it is wired, and the on-node canvas shows the last image it padded. Saved workflows load unchanged: the socket is appended and has no widget.
• Editors can be closed without saving. The Crop + Rotate + Pad editor (image, video and clip) gains a Cancel button beside Save & close, and Escape now cancels instead of saving. When something changed, Cancel asks first, and Discard puts the node back exactly as it was when the editor opened. The Workflow Note editor asks the same before Cancel, Escape or a click outside throws edits away, and a JSON edit carried back to the form no longer survives a Cancel.
• Save Image: a preview switch, and plainer wording. The saved picture now sits in the pack's preview panel with the same small PREVIEW switch as Select Frame and Mask Refine; off, the picture goes away and the node is shorter (the file is saved and listed in the queue either way). The Workflow row now reads Embed workflow: no | yes.
• The Qwen3-VL caption steps keep Text Generate's thinking on. Krea 2's text encoder is Qwen3-VL-4B-Instruct, which has no thinking mode; with thinking off, ComfyUI appends an empty think block and the model often stops at once, so Krea 2 Prompt from Image could write nothing. Krea 2 Prompt from Image, Krea 2 Outpaint, Klein 9B Outpaint and both MiniMax H3 image workflows now keep it on, and the Krea notes say why.
• Seed: switching to Random rolls a new seed right away. The mode only rewrites the seed after a queue, so the first queue after switching from Fixed used to repeat the last run and ComfyUI returned the cached result. Switching from Fixed to Step takes the first step the same way; flipping Step's direction leaves the number alone.
• Source lists preview the file under the pointer. Load Image + Pad, Load Video and the crop / rotate / pad editors replace the browser's plain drop-down, where you picked by filename alone, with a list that shows the hovered picture or plays the hovered clip (muted, with its size and length) and filters as you type. Arrow keys, Enter and Escape work too. The saved value is still the node's own image / video widget.
• Crop + Rotate + Pad: feather softens the mask only. Image Crop + Rotate + Pad and both Video Crop nodes also faded the picture into the fill colour, unlike Load Image + Pad. Outpaint models read that ramp as content: the Qwen 2.1 outpaint LoRA painted a darker band along a tilted edge, and LTX's IC-LoRA left feathered black bars unpainted. The stitcher kept the faded pixels too, so Stitch's colour match measured its drift against them and pulled the fill toward grey (colour error 3 → 18 on a straight pad at feather 24). The picture now keeps a hard edge against the fill at any feather; the mask, and the stitch blend built from it, are unchanged. Saved workflows with feather above 0 get a hard-edged image.
• Stitch Inpaint's tone match no longer pulls a rotated outpaint toward the fill colour. On a Crop + Rotate + Pad canvas that was rotated and padded (a canvas multiple's few extra pixels count), color_match read the rotation's empty corners as picture and shifted the whole fill, by up to 33 dE on a flat test. It now reads only picture pixels at each seam. Load Image + Pad and Crop For Inpaint stitch exactly as before.
• Load Image + Pad no longer adds a strip along an edge you did not pad. Rounding the canvas to Multiple put the leftover on the right and bottom whatever you padded, so the published Krea 2 Outpaint workflow (top and bottom padded, Multiple 16) grew a 6 px grey strip down the photo's right edge. Krea painted it flat and lighter, the feather opened the photo beside it to the sampler, and Stitch Inpaint's color match read the strip as a seam and darkened that band. The leftover now joins a side you padded (the right or bottom one when both are). Where you padded neither left nor right, or neither top nor bottom, a Budget scales the photo onto the multiple instead, keeping its shape, and without a Budget the photo loses those few pixels evenly from both edges; the node's canvas shows them as −N px. Saved workflows can come out slightly different: with only the left or top padded, the photo sits up to one multiple minus a pixel further right or down (15 px at 16, 63 px at 64); with a pair of edges unpadded, that side can be one multiple smaller without a Budget, and with one the photo's scale shifts a little, which can move either side by a multiple or so. The shipped examples pad every side and are unchanged. Crop + Rotate + Pad keeps the old rule for now.
• Image Crop + Rotate + Pad: stitch_blend and stitch_grow. The stitcher's blend was fixed at 32 px; it now has the clip node's two optional inputs (defaults 32 and 0), and the editor shows the same Inpaint & Stitch section with Show blend.
• Crop + Rotate + Pad: rotating keeps the crop. The crop keeps its size and stays over the same part of the picture, from the knob, the editor's slider, its number box and Reset rotation alike. The slider used to throw the crop away (a centred 1:1 crop jumped to the top), and the knob slid it off the picture as the canvas grew.
• Crop + Rotate + Pad: the size explains itself. A line under the picture names every step that sets the output size (crop → pad → align → resize), warns when the resize Step stretches the picture by more than 1% and when Align pads a strip of fill ahead of a resize, and the editor's panel lists the same steps. Ticking Resize on the node now shows the resize Step next to the megapixel budget.
• Crop + Rotate + Pad: exact aspect sizes. A crop at a locked ratio no longer loses a pixel to rounding (21:9, 9:21, source), so the size shown on the node always equals the run. A few saved crops come out 1 px larger.
• Image Crop + Rotate + Pad: API prompts saved before the resize inputs existed validate again; the four resize inputs are optional.
• Video Crop + Rotate + Pad → Clip: one way to set the length. The Fixed frames and Limit boxes are gone from the node face. OUT ends the clip, and a Length switch (off by default) sets it by a frame count instead: OUT then follows IN at that distance and the count survives a video swap. OUT always sits on the last frame the run outputs (Every nth and Snap included), where Limit used to end the clip short of it. Connecting fixed_frames or frame_load_cap hands the length to that input: the OUT handle goes away and only IN is left to set. No inputs changed; saved Limit values read as a Length, and an old Fixed frames value becomes one on the first edit.

2.2.0 - 2026-09-24

• Save Image and Save Video save to the output folder itself by default (image_..., video_...) instead of an AusBoss subfolder, and the example workflows save without a folder too. Saved workflows keep whatever prefix they already have.
• The pack makes no network requests. LoRA Loader's Civitai lookup is removed: the info-card button, its gear-menu switch and the server route. A .civitai.info sidecar already beside a LoRA, from an earlier lookup or another tool, still fills in the title, base model, Civitai trigger words and model link. The lookup's hash cache, ComfyUI/user/ausboss/lora_hashes.json, is no longer used and can be deleted.
• Recreate node and Replace with AusBoss nodes create links through one shared helper that calls LiteGraph's connectSlots directly: the same checks and callbacks as before, without the index-based connect call that the Registry scan reads as a network socket. Release preflight fails when shipped code contains an HTTP or socket client, or that connect call.

Video

• Load Video, both video transform nodes and their previews count time from the video stream's first frame. Transport streams (.mts, .m2ts, .mpg) and MP4s whose video starts late loaded the wrong window or no frames at all; seeks now start from a keyframe, and audio stays in step with the picture.
• Video Crop + Rotate + Pad → Frame gains the megapixel resize the image and clip nodes have, as optional inputs so older API prompts still run, and both it and Image Crop + Rotate + Pad append width and height outputs.
• The three Crop + Rotate + Pad faces share one set of controls: a canvas row with fill colour, feather in pixels and the resize budget where the node has one, and Open editor | Reset crop | Reset. The image node's feather was an on/off box and had no fill colour; only it had a full Reset. Reset clears rotation, crop and padding and keeps fill, feather and Align.
• The video transform panels no longer vanish when the graph is zoomed out.
• Dropping an AusBoss-saved video on Load Video restores its trim and sampling values again; current frontends' own upload handler had taken the drop.
• Select Every Nth takes an optional fps and appends an fps output divided by nth, so a thinned clip keeps its duration.
• Stitch Inpaint no longer fails when a video model returns fewer frames than the stitcher holds (LTX keeps 8n+1).
• Save Video honours ComfyUI's --disable-metadata, and shows its CRF and Metadata rows only for the formats that read them.

Image, mask and LoRA

• Load Image + Pad has a Reset padding button on its stage.
• Save Image: with on_existing set to error, the whole batch is checked before any file is written. Editing a legacy exact name no longer overwrites the filename prefix, the Browse popup no longer closes itself and closes on Esc, and Embed workflow is an off | embed pill like every other boolean on a card.
• LoRA Loader: a row skipped because its file is missing adds no trigger words (rows parked at strength 0 still do). Every number is a scrub control, including the gear menu's default strength and step and the info card's suggested range, which can be left at "any". The gear menu's Reset no longer folds separate CLIP strengths into the model strengths, and row edits enter undo history.
• Krea 2 Encode's tooltips and help describe both outpaint setups: VLM reference on for AnyPaint, off for Registered Outpaint.

Pack-wide

• Replace with AusBoss nodes keeps the saved settings of missing nodes; current frontends give their placeholders stand-in widgets that it read instead, so every replacement started from defaults.
• Clicking a node's ? badge no longer starts a drag; Run Timer, which has no title bar, has About this node in its menu.
• Copy buttons report a blocked clipboard instead of a false success (plain http on a LAN address has no clipboard API), and Show Text gets one.
• Messages that used a blocking alert() or only the console are toasts.
• Errors name the node you are using: Save Video, the Clip node, Stitch Inpaint and the padding nodes no longer borrow another node's name.
• Gear menu resets keep the settings that belong to the node itself. Compare's mode and Mask Refine's AUTO enter undo history.
• The optional live status and runtime badges sit above the left end of the title bar, clear of the frontend's own node-source badge.

