Repository navigation
Releases
Scumble 0.1.59: templates, Approve everything and a hundred steps for the assistant
Get Scumble
| Package | Download | Size | Notes |
|---|---|---|---|
| Microsoft Store (Windows) | Scumble in the Store or winget install scumble |
– | Signed by Microsoft and updated by the Store; a new version can reach the Store a few days after GitHub (Microsoft's certification). |
| Windows installer | Scumble-Setup-0.1.59.exe | 124 MB | Updates itself; not code-signed yet, so SmartScreen asks once. |
| Windows portable | Scumble-0.1.59-portable-win-x64.zip | 175 MB | No install: unpack and start Scumble.exe; keeps its data in a folder beside it and does not update itself. |
| Linux AppImage | scumble-0.1.59.AppImage | 152 MB | Early build made by CI, not tested by the author yet. |
| Linux .deb | scumble-0.1.59.deb | 120 MB | Early build made by CI, not tested by the author yet. |
| Claude Desktop extension | scumble-0.1.59.mcpb | 83 KB | Adds Scumble's MCP tools to Claude Desktop with one click; needs Scumble installed (any of the above). |
The other files of this release (latest.yml, latest-linux.yml and the .blockmap) are for the app's updater; you do not need to download them.
- The assistant: templates. A task you give more than once (upscale, a detail pass on the eyes, a grade) can be
kept under a name: Templates beside Send lists them, and a click puts one into the chat field, where you can
change it before you send it. A template is only its text; it never switches Approve everything on. - The assistant: Approve everything. A switch beside Send takes the questions away for one chat, for a task
of many steps you do not want to answer card by card. It asks once whether you mean it; after that every call
runs without a question, paid runs and files included, and each card says what it would have asked. What the
assistant refuses stays refused, and after 30 paid or queued runs in one turn the next one asks again
(Settings › Assistant sets the number). A new chat asks as before. Mind what it means: the model reads layer
names, file names and text in the picture, and with everything approved such text can make it spend money or
overwrite a file. - Fixed, from the code review of 2026-11-16 (the editor's part):
- A layer clipped to the layer under a reference picture clipped to the reference instead, and was left out of
exports and runs. - Colour and gradient fill layers vanished while you painted or moved a layer under them.
- The Scope counted visible reference pictures into the histogram and the clipping warning.
- Choosing a filter option from the preview list left the layer's own menu on the old option.
- Dragging a corner of a polygon control point could not be undone; an unfinished polygon stayed on screen in
other tools and moved into another tab. - An AI label and the sample plugin's Selection to new layer added their layer without an undo step, so Ctrl+Z
undid the edit before it. set_textfrom an external agent recorded no undo step.- Extend canvas moved a full-size paint layer that had been moved away from the corner, kept the clip of a hidden
layer whose base went into the picture, and wrote untouched layers again. - An undo of a crop or turn re-encoded every filter layer's empty pixels.
- The undo memory budget could count below zero and then let the history grow past it.
- A contact shadow of a clipped layer went under the layer it clips to, where it was hidden.
- Undoing a reference change rewrote the box descriptions inside restored snapshots.
- Relight of the whole picture could send a layer an undo had just taken away, or the old relight result through a
cached filter. - Surface maps made while you painted were fitted to the new strokes under the old picture's fingerprint; they are
dropped now, with a note. - A wrap read from a file took the geometry of the first turn or crop after it, and erased the wrong part.
- Florence-2 read a long strip at full size (a 46 MB canvas, or a failure above 65,535 px) and squashed it; it is
fitted now. A local Generate freed its memory only in the tab where it ran. - With Sample: layer, the eyedropper, the bucket and the wand on a Deep paint layer read the whole picture instead
of the layer's own paint. - Opening a
.scumbledid not tell the plugin panels and Each area once the document was all in. - Undoing a turn of the picture or a named snapshot could turn the depth and surface maps a second time.
- On the canvas backend, a text being typed when a result landed came back at the wrong size on undo.