Examples

• Three new examples. Video Reframe and Slow Motion and Dataset Frames from a Clip need no models; LaMa Object Removal needs only big-lama.pt.
• The node tour and the watermark remover take the other examples' naming: Video Node Tour (AusBoss) and Simple Video Watermark Remover (AusBoss).
• Loaders select the official file names from their download links and carry matching download info, so a fresh install finds every model the Workflow Note lists. The caption encoders link the Qwen3-VL 8B file they load, SAM3 Segment names ComfyUI-RMBG, and six LoRA Loaders save their missing-file rule.
• Krea 2 Outpaint turns VLM reference back on, the AnyPaint recipe its note describes, and works on a 1.6 MP canvas budget like Klein 9B Outpaint. At full size a big photo's canvas (4 MP) painted its new sky darker, creased at the corners and cut invented limbs at the old edge. The node tour thins with Select Every Nth and carries the rate through it.
• Every example has a valid id, opens at a zoom that draws widget text, drops other tools' leftover metadata, stores its named widget values in agreement with the loaded ones, and stops promising audio from the silent sample clip.

Repository

• README: install through ComfyUI-Manager or comfy-cli, a corrected upgrade note, working commands for the optional extras, the pack-wide tools, and links to the changelog, contributing guide and security policy. Help pages that had drifted from their nodes are corrected, and the 1.2.0 changelog heading is back.
• The Registry archive leaves out the README media (14 MB to 7 MB), and the preflight lists the real archive and holds it to a size budget.
• Publishing happens only when main's version changes, after the offline checks pass. Pull requests also run the standard-library Python tests, a JavaScript syntax check and ruff.
• A node API snapshot test fails when a released input or output is renamed, reordered, removed or newly required.
• Add SECURITY.md, CONTRIBUTING.md, issue and pull request templates, .editorconfig and .gitattributes; remove a finished checklist that was committed with 2.1.0 and two dated review reports.

2.1.0 - 2026-09-22

• LoRA absorption preserves repeated applications and upstream → stack → downstream order. Shared MODEL/CLIP paths, used auxiliary outputs, a bypassed destination or linked stack values leave the original loaders untouched.

• Rename the editor’s Reset all action to Reset transform and explain which settings it keeps. Keep the selected video frame during a transform reset so the picture and timeline stay in agreement. Document the Clip node’s original-frame output in README.

• Video Clip: add Fixed frames for an exact-length selection. Drag either handle or the highlighted band to move both ends together. Duration follows output fps; frame/audio counts stay aligned. Existing frame caps retain their maximum-limit behavior. Too-short sources report the required duration.

• Refresh related examples: Krea/Klein Outpaint and both image-guided H3 graphs describe their sources with tested Qwen3-VL caption paths. Krea reuses its own text encoder; Klein and H3 list the extra caption model download. Setup cards explain manual prompts and Krea's caption sampling settings.

• LTX Outpaint captions the unpadded original frame. H3 First + Last Frame no longer hardcodes the ending timestamp. Krea Studio includes LoRA trigger words, the LoRA Stack example saves a prompt sidecar, and the transform example compares results against the new original outputs.

• LoRA Loader: refresh restored rows when widget values arrive after node configuration, including undo, without switching workflow tabs.

• Video transform uploads stream into the input folder instead of buffering through the image-upload endpoint, allowing files over 100 MB.

• All Crop + Rotate + Pad nodes: add Crop / Pad mode, crop-only ratio locking, Reset crop, and pixel alignment on the node face. Image and frame nodes append stitcher and original outputs; the clip node appends original frames.

• Image Resize: new nodes start in megapixels mode; saved modes are retained.

• Add Qwen Image 2.1 Text to Image and Edit examples with numbered stage groups, model download cards, AusBoss size and seed controls, and PNG saving. Text to Image includes an empty optional LoRA stack; Edit uses source-matched Qwen conditioning and a before/after comparison.

• Video Crop + Rotate + Pad → Clip: distinguish source and output fps in the trim footer, including a connected numeric rate and Every nth.

• LoRA Loader: separate single-row removal from a red, trash-marked Delete all LoRAs action with confirmation. LoRA files are never deleted.

• LoRA Loader: absorb rgthree Lora Loader Stack and JPS Lora Loader nodes on either side of the connected model chain. Preserve disabled JPS rows and rgthree's linked strengths; refuse linked settings or an overflowing stack before bypassing the source loaders.

• Video Crop + Rotate + Pad → Clip: add optional force_rate, start_frame, end_frame, and frame_load_cap sockets. Connected frame bounds lock the matching timeline handle; a connected cap locks Limit. Rate conversion drops or repeats frames while preserving playback timing.

• Image Resize: add compact W/H labels beside the paired size fields.

• LTX 2.3 Video Outpaint example: disable Stitch Inpaint color matching, which can spread moving seam content into flickering dark or light bands across the generated area. Source restoration and blending remain on.

• LoRA Loader: right-click a row to insert an empty LoRA slot above or below it, preserving the order and settings of the existing stack.

2.0.2 - 2026-09-10

• The LoRA loader validates the SHA-256 used in its fixed Civitai URL and refuses redirects, so the lookup cannot follow another host.
• The completion chime is a small WAV generated in memory and played through an audio element, preserving the sound without WebAudio graph connections or a bundled audio asset.
• Save Image writes only inside ComfyUI's output folder. output_dir used to take any absolute path, so any client that can reach the unauthenticated /prompt route could make the server write an image and a caption .txt of its choosing into any folder ComfyUI can write, overwriting by default - one step from code execution through a custom node's requirements.txt. output_dir now takes a subfolder name inside the output folder, like ComfyUI's own Save Image; absolute paths, drive letters, ~ and .. are refused before the run, and the card's path preview says so as you type. Workflows that saved into an absolute folder need a subfolder name instead.
• Local path mode reads only ComfyUI's input, output and temp folders. The 2.0.1 AUSBOSS_TRANSFORM_LOCAL_PREVIEW switch that opened the rest of the disk is gone; a refused path says where the video has to be.
• Save Image checks the complete image and caption destinations, and Save Video checks both the prefix and final file, rejecting symlinks that lead outside the output folder. Windows drive and alternate-stream syntax is rejected before filesystem access.

2.0.1 - 2026-09-09

• Local path mode is opt-in beyond ComfyUI's folders. The video transform nodes' local_path used to be read by every queued run, which means any client that can reach the unauthenticated /prompt route could make the server open any video on the disk. By default the mode now reaches only ComfyUI's input, output and temp folders, for queued runs and editor previews alike; starting ComfyUI with AUSBOSS_TRANSFORM_LOCAL_PREVIEW=1 restores full-disk access on that server. Nothing else in the pack reads a path or contacts a host taken from a widget.
• Video timeline overhaul (Video Crop + Rotate + Pad → Clip / → Frame). The rail under the node's preview is now a real timeline: press or drag it to scrub a playhead on the node face, no editor needed, and the frame picker's face got the same rail for its output frame. IN and OUT are frames on the source's grid (the boxes take frame numbers; the stored seconds are derived and resolve back to exactly those frames), the playhead rides on a dragged handle so the stage shows the first or last frame the run keeps and stays there on release, and the bright part of the selection is what reaches the output after the frame limit and Snap. The editor's slider is replaced by the same rail, plus Set IN / Set OUT buttons and I / O keys. The Snap select no longer clips at the node's minimum width.
• Scrubbing no longer flickers. The storyboard tile only stands in while it is nearer the target than the frame already on the stage (with one keyframe per clip it used to be frame 0 on every pointer move), and storyboard tiles are now decoded at their target times instead of the keyframe before them. Two backend fixes: asking the scrub session for the frame it already decoded returned the frame after it, so every drag release landed one frame late; and served preview frames are kept in a small cache so revisiting a region costs no decode.
• Swapping the clip no longer breaks the canvas. Choosing another video used to reset the fill to grey, the feather to 24 and clear the padding - silently, since none of those showed on the node - which is precisely what the LTX outpaint LoRA cannot work with. A new source now keeps fill, feather, the resize budget, Snap and Limit, and is padded to the lit format chip; only rotation, crop and the trim window start over. Both video nodes also gained a canvas row under the format chips (fill swatch, feather, resize budget), so those values are visible and editable on the node face. A fresh clip node now starts outpaint-ready - black fill, feather 0 - and the editor's Reset all returns to the node's own defaults instead of the grey, feathered canvas the image nodes keep.
• Aspect lock. Tapping a lit format chip locks the format (padlock on the chip); crop and padding drags then keep the canvas at that aspect with the other axis's padding following, and a padding handle stops where the other axis would have to go negative. A third tap clears the bands and the lock. Lock aspect in the editor's Crop section is the same switch.

2.0.0 - 2026-09-07

• Breaking: LM Studio Chat is removed. The Registry identified an unrestricted server-side request through its workflow-controlled endpoint. The node, HTTP helper, frontend, help page, and tests are removed. Existing workflows using AUSBOSS_NODES_LmStudioChat must replace that node; a Text node can supply a fixed prompt where live chat is unnecessary. Other public mapping keys remain unchanged.

• Added Resolution Master, Seed, Workflow Note, Run Timer, and Video Crop + Rotate + Pad → Clip, alongside 15 organized example workflows.

• The README now includes eight visuals using actual editor screenshots and tested renders, including an animated video outpaint comparison.

• Release PRs now run offline validation and frontend tests in Actions. A separate Registry approval check reports the exact version's status and review reason after publishing, with manual rechecks that do not republish.

• Klein 9B Outpaint now samples the encoded padded image through a noise mask and uses a 64 px feather. Controlled comparisons showed a cleaner seam than the previous empty latent and 24 px feather; Stitch Inpaint still preserves the protected source exactly.

• Completed the public-pack readiness review: grouped README sections cover every public node; all example graphs have setup notes, stage groups, and matching thumbnails. Added the model-free Image and Video Transform example for both editors and Align Image. Corrected installed model selections, literal line breaks in the LoRA example note, card-driven layout overlaps, and the tour's sample and timing instructions.

• Load Image + Pad now uses a widget card for source/upload, fill, feather, alignment, and budget, with individually linkable exact padding controls. Text uses a growing multiline card, and Krea 2 Outpaint Model Patch exposes both reference placement and KV caching in its card.

• The watermark example keeps SAM3 detection while removing rgthree presentation helpers and the Easy Use frame-count display. Its setup card names ComfyUI-RMBG, and SAM3 unloads between branches to leave GPU memory for LaMa.

• Release preflight now checks saved links in both directions, stage containment, title-bar overlaps, and setup-note formatting. Backend tests have a process-isolated runner so ComfyUI stubs cannot leak between test files.

• LoRA Loader 🆎 now stops on a missing LoRA by default: an enabled row whose file cannot be found fails validation before sampling with an error that names the file and says where the switch is, instead of warning and rendering without it — what scripted and agent-driven runs need. Gear menu → Stop on missing LoRA (API input on_missing, "error" by default or "skip") restores the old warn-and-skip behaviour per node; graphs saved before the input existed get the default.

• Added rendered MiniMax H3 Text to Video and First + Last Frame examples, plus a model-free Resolution Master example. Corrected H3 duration rounding to never undershoot the requested seconds. Tested and corrected the Klein PixaOutpaint prompt and conditioning; the protected source interior remains pixel-exact.

• Resolution Master 🆎 joins the public pack: an explicit orientation toggle, ratio chips, canvas handles, sub-1 MP budgets and an empty latent output. Corrected zoomed drags, bounded extreme ratios, linked-axis controls and restore behavior; removed the external font request.

• Five new showcase workflows. Klein 9B Inpaint (a painted mask through Mask Refine 🆎 and Crop For Inpaint 🆎, Klein 9B edits only that crop, Stitch Inpaint 🆎 puts it back), Klein 9B Edit (Image Resize 🆎 sets the 1 MP working size and Image Size 🆎 keeps the output the source's shape), Klein 9B Outpaint (Load Image + Pad 🆎 on any side at once, PixaOutpaint in the LoRA Loader 🆎, a padded reference and masked Flux 2 starting latent), MiniMax H3 Image to Video (Image Resize 🆎 to H3's 768 px canvas, Float 🆎 + Math Expression 🆎 for the model's 17k+5 frame grid, the turbo LoRA model-only, Save Video 🆎 with the model's own audio) and Krea 2 Prompt from Image (core Text Generate on the Krea 2 text encoder, Show Text 🆎, the prompt wired into Krea 2 Encode 🆎). Each opens with a Workflow Note 🆎 and ships a thumbnail; a masked copy of the sample picture joins example_workflows/inputs/. Every node in the pack now appears in at least one example, including the node tour.

• Krea 2 outpaints on every side at once. The Krea 2 Outpaint example is rebuilt around yijunwang2's AnyPaint LoRA (krea2_anypaint_rank32). The old one-axis rule was never a Krea 2 limit: the patch's registered placement is the contract of the same author's Registered Outpaint LoRA, which was trained on a source spanning one whole canvas axis - and the shipped example did not load that LoRA at all. AnyPaint is trained the other way round: the grey-padded canvas is the reference, spread over the whole frame, and the sampler's noise mask keeps the known pixels, so left, right, top and bottom can grow in one pass. Krea 2 Outpaint Model Patch 🆎 gained an optional placement input for it - source rectangle (the default, the old behaviour) or whole canvas - and its help page explains which LoRA wants which. Saved graphs load unchanged.

• Outpaint seams: no more dark band, no more corner diagonals. A four-side outpaint used to show a faded band along the source's edges and 45° lines running out of its corners. Measured on the pier sample: the model paints the new area a few percent off the source's tone on every side (fill colour makes no difference), the feathered sampler mask smears that step into the band inside the source, and the pad feather merged its per-side ramps with a maximum, whose iso-lines crease at the corners and print as diagonals. Stitch Inpaint 🆎's color_match now reads the drift ACROSS each seam - the generated pixels just outside against the original pixels just inside, line by line - instead of inside the mixed band, and blends the sides by distance instead of picking the nearest, so the correction turns the corner smoothly and the band gets exactly the share the sampler gave it. Load Image + Pad 🆎's feather (and every stitcher built from it) now unites its ramps instead of taking their maximum: identical along a side, bilinear in the corner. Both outpaint examples run with color_match 1 again; their seam steps drop from 2-6 % to under 1 % of luminance.

• Video Crop + Rotate + Pad → Clip 🆎. The transform nodes' third member: one rotate, crop and pad applied to every frame of a trimmed video, with Load Video's start_seconds / end_seconds, every_nth and max_frames, and the same outputs (frames, the window's audio, frame count, fps, size, duration) plus the generated-area mask — so a clip can lose a border or a tilt, or grow the black bands a video outpaint paints into. It shares the frame picker's editor (the timeline picks the preview frame; scrubbing never re-runs the clip) and gains the image node's megapixel resize, which the editor now shows for any node that carries those widgets. Frames go through the transform in chunks with the queue's progress bar and cancel serviced between them. It also emits a stitcher (a ninth output, after the existing eight) built from the final resized frames, so Stitch Inpaint pastes the source back over a generated clip without a Crop For Inpaint node; the editor's right sidebar gained an Inpaint & Stitch section - blend ramp, a stage overlay of where the paste lands, and grow behind a disclosure - kept separate from the padding feather. The feathered paste mask is now one shared helper for every stitcher producer. A Snap select on the trim strip (frame_snap: free / 8n+1 / 4n+1) drops trailing frames to a count the video model keeps, so frame_count, the audio window and the stitcher match the clip that comes back; the showcase workflow wires the clip's size and count into the empty latent and snaps at 8n+1.

• Transform nodes: drop a video, pad to a format in one tap, a sharp preview at any zoom, trim without losing the playhead. A video file dropped onto Video Crop + Rotate + Pad (frame or clip) uploads and becomes the source, as core's upload widgets do for their nodes. A row of format chips right under every transform preview (16:9, 9:16, 1:1, 4:3 ...) pads the whole source to that aspect with centered fill bands - the editor's Pad to aspect without opening it; the lit chip clears it again. The node preview's backing store now follows the graph zoom (up to 4x), so zooming in on the node no longer shows a blurry, pixelated stage. Dragging an IN/OUT handle now peeks at the frame under the handle and returns to the playhead on release, instead of moving the scrub bar and the saved preview position along with the trim.

• Stitch Inpaint takes a shorter generated batch. A stitcher built from more frames than came back is trimmed to the leading ones with a console note instead of failing after the run - video models keep 8n+1 or 4n+1 frames and drop the tail. More inpainted frames than source frames is still an error.

• Save Image redesigned as one card. Folder with a Browse into the output folder, Filename, a live path preview that turns into the first file written after a run, the name tags as chips - Counter (on, the next free number for that stem, never a collision), Date, Time, Size and Batch # - PNG / lossless WebP (new) / lossless JXL pills, and an Embed workflow switch. Two link-only inputs: filename takes the exact name from upstream (its image extension swapped for the chosen format's, a blank value stops the run rather than inventing a name) and caption_text writes the paired .txt. The old widgets stay underneath in their saved order, so existing workflows load as they were and a legacy exact_name shows in the field tagged exact. Local names lose core's trailing underscore (shot_00001.png, not shot_00001_.png); numbering restarts at 00001 beside the old files without touching them.

• Align Image drops its offset outputs. offset_x / offset_y were never wired in practice; the node stops at image, width, height. A saved workflow that linked them loses those two links.

• Select Frame, Mask Refine and LaMa Inpaint previews can be switched off. A thin bar sits between the settings and the picture - the node's tools (Mask Refine's AUTO) on the left, a small preview switch on the right. Off, the picture's box is gone and the node is that much shorter; the switch stays where it was, and the node writes no file to the temp folder (a new optional preview input carries the choice).

• Units live inside the scrub box (px, MP, ×), ahead of the chevrons, and every single-field row in a card reserves the same slot for one whether it has a unit or not, so the numbers down a card share one centre line - a px no longer nudges its number left of the row above.

• Widget cards: the classic canvas widgets are gone from every node face. Image Resize, Color Match, Stitch Inpaint, Crop For Inpaint, Mask Refine, Save Image, Save Video, Load Video, Align Image, Frame Interpolate, Math Expression, Select Every Nth, Split Batch, Select Frame, Merge Batches, Integer, Float, LaMa Inpaint, Krea 2 Encode and the Krea 2 Outpaint Model Patch now carry one compact card in place of the full-width rows with an arrow at each end: numbers are the pack's scrub control, short choices a segmented pill, long ones a select, booleans an off | on pill as wide as the other controls, hex colors a swatch. Rows that only mean something in one mode (a resize target's size, a pad color) show in that mode; the rarely-touched settings fold behind one disclosure (Crop For Inpaint's targets and extend, Save Image's exact-name and folder, Mask Refine's More). The widgets underneath are untouched - saved workflows, the API format and widget-to-input links work as before, and a row driven by a link dims out. Shared in js/shared/widget_card.mjs.

• Links still land on card rows. The frontend only draws a widget's input socket while a link is being dragged or once it is connected, at the widget's own place - which a hidden widget no longer had. Each row now lends its position to the widget it stands for: drag a link over a card and a socket appears beside every row that can take it, drop it on the row (the socket, the label or the field itself) and that widget is driven by the link, exactly as with the classic widget. One socket per row, so no row holds two linkable fields: Image Resize's width and height, Frame Interpolate's two rates and Save Video's fps and crf keep their side-by-side layout but their sockets sit with the node's inputs (under image / mask), where a link greys the field out like the old converted input; Crop For Inpaint's target and extend values, Mask Refine's levels and Load Video's size and every-nth / limit are single rows now, one socket each.

• Math Expression: a, b, c are sockets. The three values arrive on the node's left edge and take a FLOAT or an INT (an image width wires straight in); the card is the expression alone. An unwired value reads as 0. A workflow that had typed a constant into a, b or c should carry it in the expression instead - those boxes are gone.

• Krea 2 Encode is one card. The prompt, the negative prompt and the VLM reference switch sit in a single panel; the prompt box takes the node's spare height when it is dragged taller. The node is unchanged underneath: a plain two-prompt encoder without a VAE and reference, and an edit/outpaint encoder with them (checked both ways on a Krea 2 turbo render).