- The assistant's switch of recipe or tab took a turn snapshot of the tab in front, which Undo this turn then
took back too. - A box description cut at 400 characters could cut a reference marker in two and send it as text.
- Opening a PSD or ORA made its bottom layer the background even when it was in a group or had a layer clipped to
it; the clipped layer then spread over the whole picture. - The range bar's Pick and Invert left the red preview on the canvas and the dialog's own Invert box out of step;
a cancelled Pick on a filter's limit could put a stale undo step on the next slider drag. - A wrap's live link was cut by an undo of a crop or turn, and did not follow a recolour, a flip or a stroke on its
source. - Relight of a whole strip steeper than 8:1 stretched the picture it sent; it is refused with a note now.
- A click with a selection tool and nothing selected put an empty step on the undo list (and dropped the redo
steps); a click that deselects is one step, not two.
- A layer clipped to the layer under a reference picture clipped to the reference instead, and was left out of
- Fixed: the assistant stopped longer tasks in the middle with "the step limit of this turn was reached". A turn
ended after 25 tool calls, and every screenshot and every read of the layers counted. The limit is now 100 by
default (Settings › Assistant, up to 1,000), the note has a Continue button that goes on with the same task,
and the model is told that its turn was cut short, so it carries on instead of starting over. A limit you set
yourself stays as it is.
Scumble 0.1.58: paint that stays on its surface, and helper jobs that end when ComfyUI is gone
Get Scumble
| Package | Download | Size | Notes |
|---|---|---|---|
| Microsoft Store (Windows) | Scumble in the Store or winget install scumble |
– | Signed by Microsoft and updated by the Store; a new version can reach the Store a few days after GitHub (Microsoft's certification). |
| Windows installer | Scumble-Setup-0.1.58.exe | 124 MB | Updates itself; not code-signed yet, so SmartScreen asks once. |
| Windows portable | Scumble-0.1.58-portable-win-x64.zip | 175 MB | No install: unpack and start Scumble.exe; keeps its data in a folder beside it and does not update itself. |
| Linux AppImage | scumble-0.1.58.AppImage | 152 MB | Early build made by CI, not tested by the author yet. |
| Linux .deb | scumble-0.1.58.deb | 120 MB | Early build made by CI, not tested by the author yet. |
| Claude Desktop extension | scumble-0.1.58.mcpb | 83 KB | Adds Scumble's MCP tools to Claude Desktop with one click; needs Scumble installed (any of the above). |
The other files of this release (latest.yml, latest-linux.yml and the .blockmap) are for the app's updater; you do not need to download them.
- Deep paint: Stay on surface. A stroke of the brush or the eraser and a bucket fill on a Deep paint layer now
land only on the surface they begin on: paint a wall and the arm in front of it stays free, paint a table and the cup standing on it
stays free. The switch is in the options bar while a Deep paint layer is active and is on by default; off, they
paint every surface under them, as before. The surface is read from the surface maps (it ends where the depth
jumps and where something rests on it), so its border is as fine as the maps are: a few pixels on a 2,000 pixel
picture, coarser on a very large one. A part of the same surface that is cut off entirely by what stands in front
is not reached; begin a second stroke there. - Fixed: a selection by text stayed on "Segmenting ..." for good when ComfyUI was restarted while it ran (issue
#15): the tab refused a new selection and Generate new until Scumble was restarted, and Cancel knew nothing of the
job. A helper job on your ComfyUI (a selection by text, a background removal, the object map, prompt upsampling
through a node) now ends its wait when the server no longer holds it (a restart, a cleared queue), when it is
interrupted there or when the connection is lost, and the status line says which; a result the connection missed
is read from the server's history. A helper job that runs longer than a moment shows in the title row with its
seconds and Cancel, which takes the job off your ComfyUI and frees the document at once;cancel_runstops it
too, and a mask that still arrives after a Cancel is not applied. - Fixed: GPT Image on Comfy Cloud failed about every second run. The editor's random seed can be larger than the
Partner Node takes, and Comfy Cloud refused the workflow after the pictures were uploaded. The seed now goes inside
the node's range (GPT Image 2, 2.5 and 2.6). - Fixed: Seedream 5.0 lite and pro on Comfy Cloud refused near-square and small crops. At Highres fix › Maximum a
square crop went larger than the model takes (lite 16.8 MP, pro 4.6 MP), and a small crop on lite went under its
3.7 MP floor. The crop is now planned inside the model's own range; crops that fit before go as before. - FLUX.1 Fill is no longer offered for Generate new on Black Forest Labs, fal, Replicate and WaveSpeedAI: it is an
inpaint model whose endpoints require the picture and the mask, so those runs could only fail. - Qwen Image 3.0 on Oxen.ai sends Thinking off beside the prompt expansion that is off (Oxen's schema ties the two).
- Fixed: prompt rewriting with a local reasoning model could put the model's thoughts into the prompt when the
answer was cut off inside its<think>block; such an answer is now an error that names the reason. A GPT-4 model
you added under Settings › Language models for OpenAI is no longer refused (it was sent a reasoning setting it
does not take). - Boxes: an
@imgtoken in a box's description or text that names no reference layer now stops the run before
anything is sent, as it does in the prompt (it went to the model as written, in a paid run). - Workflow import: a setting wired through a Reroute or a bypassed node, or into several inputs, becomes a Settings
row for each input. Before, such an input kept a link to the canvas node that a run never filled, and the recipe
failed on ComfyUI. - ComfyUI window: Save as new recipe from another workflow tab starts from that graph alone. It used to take the
open recipe's task, picture rules, model list, LoRA rows and row values along, which could leave the new recipe
unusable. - Assistant: deleting a chat stops the answer that is still running in it, and a deleted chat no longer comes
back in the list when that answer ends. Asking the assistant to describe the picture no longer shows
"Undo this turn" (nothing changed) and no longer takes the previous turn's undo away; a layer's role set by the
assistant is one undo step, not two. - Assistant: clipping a layer you made to the layer below, or releasing it, now asks first, as changing its
opacity or blend mode does. On a layer the assistant made itself it goes on without asking.
Scumble 0.1.57: outpainting in one step, layers that name themselves, and one Ctrl+Z per result
Get Scumble
| Package | Download | Size | Notes |
|---|---|---|---|
| Microsoft Store (Windows) | Scumble in the Store or winget install scumble |
– | Signed by Microsoft and updated by the Store; a new version can reach the Store a few days after GitHub (Microsoft's certification). |
| Windows installer | Scumble-Setup-0.1.57.exe | 124 MB | Updates itself; not code-signed yet, so SmartScreen asks once. |
| Windows portable | Scumble-0.1.57-portable-win-x64.zip | 175 MB | No install: unpack and start Scumble.exe; keeps its data in a folder beside it and does not update itself. |
| Linux AppImage | scumble-0.1.57.AppImage | 152 MB | Early build made by CI, not tested by the author yet. |
| Linux .deb | scumble-0.1.57.deb | 120 MB | Early build made by CI, not tested by the author yet. |
| Claude Desktop extension | scumble-0.1.57.mcpb | 83 KB | Adds Scumble's MCP tools to Claude Desktop with one click; needs Scumble installed (any of the above). |
The other files of this release (latest.yml, latest-linux.yml and the .blockmap) are for the app's updater; you do not need to download them.