• A node that loaded shorter than its new card grows on load. A workflow saved before a card or the preview bar existed kept the old node height, which squeezed the panel under its floor and clipped it flat (LaMa Inpaint's preview box); the node now takes at least the height its widgets ask for.

• LoRA Loader: the bar scale follows the enabled rows. Switching a row off hands the shared strength scale to the strongest row still in play; the dimmed row's own bar clamps at the edge meanwhile.

• Workflow Note 🆎: the card for sharing a workflow. A big title, a Markdown how-to, the models the workflow needs with Download buttons grouped under the 📂 ComfyUI/models/<folder> they belong in, the node packs it depends on, and the author's links — the note every shared workflow carries, without a second pack. Every model file and node pack on the card is checked against the install that opened it (teal dot and an installed pill when found, subfolders included; red dot and the Download button when not), so the person who downloaded the workflow sees what is left to do before the first run. A pencil opens the editor (form or raw JSON; Detect from this workflow fills the pack list from the open graph); the Banner layout turns the same node into a title label. Stored as JSON in one STRING widget; rendered as text, never HTML; only http(s) URLs become links.

• Seed 🆎: the seed that remembers what ran. One INT wire, one card: the number (click to type, ⧉ copies), a Random / Fixed / Step segment that drives the sampler's own after-generate control, and the buttons a run needs — New seed rolls and pins one, Use last run puts back the seed the last run actually used (reported by the backend, so it is right even after Random has already advanced the box) and pins it, ▾ lists the last eight. The history saves with the workflow.

• Run Timer 🆎: a stopwatch for the whole queue. One black readout with no title bar, no pack badge, no wires and no padding — the node is the display, painted by the node itself on the classic canvas, so it drags from anywhere and resizes by the corner with the digits scaling along (a DOM panel stands in under the Nodes 2.0 renderer). Starts on execution_start, ticks, holds the total at the end (amber while running, teal done, red failed); the last time saves with the workflow and the right-click menu lists the few before it.

• Stitch Inpaint: color_match. An outpaint often comes back a step lighter or warmer than the picture it extends, and the seam shows even when the content continues perfectly. The new 0..1 control measures that drift in the feathered band — where the sampler's pixels and the true pixels overlap, so it reads the model's own shift rather than comparing unrelated content — per padded side and per line along the seam (sky and water drift differently), smoothed — and pulls the pasted region back onto the original's tone before the blend. Source pixels stay bit-identical. Off by default; the outpaint showcase ships with it at 1.

• Panels hide their storage widgets under Nodes 2.0 too. The Vue renderer ignores the classic canvas collapse and filters widgets on its own options.hidden flag, so the Seed card, the Workflow Note, the LoRA loader and Load Image + Pad all showed their raw storage widget above the panel there. The shared hide helper now throws both switches. Run Timer's Nodes 2.0 fallback also drops the renderer's 225px minimum width and its footer badge, so the readout stays the whole node.

• The Registry archive stops shipping the workshop. A documented .comfyignore keeps tests, offline scripts, CI and the developer docs out of the published zip (comfy-cli's packer honours it), which also removes the test fixtures the Registry scanner kept flagging. release_preflight.py now checks the file covers those paths and never swallows a runtime one, parses every example workflow, and pairs each thumbnail with its workflow; the stale thumbnail generator is gone.

• Showcase workflows in example_workflows/, four graphs on ComfyUI core plus this pack and nothing else. Krea 2 Studio: draft, then refine - the plain Krea 2 Turbo draft goes through a learned 4x upscaler, Image Resize lands it on a 2.3 MP budget, and a second KSampler repaints it at denoise 0.15 with the same prompt and seed (measured against plain resize, learned-only, pixel-pass and latent-upscale routes on a portrait, a tarot illustration and a street scene: 0.12-0.2 is clean, 0.3 speckles skin, latent upscale is worst); Color Match keeps the draft's tone, a gentle core sharpen finishes, Compare A/B shows final against draft, and every stage bypasses on its own (each combination verified to execute); about 15 s on a 5090. An LTX 2.3 video outpaint that turns a 9:16 clip into 16:9 - the clip node pads the black bars an outpaint IC-LoRA paints new scene into, core LTX nodes only (LTXV Add Guide + Get IC-LoRA Parameters, no Lightricks pack), then Stitch Inpaint puts the original frames back from the clip's own stitcher with color_match and Save Video keeps the clip's audio; about a minute for 97 frames at 1280×704 on an RTX 5090, with a 9:16 sample clip in example_workflows/inputs/. Plus the Krea 2 one-axis outpaint and a Krea 2 text-to-image with the LoRA Loader, each with a matching preview JPG and properties.models download metadata on the loaders so ComfyUI's own missing-model check offers the files too.

1.3.0 - 2026-08-31

• LoRA Loader: strength bars you can grab. Every named row paints a center-zero bar behind its name — teal right of center for positive model strength, muted red left for negative, a brighter cap at the value's edge — on one shared scale (the stack's largest magnitude, floored at 1.0) so the everyday 0..1 range reads absolutely and one strong row rescales the whole stack instead of clipping. The name is also a scrub surface: drag it to change the strength, bar riding along; a plain click still opens the picker. Both have gear-menu off switches.

• LoRA Loader: absorb the loader chain. A gear-menu action walks the model chain on both sides of the node and lifts every recognized loader — core LoraLoader / LoraLoaderModelOnly, rgthree's Power Lora Loader, Pixaroma's loader, another AusBoss loader; Reroutes walked through — into the stack, appended below your existing rows in chain order, then bypasses the originals, so an old workflow's loader daisy-chain collapses into one node without changing what the graph computes. Names resolve against this install's list, duplicates are skipped not doubled, a fan-out stops the downstream walk, and a row imported with unequal model/CLIP strengths flips that node into separate-strengths mode so the difference stays visible.

• LoRA Loader: moved and missing files just work. A row whose file moved folders resolves by name at run time (exact → unique case-insensitive path → unique basename, one console note) and shows a dashed border naming the file the run will use; a genuinely missing LoRA warns once and skips its row instead of failing the whole run, and validation no longer blocks the queue over a missing file. The bar's new reconnect button — and ComfyUI's own R refresh, quietly — re-checks the list and rewrites repaired rows in place. Thumbnails, the info card, and range lookups all use the resolved name.

• LoRA Loader: row awareness. Rows show just the file name by default (full path in the tooltip, folders kept in the picker; gear switch to restore), and any LoRA loaded on two rows wears an amber duplicate ring — same basename under different folders is deliberately not flagged, since the stack would truly load both files.

• LoRA Loader: the master pill remembers. It now cycles mixed → all on → all off → back to the mixed setup it destroyed; every hand-made row toggle refreshes the memory, so an accidental master click is always one more click from home.

• LoRA Loader: the control bar rides the slot band. The templates/master/reconnect/gear cluster moved up into the empty middle of the output-slot band, cutting ~52px of dead space from every node; slot dots stay wirable beside it. The add button is now + LoRA and lives inside the stack container, pinned to its bottom edge. Strengths are unified by default on every new node — the separate-strengths switch is per-node and no longer leaks into the stored default.

• The README stops advertising a stale release. The front-page badge said 1.0.0 through two releases; it is now a dynamic shields.io badge that reads the version out of pyproject.toml on main at view time, so it can never go stale, and release_preflight.py gained a third check that keeps it the dynamic kind (a hardcoded badge fails preflight). The README also names the real ComfyUI-floor key (requires-comfyui).

• Image Crop + Rotate + Pad: resize the output to a megapixel budget. A new resize block (off by default) scales the transformed result to a pixel budget with core Scale Image to Total Pixels semantics — megapixels × 1024², aspect preserved, each dimension rounded to resolution_steps — using the chosen filter (lanczos, area, bicubic, bilinear, nearest-exact); the mask always resizes bilinear so feathered edges cannot ring. The big editor gains a Resize output section and its status names the exact resized size; appended after the stable V1 widgets, so saved workflows keep loading.

• Image Crop + Rotate + Pad: a quick row under the canvas. Reset (rotation, crop, and padding in one click), a Feather on/off that remembers the amount it turns off, and the Resize toggle with its megapixel box — the everyday knobs without opening the editor.

• Image Crop + Rotate + Pad: the size readout moved off the pixels. The output dimensions now sit centered just below the image (flipping above when the bottom edge leaves the stage) instead of overlapping the corner of the picture being judged, and they show the resize target too: 576 x 1024 → 768 x 1344.

• Scrubbable numbers become a pack standard. The LoRA loader's strength box grew into a shared control (js/shared/scrub_input.mjs): drag the value to scrub it, click to type an exact value, chevron arrows step, Shift is always the fine step. First adopters are the Image Crop + Rotate

  • Pad megapixel boxes (quick row and editor) and the editor's resolution steps; new numeric fields use it by convention.

• LoRA Loader: drag-to-reorder actually drops now. The row preview reparents the row mid-drag, and a reparent silently releases pointer capture — so the pointerup never landed, the drop never committed, and the row rode the cursor until a re-render snapped it back. The gesture now listens on the window for its whole lifetime instead of trusting capture.

• LoRA Loader: step arrows on every strength box. A third way to set a strength next to scrubbing and typing: small up/down chevrons step by the configured step (default 0.05); Shift steps by 0.01. Out-of-range tinting and the suggested-range tooltip carry over to the new box.

• LoRA Loader: layout matches the hand. The templates, stack toggle, on-count and settings now sit together in one bordered cluster directly on top of the row stack, and + Add LoRA is pinned to the node's bottom edge, where it stays however tall the node is dragged — the row stack flexes in between. The third output was renamed trigger_words → triggers for a narrower slot label; output links ride slot indices, so saved workflows reconnect unchanged.

• LoRA Loader: Civitai lookup actually completes. The fetch read the response with a single StreamReader.read(n), which returns whatever the buffer holds — the first ~1KB TCP chunk — not the full body. A real hit is ~150KB of JSON, so every successful lookup died mid-parse as "Civitai lookup failed" while only the 404 path worked. The body is now accumulated to EOF with the size cap enforced per chunk.

• LoRA Loader: calmer picker hover. Moving the mouse down the LoRA list used to rebuild the whole list on every row crossed — repositioning the popup, blinking the hover thumbnail, and risking the click landing on a detached row. The highlight now moves by class swap only.

• Fixed the clipped flat edge under the LoRA stack — and the same latent bug pack-wide. The frontend mounts every DOM widget's element inside a ~10px frame, so the element gets ~20 fewer CSS pixels of height than the layout allocates; the LoRA panel demanded its exact pixel sum with only 10px of slack, so the stack's rounded bottom border was clipped flat on every render. The panel now follows the node's height (fillNodeHeight) with a floor that carries the frame allowance — now a shared WIDGET_FRAME constant — and the same allowance fixed Load Video's trim strip (clipped by 22px at minimum height) and Compare's caption row (shaved by 2px). The Video Crop + Rotate + Pad fallback panel also gained the width guards every other panel already had.

1.2.0 - 2026-08-27

• New: Replace with AusBoss nodes 🆎 (prototype). A canvas-menu and command-palette action that finds third-party nodes in the open workflow — missing-node placeholders and installed types alike — and offers to swap the ones this pack can stand in for: VHS Load Video/Video Combine, KJNodes ColorMatch and GrowMaskWithBlur, LayerStyle LaMa, and the easy/Derfuu image-size nodes. Nothing changes silently: a preview lists every candidate with a per-node opt-out, widget values translate across (VHS format ids to Save Video 🆎 formats, KJ color methods to lab/mkl/histogram, frame trims to seconds where the rate is known), and anything that cannot carry losslessly is flagged in the preview instead of guessed — a frame-index trim without a known fps stays at the default with a "check trim" note. Each swap remaps links by declared name/slot mapping, rolls back on failure, and the whole apply is one undo step. Nodes with no equivalent (model loaders, WanVideo pipeline nodes, controlnet preprocessors) are listed with the reason they are refused; multi-node swaps are declared but wait for phase 2. The map and its widget translators live in js/shared/replace_map.mjs under node:test coverage.

• New node: Image Resize 🆎. The single most common reason a shared workflow drags in a heavy pack, as one dependency-free node: target an exact width+height, a longest or shortest edge, a megapixel budget, or a scale factor, then stretch, fit, cover-crop, or pad when the aspect changes — pad fills with a color and marks the new bars as 1.0 in the mask output, the pack's usual generated-area contract. divisible_by snaps the result to a clean multiple (16 for WAN), 0 in a size widget keeps the source, an optional mask rides through the identical transform, and resampling is lanczos (PIL, in float), bicubic, bilinear, nearest, or area.

• Added the Utility group — nine small nodes for the gaps that used to mean installing a 200-node pack for a text box: Text 🆎, Integer 🆎 and Float 🆎 (typed constants on their own wires), Show Text 🆎 (shows the string it receives on the node face — selectable for copying, saved with the workflow — and passes it through), Math Expression 🆎 (arithmetic over a/b/c with FLOAT and INT outputs, parsed with ast against a whitelist and never eval'd, so a shared workflow cannot smuggle code through it), Select Every Nth 🆎, Split Batch 🆎 and Merge Batches 🆎 (IMAGE-batch thinning, splitting and joining, with an explicit resize policy instead of a silent one when sizes differ), and Free Memory 🆎 (a wildcard passthrough that unloads comfy's models and empties the CUDA cache between heavy stages — every step fail-soft and imported at run time, so a core API move skips the step instead of deleting the node).

• Video I/O polish: format-aware Save Video, drop-to-restore, an honest Load Video label. Save Video's face now follows the chosen format — crf hides for mov prores, mkv ffv1 and gif, which ignore it, and save_metadata hides for gif and webp, which cannot carry it; hidden widgets keep their position and value, so saved workflows are untouched and switching back restores the number. Dropping a video onto a Load Video node (as opposed to onto empty canvas, which still restores the whole embedded workflow) makes it that node's source — copied into the input folder, identical re-drops reusing the existing file — and an AusBoss save also restores the trim, every_nth, max_frames, sizing and FRAME values its embedded workflow stored. And since the preview cannot re-render frame drops, its label now reports what one Run will actually load — 0:04.0 of 0:10.0 · 48 frames @ 12 fps, or 1 frame at 0:05.2 in FRAME mode — computed from the source's probed frame rate with every_nth and max_frames applied, and omitted rather than guessed when a deciding value arrives over a link.

• LoRA Loader: rows drag to reorder. Every row grew a dotted grip; drag it and the stack shuffles live under the pointer, committing to the serialized widget only on drop — an abandoned drag never dirties the workflow. LoRAs apply in row order, so order is part of the recipe; the right-click Move up/down items remain for one-step moves.

• LoRA Loader: hover thumbnails on rows, not just in the picker. Hovering a row's name floats the LoRA's sidecar preview image beside the cursor, the same way the picker list already did; both now share one floating element that follows the pointer and flips sides at the screen edge. It appears only once the image has actually loaded, so a LoRA with no preview file shows nothing instead of an empty bordered box, and nothing on the node ever shifts.

• LoRA Loader: the info card states the file's size and modified date, read from disk by the same route that serves its trigger words — quick ground truth for "which of these two 143 MB epochs is the newer one".

• New node: Save Image 🆎. PNG or lossless JPEG XL (optional pillow-jxl-plugin, listed as the pack's jxl extra), workflow embedding on or off — and an exact_name mode that saves under exactly that filename with no counter suffix, so a caption or edit pass keeps the source file's name: photo123.jpg in, photo123.png out, paired with the photo123.txt sidecar the optional caption input writes. on_existing decides overwrite/skip/error, output_dir accepts a subfolder of ComfyUI's output or an absolute dataset path, and the node returns the saved file_path. Classic prefix+counter saving that never overwrites remains the default.

• The registry description now names the nodes. Manager search matches against the description text, so "a curated collection of polished nodes" told searchers nothing — it now lists every node family, and the keywords grew to match.

• Image Resize never invents pixels unless explicitly asked. In every target mode except width+height the box is derived from the source's own aspect, so there is nothing to letterbox against — yet pad used to answer a divisible_by snap by inventing a sliver of bar (one bottom row on an 855×480 source at longest_edge 512, /16). All scale-derived modes now resolve the snap the way fit always has: an invisible sub-half-step resize, a black mask, nothing for the user to think about. Bars — and white in the mask — can only appear when a width+height box that disagrees with the source is combined with pad, which is the one place they are the explicit request.

• Added example_workflows/ausboss_node_tour.json — a model-free tour that wires 17 node types into one runnable graph: load and thin a clip, split and rejoin the batch, pad-resize with the bars masked, de-flicker against the first frame everywhere except those bars, retime back to double rate, free VRAM, and save — with the fps, frame budget, filename prefix, and CRF all arriving over wires from Math Expression, Image Size, Text, and Integer nodes, the saved path landing in Show Text, and first vs last frame in the A/B panel. Pick any video and Queue; nothing else is required.

• The flagship node color deepened to slate-teal. The pack-wide default scheme now pairs a deep slate-teal title (#14424d) with a softly lifted near-black body (#161f21) — quieter on a busy canvas than the original bright teal. Nodes in saved workflows still wearing the old pair upgrade automatically on load (LEGACY_SCHEME_PAIRS); colors a user picked by hand are untouched, as ever.

• An optional completion chime. A new off-by-default setting (🆎 AusBoss → Notifications → Completion sound) plays a soft two-note WebAudio chime when the prompt queue empties, so a long video render can run unwatched in another window. No audio asset ships; the tone is synthesized on the spot.

• Stale text and metadata cleaned up across the pack. Image Compare A/B's description and tooltip now describe the A/B toggle instead of the removed hold mode; Krea 2 Outpaint Model Patch gained the RETURN_NAMES its socket label was missing; seek_mode, crop_x/crop_y, and the four pad_* inputs gained the tooltips the pack's own rules require; Frame Interpolate no longer answers a "rife" search it cannot honor; and LaMa Inpaint moved into the Inpaint category beside its crop/stitch companions.

• Four nodes gained the outputs real workflows kept asking for. Stitch Inpaint now also returns blend_mask — the feathered paste band in original-image coordinates, so the docs' own stitch-then-Color-Match loop wires directly. Save Video, which had no outputs at all, returns the saved file's absolute path for chaining. Select Frame accepts negative frame numbers counting from the end (-1 is the last frame — the feed-the-last-frame-to-I2V move). Image Size adds a count output for the batch size. All outputs are appended, so saved workflows load unchanged.

• Six help pages caught up with their nodes. Color Match's page (which still described a single-method node), Align Image, Crop For Inpaint (ten undocumented inputs), Load Image + Pad (the Krea 2 outputs), LM Studio Chat (the history output and gear-menu controls), and Load Video (every_nth / max_frames) now match what ships.

• Fixed tests/test_inpaint_crop_helpers.py ending its direct run at a mid-file unittest.main(): the two canvas-stitcher test classes defined below it never executed. The block moved to the end of the file; all 68 tests (up from 57) run and pass.

• Load Video can pick a single frame. A FRAME button on the preview turns the trim strip into a frame picker: only the frame at the marker loads, as a one-image batch ready for image workflows. Click or drag the rail to scrub, or type an exact time into the AT field; playback runs the whole source freely while picking, and the trim window comes back untouched when the toggle turns off. Backed by a single_frame widget appended after the existing inputs, so saved workflows keep loading unchanged.

• Image Compare A/B: nothing is drawn over the picture any more. The status chip that sat in the top-left corner carried the resolution and a hint about how to use the panel; it covered part of the image to say something you only need once. The resolution moved to a caption centred under the panel, and when the two sides are different sizes both are named there - the panel scales them to fit, so nothing else on screen would show it. The empty and error states still use the middle of the stage.