-
Florence-2 as a helper model (Microsoft, MIT): download it in Settings › Helpers (277 MB, every file checked by
its sha256); it runs in the helpers' process, on the graphics card where it can. New results, pastes, cut-outs and
reference images get a short name from what it sees ("Woman sitting on the hood of a blue car" instead of "Result 3")
a moment after they land: only a name that says nothing (Result, Paste, IMG_1234, image) is replaced, never one you
typed meanwhile, and the name is no undo step (Name new layers automatically turns it off). Image › Describe
Selection (and in Detail) puts a caption of the selection or the picture in a note to copy or use as the prompt;
Shift+click puts it into the prompt field. MCPdescribe, 124 tools. -
A Generate result is one undo step of its own: Ctrl+Z takes exactly that result back, also when it lands after
you went on working and for each area's result, and Ctrl+Y brings it back. Before, Ctrl+Z took back whatever came
before the result (after a layers step, the result too). Each result takes one of the history's steps (30 by default). -
Split into layers through OpenRouter for Seedream 5.5 (up to 32 layers and a base, with the count and Text
recognition). Written from OpenRouter's endpoint list and Image API guide, not run live: how the layers come back
through OpenRouter is not documented, and an answer with the picture only says so in red. -
Fixed: a Split into layers run was lost when its window reloaded or crashed during the run: Scumble keeps the
base and every layer and opens them as a document of its own (at the base's size), the layers in one group. -
File › Open and
load_imageno longer send the file to the window in one message: the window reads it from a
link only it can use, so opening a large PSD, TIFF or PNG no longer freezes the window (a 15k PSD was 1.8 s), files
over 2 GB open, and a missing file or a folder is reported in plain words. -
Fixed: a save while a layer was still uploading named the layer's old file; it now waits for the upload.
-
Fixed: a filter layer with a limit and a blend mode or opacity whose filter changed nothing (untouched Curves,
for example) was blended with itself outside the limit in exports and runs; the screen was right. -
Relight on the surface maps uses the depth map's normals where MoGe-2 saw no surface (the sky), as Wrap does.
-
Settings › Helpers: a Deep paint normals select picks MoGe-2 ViT-B or ViT-S.
-
A surface or depth map whose upload failed is stored again; a save says so instead of leaving the map out.
-
The Limit row's and Select range's bar keep a colour, tolerance or Invert you just changed when you drag it.
-
Comfy Cloud: LoRA rows kept for a recipe whose route takes no LoRAs give a plain status line, not a red one.
-
An ORA saved over by another program while it opens says " changed on disk while it was opened; open it again".
-
A cancelled save reads as a plain line again, not as an error.
-
A document with a depth map no longer carries the model's raw numbers in its saved state (about 11 MB of text per
autosave; after Depth from layer on a large picture a save could refuse); files saved before shrink on the next save. -
A mask made on a filter layer after a resize, turn, crop or extend is the picture's size again (it was squashed,
offset or the wrong shape, in exports and paid runs too). -
After a crop, resize or extend, a Deep paint layer's paint is stored at once, so a crash right after no longer loses it.
-
Opening a document and opening, dropping or pasting a picture no longer mix layers when one starts while the other
is still loading. -
A document with a Deep paint or LUT file that could not be read saves again (the layer stays, without that file).
-
Undoing Bold, Italic or a text edit right after typing shows the right text on the canvas and in exports.
-
The mask row's Apply and Remove buttons: an applied mask on a text layer stays applied after the next text edit.
-
A result that lands while you type in a layer's field or hold a slider in the layer list no longer takes that edit
back when you undo the result. -
Split into layers on a very large picture (above about 268 MP) no longer loses the paid answer, and at 15k the
layers land without a 600 MB canvas per layer. -
fal.ai, Replicate and WaveSpeedAI: the API key goes only to https status and result links on the provider's own
host; WaveSpeedAI's upload link gets the picture only over https and never to this machine or its network. -
BFL, Magnific and ToAPIs: a paid job is no longer given up when the machine slept past the run's time or the network
was gone for a few seconds; a Cancel during the download says the job had finished and is charged. -
Select range in Add mode with a range nothing matches leaves the selection's box as it was.
-
A 3D object or mannequin edit that would write a locked layer is refused before anything changes.
-
Parameters in the ComfyUI window: a row you untick keeps the recipe's value instead of the node's default.