• Image Compare A/B's HOLD became an A/B toggle. Press-and-hold meant the comparison only existed while a mouse button was down. Each click now swaps the whole panel and the button says which side you are looking at, so you can flick between them - which is how a small difference actually becomes visible. A workflow saved on hold opens on the toggle it became rather than dropping back to slide.

• Updated example_workflows/simple_video_watermark_remover.json: the overview still credited SimpleWatermarkRemover (the deprecated compatibility id) rather than LaMa Inpaint 🆎, still called the node Refine Mask, and its GitHub link pointed at github.com/auboss, which does not exist. It also gained model-setup and requirements sections, consistent node titles, and the mask settings its single-frame test was missing - that path shipped with every Mask Refine control at zero, so the frame it previewed was not the result the full run would produce.

• Renamed Refine Mask 🆎 to Mask Refine 🆎. The mapping key (AUSBOSS_NODES_RefineMask) is unchanged, so saved workflows are unaffected, and "refine mask" is now a search alias. ComfyUI's node search cuts at 64 results and only ranks a name that STARTS with what you typed near the top - as "Refine Mask" this node came 121st for "mask" and was never on screen. It is 15th now. Search aliases cannot fix this on their own: the frontend indexes them but drops them from that ranking because they are a list rather than a string, so they decide whether a node is found at all, not where.

• Mask Refine, LaMa Inpaint and Select Frame preview their own result, in a panel that now follows the node's height like every other stage in the pack. It was pinned at 140px, correctly, while it showed a small thumbnail of the node's INPUT; showing the result makes it a viewport onto a picture, and the guard that classifies panels as growing or fixed has been updated to match - moving a panel between those two sets is now explicitly the call to review whenever what a panel displays changes, not only when one is added. The panel used to show whatever fed the node's input, which is nothing at all when a segmentation node is upstream - Mask Refine's panel was blank however the mask turned out. Each now returns its result as a preview and the panel shows that, falling back to the input thumbnail before the first run. ComfyUI's own preview for these nodes is stood down, so the picture appears once, in the panel, instead of also underneath the node.

• Fixed: LaMa Inpaint finished a single image with an empty panel. Frame previews were only attached when the batch had more than one frame, on the assumption that one image would preview through the output path - which does not exist for a node that is not an output node.

• Mask Refine opens on expand and blur, with fill_holes, smooth, black_point, white_point and edge_refine behind a MORE button. Hidden widgets keep their values and their saved order, so nothing changes for a workflow that set them. A new AUTO button sets expand and blur from the mask's size, scaled off the 8/4 that was hand-tuned for the video watermark workflow at 576px: a feather is a fraction of the picture, not a fixed pixel count.

• Save Video gained formats, pingpong and a metadata switch. Alongside mp4 h264/h265 and webm vp9 there are now mp4 h264 nvenc and mp4 h265 nvenc (GPU encoding), webm av1, mov prores (ProRes HQ), mkv ffv1 (bit-exact lossless), gif and webp. pingpong bounces the clip for a seamless loop without repeating the turnaround frames, and save_metadata turns off embedding the prompt and workflow. Both are appended after format, where positional widget values cannot shift. fps steps in whole frames now but stays a FLOAT - an INT would refuse Load Video's fps link and would round 29.97 to 30, drifting picture away from audio. mov files now carry the embedded workflow too; they silently did not before.

• Removed Save Video's ↻ reload button. It only ever re-fetched the preview already on screen, and after a page reload there was no saved file in memory for it to fetch at all.

• Frame Chooser 🆎 was removed from the public pack (AUSBOSS_NODES_FrameChooser is gone). Pausing the graph for an interactive pick is a bigger surface than the rest of the pack - a server route, a resumable pause, a panel that has to survive a reload - and it has open issues that are not worth holding a release for. A saved workflow using its id will report it as missing.

• Renamed Video Crop + Rotate + Pad 🆎 to Video Crop + Rotate + Pad → Frame 🆎. The old name reads like it transforms a clip; it takes one frame out of a video and returns a single image, and the editor's timeline is there to find that frame rather than to trim a range. "grab frame", "extract frame" and "frame from video" join the search aliases. The mapping key AUSBOSS_NODES_VideoCropRotatePad is untouched — it is the workflow-compatibility contract — so saved workflows are unaffected.

• Video Crop + Rotate + Pad 🆎: the preview no longer stretches when the node is resized. Its canvas is CSS-stretched to whatever box the panel is given, but the backing store is only re-sized when something draws — and the panel's resize observer sat inside an image-only branch next to the interactive drag handlers, so the image node redrew on resize and the video node, a passive preview with no other reason to redraw, kept painting its last frame into a box that had changed shape. Latent until panels started taking the node's leftover height, because before that the box could not change shape without a width change forcing a relayout. A test now fails if the observer is gated on node kind again, or if a canvas-painting panel ships without one.

• Added Krea 2 Encode 🆎 and Krea 2 Outpaint Model Patch 🆎. Together they make Krea 2 outpaint the source instead of painting something next to it: the encode attaches reference latents to the positive conditioning (and emits the negative from the same node, so a turbo graph at CFG 1.0 stops carrying a second text encode that does nothing), and the patch registers those reference tokens into the target grid at the rectangle the stitcher records. Reference latents with no position are only a style hint — the model borrows the look and reinvents the content, which is the failure this pair fixes.

  The patch reaches into comfy's flux attention layers, so it imports them when you run it rather than at startup: a core release that moves one surfaces as an error on that node instead of deleting it from the menu and leaving saved workflows reporting it missing.

• Krea 2 Outpaint Model Patch 🆎 warns when the source is padded on both axes. The model places the source spanning one whole canvas axis; the reference pipeline splits anything else into two passes, and doing it in one is not a slightly worse result — the extended region breaks up. The warning reports the spare pixels on each axis, because the usual cause is not a deliberate second pad but canvas_multiple rounding the other axis up by a few pixels, which breaks the span just as completely.

• Load Image + Pad 🆎 gained a reference output — the unpadded source, fitted to a 384px long edge and a multiple of 16, ready for reference conditioning. It is appended after stitcher, not inserted, because a workflow stores links by output slot index and anything else would silently rewire every saved graph.

• The stitcher now records where the source sits on the canvas, as source_bbox in pixels and bbox_normalized in 0..1. Padding knows the rectangle exactly, so it rides along with the canvas rather than on a parallel wire that can be left unplugged. Stitching itself never reads either key, so an older stitcher still stitches identically.

• Removed Drop Shadow 🆎 (AUSBOSS_NODES_DropShadow). The result never looked like a real cast shadow — a mask offset, grown and blurred has no contact darkening and no perspective, so it read as a sticker halo rather than something in the scene, and that is a limit of the approach rather than a tuning problem. A saved workflow containing the node will report it as missing on load; delete it, or composite the shadow in an image editor.

• Removed Pad Image 🆎 (AUSBOSS_NODES_PadImage). Load Image + Pad 🆎 does the same job from the same stage and starts from the file, so keeping a second node whose only difference was taking an IMAGE wire earned its slot in the menu twice over from one idea. A saved workflow containing it will report it as missing on load. The padding helpers, the modes and the mask are unchanged — they were always shared, and Load Image + Pad keeps all of them plus feather, canvas rounding and the megapixel target.

• Load Image + Pad gained a stitcher output, so an outpaint can put the source back exactly. Feed it to the existing Stitch Inpaint 🆎 with the sampled result and every pixel outside the padded band comes back bit-identical to the input — only the new padding is the model's work, which is what stops a full-canvas sample from quietly resoftening the whole photo. No new node: padding now builds the same stitcher shape Crop For Inpaint emits (the crop is simply the whole canvas), so one stitch node serves both. The output is appended last, so saved workflows keep their existing links.

• Compare: the A/B stage now grows when you drag the node taller, not only wider. Its panel declared a computeSize, and the widget layout gives any widget that defines one a fixed height and leaves it out of the leftover-space split — so the stage was sized purely from the node's width (capped at 520px) and extra height became dead space under the image. It now declares only a minimum, which puts it in the split and lets it take the height that is left. The node also opens at a 16:9-ish default instead of inheriting one from the removed calculation; saved workflows keep their own size.

• LoRA Loader: the "wrong base model" warning now measures the result instead of guessing from names. It compares comfy's applied-patch count across the model and CLIP before and after each row, and warns — naming the LoRA — when a row patched nothing, which is what a mismatched LoRA actually does. The old check keyed off comfy's model class name against a hardcoded table that covered 8 of the 81 classes comfy ships, so on anything newer than Flux (Krea 2, Qwen, WAN, LTXV, Z-Image, Chroma, HiDream, …) it read "unknown" and silently skipped the check; an SD 1.5 LoRA on Krea 2 did nothing to the image and said nothing about it. The new check needs no table, no LoRA metadata, and works on every model family.

• LoRA Loader: a LoRA's declared base model is read only from the declarative metadata keys. ss_sd_model_name is the trainer's source filename, and mining it for substrings labelled any ..._v1.safetensors as SD1.5 and anything with xl in the name as SDXL — a wrong label on a working LoRA. Families newer than the known-name table now show what the file declares (krea2) instead of nothing.

• LM Studio Chat: reasoning that the server returns in its own reasoning_content field now reaches the thinking output. Reasoning models report two ways — inline <think> tags inside the content, or that sibling field (what LM Studio sends for gemma/qwen-style hybrids) — and only the first was read, so with those models the entire reply landed in a field the node never looked at and text came back silently blank.

• LM Studio Chat: an empty answer now says why instead of returning "". When the model produced reasoning but no answer, the node reports whether it ran out of tokens mid-thought (naming the max_tokens budget it hit) or simply answered nothing, and points at the fix. A reasoning model can spend an entire small token budget thinking, which read as "the node is broken". instead of magic numbers. Each of Top-p, Top-k, Min-p, Repeat penalty and Presence penalty has a checkbox, a reset button that appears once you change it, and — for the 0-1 ones — a slider beside the number. Ticking one on starts from LM Studio's own default (top-p 0.95, top-k 40, min-p 0.05, repeat 1.1) rather than the value that means "off", and ticking it back on returns the value you had before. Unticked still sends nothing, so the payload, the widgets, and existing workflows are byte-for-byte unchanged; what changed is that "off" now looks off instead of requiring you to know that top-p 1 happens to mean off.

• Settings menus: editing one row no longer reverts the others. The menu seeds from the open node's values, but each save handed back a value set rebuilt from stored defaults, so changing any one setting silently replaced every other row with whatever the stored default was — visible as a value quietly reverting on a node whose saved workflow differed from those defaults.

• LoRA Loader: fixed the templates popover — a blanket .ausboss-lora-menu button rule outranked the specialized buttons inside it, forcing them to width: 100%, which clipped Save off the panel edge and stretched each saved row's delete × across the whole row. The rule now targets only the menu's direct children, which are the plain list rows. Action buttons also read as buttons: matching 26px height with the name field, a border, and pressed/keyboard-focus states.

1.1.1

• Registry metadata only; no node behaviour changes. The ComfyUI version floor now actually reaches the registry: comfy-cli reads it from a [tool.comfy] key named requires-comfyui, and the pack had been declaring supported_comfyui_version, which is silently ignored - so 1.1.0 published with an empty floor and Manager would not have refused an install on an incompatible core.
• Declared Operating System :: OS Independent so the registry records the supported-OS list. Accelerator classifiers are deliberately omitted rather than asserting untested hardware.
• License publishes as a readable name instead of the literal string {"file": "LICENSE"}. The LICENSE file itself is unchanged, still MIT.

1.1.0

• Image Crop + Rotate + Pad: the node's compact preview is now the editor stage in miniature — grab the crop squares, pad diamonds, and rotate knob (with a live degree readout) right on the node, and the panel grows with the node. Fit-only there, so the wheel keeps zooming the graph; the full editor keeps zoom, pan, and the sidebars, and both surfaces run the same hit-test and drag code, so they cannot drift. The video node's panel stays a passive preview.

• Added AUSBOSS_NODES_LoadImagePad: a Load Image with an on-node outpaint canvas — drag any edge of the final rect to set that side's padding (the whole edge is the handle; the badge shows the true output size after rounding), with the four Pad Image fills, a mask feathered inward across the seam, canvas-multiple rounding, and a megapixel target that rescales the source before padding so the mask seam stays crisp. Outputs image, mask, and the final width/height as INTs.

• Pad Image: the same on-node handle canvas — drag the final rect's edges to set the padding over a live preview of the input image (fed by execution, so it fills in after the first run; a wireframe stands in before that). Widgets, outputs, and saved workflows are unchanged.

• Drop Shadow: a blend choice — normal (the old mix toward the color) or multiply, which darkens the backdrop by the color and keeps its texture — plus a shadow_mask output carrying the effective shadow alpha for compositing downstream.

• Align Image: offset_x/offset_y INT outputs locate the original's top-left inside the aligned output (positive after pad, negative after crop, 0 after resize), so an un-align crop after sampling needs no manual math.

• Color Match: reference_mode: first_frame matches every frame of a batch to the batch's own first frame — the one-node video flicker fix; the reference input is optional in that mode.

• Crop For Inpaint: target_megapixels rescales the crop to a sampler-friendly area (explicit target_width/height still wins), rescale_algorithm picks the resize filter for both directions of the round trip (recorded in the stitcher), and extend_left/right/up/down grow the frame itself for outpainting — the new bands are replicate-filled, masked for painting, and become part of the stitched output.

• Save Video: a format choice — mp4 h264 (default, unchanged), mp4 h265, or webm vp9 with Opus audio (resampled to 48 kHz when needed). CRF applies to all three; bt709 tagging and the embedded workflow ride along.

• Load Video: every_nth keeps one frame in N (the fps output divides to match, so real time survives downstream), and max_frames stops the decode after that many kept frames instead of loading and discarding — long clips no longer have to fit in memory.

• LoRA Loader: the bar's ▤ button saves and applies named templates of the whole stack (browser-persisted, case-insensitive replace, sorted menu). A LoRA whose metadata names a different base-model family than the connected checkpoint now logs one clear console warning at apply time.

• LM Studio Chat: a history input/output pair (AUSBOSS_CHAT_HISTORY) chains multi-turn conversations across chat nodes — reasoning blocks and image payloads are deliberately not replayed — and a json_schema widget forces structured JSON replies via LM Studio's response_format.

• Cleanup: the dead module-level NODE_ID variables left over from before the literal-mapping-key convention are gone from every node file.

• Removed: AUSBOSS_NODES_SelectFrameRange and the Video Bundle family (AUSBOSS_NODES_VideoBundle, AUSBOSS_NODES_VideoUnbundle, AUSBOSS_NODES_VideoBundleEdit). Core ComfyUI's ImageFromBatch covers contiguous-range selection, and the core VIDEO wire (CreateVideo / GetVideoComponents) is now the ecosystem-standard way to move a whole video on one connection, which is what AUSBOSS_VIDEO existed for. Workflows using the removed keys keep their other nodes; replace those four with the core equivalents. Select Frame and Image Size stay.

• Display names traded the " (AusBoss)" suffix for the pack's 🆎 signature — "LoRA Loader 🆎", "Color Match 🆎". Typing "ausboss" still surfaces everything through the 🆎 AusBoss category, the AUSBOSS_NODES_ id prefix, and each node's search aliases; mapping keys are untouched, so saved workflows load exactly as before.

• Fixed: the LM Studio Chat endpoint toolbar rendered as a sliver that clipped its buttons — the DOM-widget wrapper takes its height from the getMinHeight option, which the toolbar never declared.

• Every AusBoss node now wears the brand look out of the box: the teal-title scheme the video nodes shipped with became the "AusBoss" row in the appearance table and the pack-wide default, and Load/Save Video stopped hard-painting themselves so the appearance setting (and the per-node color menu) governs them like everyone else. "Theme default" remains available for anyone who wants uncolored nodes.

• Every AusBoss node grew a quiet "?" badge in the title bar; clicking it opens a card built from the node's own DESCRIPTION and input/output tooltips, so the docs on screen are exactly the docs in the source.

• New gear-settings menus, persisted in the browser: the LoRA Loader's gear holds default strength, strength step, separate model/CLIP strengths, the trigger-word separator, hide-extension, thumbnail, and Civitai-lookup preferences; LM Studio Chat's gear holds the advanced sampling knobs (top-p, top-k, min-p, repeat and presence penalty), thinking control with custom reasoning tags, LM Studio idle-unload TTL, and a free-ComfyUI-VRAM-first switch. The LM Studio values ride hidden standard widgets, so they save with the workflow and reach the API like any widget.

• LoRA Loader restyle: one control language — a full-width filled Add button, a master-toggle bar with the gear, and the rows inside an inset stack container with a dashed empty state; the serif "i" is gone.

• Align Image: crop mode gained crop_position (center/top/bottom/left/right) choosing which part of the frame survives; the widget only shows while mode is crop.

• Color Match: a method choice — lab (the old behavior), rgb, mkl (full covariance mapping), and histogram (exact per-channel distribution) — plus invert_mask, and tooltips that spell out the mask contract (it scopes the fix and passes through unchanged, which is why there is no mask output).

• Crop For Inpaint: the selection can now be inverted (invert_mask), grown or shrunk (mask_grow), and edge-softened (mask_blur) before cropping, and context_pixels adds flat margin on top of context_factor's growth.

• Fixed: A/B Compare could get stuck in HOLD — in hold mode the stage captured the pointer on press, which retargeted the release and ate the mode button's click. The modes are now two dedicated SLIDE/HOLD buttons, and presses that start on the toolbar never reach the stage behaviors.

• Fixed the LoRA panel (and every DOM panel) overflowing after a node was resized narrower: the frontend sizes a panel's wrapper as widget.width ?? node.width, and LiteGraph's layout plants widget.width during draws - once planted, it outranks the node width forever, so the wrapper kept an old, wider width and parked the row's controls outside the border. All six panels now discard those writes (keepDomWidgetWidthAuto), so the wrapper tracks the node in both directions. Diagnosed from a live browser measurement and verified end-to-end against a planted stale width.

• Fixed: the LoRA Loader's strength box and info button could hang past the node's right edge. The panel now sizes its padding inside the widget's box, clips anything oversized, and declares its minimum width to the layout so the node cannot be resized out from under the row. A pack-wide test now requires every DOM panel to carry the same guards.

• The same containment sweep covered every DOM panel: Frame Chooser gained the resize floor older frontends read (minNodeSize), the input-preview thumbnail clips at its root, and the pack-wide test now requires the full guard set - border-box, an overflow clip, and a minimum width on both frontend layout paths - of every panel.

• Added AUSBOSS_NODES_AlignImage: snap an image's width and height to a clean multiple (16, 32, ...) by nearest-resize, center-crop, or replicate-pad, with the new size as INT outputs — for Qwen image models and anything else that wants cleanly divisible sizes.

• Added AUSBOSS_NODES_ImageSize: width, height, longest edge, and shortest edge as INT outputs.

• Added AUSBOSS_NODES_LmStudioChat: prompt + optional image to a local LM Studio (or any OpenAI-compatible) server. Empty model uses whatever is loaded, <think> blocks land on their own output, the seed re-rolls the cached reply, and every error names what to fix. Stdlib HTTP - no new dependencies.

• Performance: mask dilation and erosion run as one separable pass instead of one 3x3 pool per pixel of growth - bit-identical, measured 6-7x at 32 px, and it feeds Drop Shadow's grow, Refine Mask's expand, and the matting trimap. The trimap's morphology is also built once per batch instead of once per frame (27x on that stage: 11.0 s -> 0.4 s for 48 frames of 832x480). Pad Image's pillarbox backdrop blurs at quarter resolution when the blur is heavy (6-7x on the stage that was ~90% of the node; mean difference ~0.001 in a backdrop that is then dimmed - light blurs keep the exact full-resolution path). Frame Interpolate estimates optical flow exactly once per source pair however small batch_size is, where a batch_size of 1 used to re-solve each pair once per output frame (4x the RAFT work at 24 -> 120 fps); the guide-image batch is no longer duplicated up front, and the scene-cut scan drops a full-chunk temporary.

• Fixed: a Frame Chooser pause survived being cancelled by an ordinary workflow load. LiteGraph clears the graph by removing every node, so undo, switching workflow tabs, Clear Workflow and opening another file all fired the teardown that a deleted node uses to release its pause - silently interrupting a run that was still going, with no way to get it back.

• Fixed: answering "keep all" wrote the whole batch out as 1,2,...,N into pick_list, pinning a batch-size-independent answer to one batch. The next run then dropped any frames past the end of that list, or failed outright on a shorter clip. It now writes the empty answer it was given.

• Fixed: pick_list is only written back under keep last selection. It pre-answers the node, so filling it in automatically meant a chooser left at the default always pause paused exactly once and never again.

• Fixed: Refine Mask's smooth no longer flattens a mask that is already soft; it applies only the change it made to the jaggy edge. Binary masks are bit-identical to before.

• Fixed: Refine Mask's blur now reaches the matting edge-refine solve instead of being thresholded back out of the trimap.

• Fixed: Frame Interpolate's copy path honours batch_size. It gathered every copied frame at once, which on a long clip at an integer multiple was a multi-gigabyte allocation on the frames device that no setting could bound.

• Fixed: live status and runtime badges appear for nodes inside a subgraph, and a subgraph pause renders on the node that actually paused. Colon-prefixed execution ids are resolved by walking the subgraph chain rather than by stripping the prefix, which could match an unrelated node with the same number.

• Fixed: choosing the Custom node colour scheme while the stored colour is unreadable no longer strips the colour off every AusBoss node.

• Fixed: an unreadable colour in Drop Shadow or Pad Image names its own node and widget in the console instead of blaming Transform, and one node's warning no longer silences another's.

• Fixed: a malformed Frame Chooser answer returns 400 rather than 500 - a non-ASCII token and a JSON body that is not an object both used to throw out of the route and strand the pause.

• Fixed: pick_list rejects Unicode digits. ² raised a bare ValueError during validation and ٣ was silently read as frame 3.

• Fixed: scripts/validate_nodes.py parses NODE_MODULES instead of matching it anywhere in the text, and an unregistered node module is now an error rather than a warning - it used to exit 0 with no output at all. It also refuses mapping keys declared outside nodes/, which is how the registry test fixtures came to be advertised to ComfyUI-Manager as installable nodes; they are now .py.txt.

• Added AUSBOSS_NODES_ColorMatch: LAB mean/std transfer that harmonizes an inpainted or stitched region against its source, with optional mask and strength.

• Added AUSBOSS_NODES_PadImage with color, edge, edge-pixel, and pillarbox-blur fills, returning a mask over exactly the new padding for outpainting.

• Added AUSBOSS_NODES_DropShadow for padded and reframed compositions.

• Added AUSBOSS_NODES_FrameInterpolate: fps-based interpolation (24 to 30 works, not just whole multiples) with blend and optical-flow methods, bounded memory, and scene-cut detection that holds across hard cuts instead of morphing.

• Refine Mask gained a jaggy-melting smooth control, black/white point levels, and optional guided filter and matting edge-refine tiers.

• Stitch Inpaint gained an optional fix_edge_halo toggle that removes the rim left by compositing a feathered seam twice; pixels outside the blend stay bit-identical either way.

• Frame Chooser gained a countdown with timeout policies, reload recovery, a pick_list pre-answer for headless reruns with automatic writeback, stale-answer rejection, a keyboard map, and a notice when a pause begins out of sight. A pause now resolves exactly once: whichever of a keep, a cancel, an expiring countdown or a second tab gets there first decides it, and the ones that lose are refused instead of overwriting the decision or reporting a success they did not have.

• Load Video exposes a lazy core VIDEO output and Save Video accepts a core VIDEO input, so the pack interoperates with ComfyUI's own video nodes; a connected video's frame rate wins over the fps widget.

• Video decode and encode now run off the executor thread with per-frame progress, so long jobs no longer block the UI.

• Added a live per-node status badge (frame i/N during a LaMa video inpaint), About-page badges, a toast for the stale-frontend warning, and a Custom node color scheme.

• Fixed registry discovery: mapping keys are now string literals, so ComfyUI-Manager can see the pack's nodes and offer to install it for a shared workflow. scripts/validate_nodes.py now enforces the whole registry contract - each mapping assigned exactly once at module level to a non-empty dictionary literal with string-literal keys, its name never used again (no update(), no del, no aliasing it into a variable that mutates it later), matching keys across the two mappings, and no key claimed by two modules. Permanent public ids are checked against the keys parsed out of those literals rather than any mention of the id in the file.

• Added AUSBOSS_NODES_FrameChooser: pause the graph on a clickable filmstrip and keep only the frames you pick, with a no-pause "keep last selection" mode.

• Added AUSBOSS_NODES_CropForInpaint + AUSBOSS_NODES_StitchInpaint: native-resolution masked inpainting with a bit-exact paste-back contract and video batch broadcasting.

• Added the AUSBOSS_VIDEO bundle wire (Video Bundle / Unbundle / Bundle Edit) carrying frames, audio, fps, and derived info on one connection.

• Added AUSBOSS_NODES_Compare: slide or hold A/B image comparison that passes A through.

• LoRA Loader: master on/off pill with a mixed state, folder-grouped picker with hover preview thumbnails and shared-prefix stripping, and per-LoRA suggested strength ranges that tint out-of-range values.

• Load Video: trim IN/OUT are typed timecodes (h:mm:ss.s), decodes are memory-guarded with a clear oversized-trim error, seeks respect stream start time, and audio extraction is lazy.

• Save Video: output is tagged bt709 with a matching conversion matrix, and dropping a saved mp4 onto the canvas restores its embedded workflow.

• Widget values for the transform and video nodes now serialize by name with validated migration from older positional workflows.

• New Chrome settings: favicon/tab-title queue status and per-node runtime badges.

• Added AUSBOSS_NODES_LoadVideo with a single responsive player, draggable IN/OUT trim range, bounded playback, matched audio, and info outputs.

• Added AUSBOSS_NODES_RefineMask with expand/shrink, hole filling, feathering, and an inverted output.

• Added AUSBOSS_NODES_SaveVideo writing H.264 mp4 with muxed audio, embedded workflow metadata, and a responsive loopable result viewer.

• Added AUSBOSS_NODES_SelectFrame with one-based, range-checked batch selection.

• Added AUSBOSS_NODES_LaMaInpaint with bounded-VRAM video processing and explicit models/lama checkpoint discovery.

• Added a compatibility alias for the published SimpleWatermarkRemover workflow contract.

• Added the repaired Simple Video Watermark Remover example workflow.

• Added an AusBoss node color setting (Settings → 🆎 AusBoss) with curated schemes that recolor every AusBoss node live; hand-colored nodes keep their own colors.

• Added AUSBOSS_NODES_LoraLoader: a stacked multi-LoRA node with drag-to-scrub strengths, a keyboard-first searchable picker, per-row trigger words (file metadata, one-click Civitai fetch, and your own saved words), and a trigger_words output.

• Added AUSBOSS_NODES_SelectFrameRange returning a one-based sub-batch plus its actual frame count.

• Added a right-click Recreate node (AusBoss) utility that rebuilds a node from the current definition, preserving values and links, with full rollback.

• Added adaptive title ink, a per-node right-click AusBoss color override, and subgraph-aware color sweeps.

• Added live previews: LaMa video inpaints stream each finished frame to the node face, and Refine Mask / LaMa Inpaint show their upstream input before the graph runs.

• Added %date:...%-style filename tokens to Save Video, tolerant fill_color parsing (hex, CSV, floats, CSS names), an Alt+E open-editor command, and user-editable aspect presets via ausboss_presets.json.

• DOM panel edits now register with ComfyUI's undo/modified tracking, and a console warning fires when the browser runs stale cached pack JavaScript.

• Added scripts/release_preflight.py catching pyproject BOMs and JS/Python version drift.

1.0.0

• Added AUSBOSS_NODES_ImageCropRotatePad.
• Added AUSBOSS_NODES_VideoCropRotatePad.
• Added a shared full-screen rotate, crop, and pad editor with compact node previews.
• Added exact video-frame preview routes and input-folder/local-path modes.
• Added generated-area masks covering transparency, rotation voids, and padding.
• Added rich node help, example workflows, automated backend/frontend tests, and Registry metadata.
• Editor rotated-size math now matches Pillow's expand=True output exactly at every angle.
• Video frame preview uses keyframe seeking with a sequential fallback, so scrubbing long videos stays fast.
• Editor previews of local paths outside ComfyUI's folders are opt-in via AUSBOSS_TRANSFORM_LOCAL_PREVIEW=1; queued workflows are unaffected.
• Timeline scrubbing now renders immediately with a latest-wins request pump and reduced-size scrub frames, landing a full-resolution frame on release; playback and held arrow keys use the same light path. Server caches per-file video metadata so each scrub frame opens the container once.
• Video decodes run off the web server's event loop in persistent per-file decoder sessions (with an idle reaper that releases file locks), so a slow decode can never stall the ComfyUI UI and stepping forward decodes only the frames in between.
• A keyframe storyboard builds in the background after a video is selected; dragging the timeline shows the nearest storyboard tile with zero network latency, then the exact decoded frame replaces it.
• The rotation handle moved to the source's top-right corner, drawn with a crisp vector rotate glyph, clear of the top padding handle.