-
A change made while a document was still settling after an open or a restore keeps the tab marked as changed, and
edits to plugin data alone (Boxes rows, 3D placements) mark it too, so closing it asks to save. -
View › Reload pressed while the window is saving before closing no longer cuts that save short; Settings ›
Rendering's Restart waits for a document save and asks first while a render on your ComfyUI runs. -
Clean up in Settings › Local files keeps the files of an autosave the window could not restore, and stops when it
cannot read it. -
A paid result brought back from an earlier session is deleted from its kept place only after its document is stored.
-
An assistant call that is refused no longer throws away your redo steps or leaves an empty undo step;
add_surface_layeris an undo step of its own. -
Plugins: refreshing a filter layer after changing its mask, or a Deep paint layer after changing its pixels, stores it.
-
Saved assistant chats no longer change long text that only looks like a picture, and a chat whose screenshot file
is gone can go on. -
A crafted PNG or ORA with an oversized palette or transparency chunk is refused instead of exhausting memory.
-
Fixed: Comfy Cloud results opened as a new document instead of landing as a layer. Comfy Cloud took the
job, but it now answers with job links Scumble's Cloud client sent no key to, so the first status read was refused
("Missing or invalid API key"); the run gave up and the finished picture came back later as a document of its own.
Status reads now always go to cloud.comfy.org with the key, and the result lands on the layer stack again. -
Fixed: Seedream on OpenRouter refused about half of all runs with "seed must be between -1 and 2147483647":
the editor's seed runs up to 4,294,967,295, and Seedream 5.5, 5.0 pro and 5.0 lite take half that. Their seed now
stays inside the model's range. -
Outpainting in one step: tick Generate the new border in the Canvas tool, and applying an extend starts
Generate with your selected recipe on the new border. A recipe on an API provider asks first, since it is a paid
run; the result and the extend are one Ctrl+Z each. Agents useextend_canvaswithgenerate: true. -
Scope and clipping warning: a Scope section in the Image tab shows the live histogram of the visible picture
and how much of it is pure black or pure white. Alt+J (View › Clipping Warning) marks pure black blue and pure
white red on the screen while you grade; exports and the document are never touched. -
A frame from a video: File › Open and a drop on the window take .mp4, .m4v, .webm and .mov files and show a
small player with a slider and one-frame steps; the frame you stop on becomes a new picture, a layer or a
reference layer at the video's own size. Videos the built-in player cannot decode (ProRes, some HEVC) are refused
with a plain line. -
Contact shadow (Layer menu and a layer's right-click menu): a cut-out, a keyed subject or a pasted picture gets
a soft shadow of its own shape under it, darkest where it meets the ground and fading away from it. Distance,
softness, opacity and direction show live in a small dialog, a dep...
Scumble 0.1.56: Deep paint, paint that takes the picture's own light
Get Scumble
| Package | Download | Size | Notes |
|---|---|---|---|
| Microsoft Store (Windows) | Scumble in the Store or winget install scumble |
– | Signed by Microsoft and updated by the Store; a new version can reach the Store a few days after GitHub (Microsoft's certification). |
| Windows installer | Scumble-Setup-0.1.56.exe | 124 MB | Updates itself; not code-signed yet, so SmartScreen asks once. |
| Windows portable | Scumble-0.1.56-portable-win-x64.zip | 175 MB | No install: unpack and start Scumble.exe; keeps its data in a folder beside it and does not update itself. |
| Linux AppImage | scumble-0.1.56.AppImage | 152 MB | Early build made by CI, not tested by the author yet. |
| Linux .deb | scumble-0.1.56.deb | 120 MB | Early build made by CI, not tested by the author yet. |
| Claude Desktop extension | scumble-0.1.56.mcpb | 82 KB | Adds Scumble's MCP tools to Claude Desktop with one click; needs Scumble installed (any of the above). |
The other files of this release (latest.yml, latest-linux.yml and the .blockmap) are for the app's updater; you do not need to download them.
- Deep paint: paint on the picture's surfaces. Image › Surface Maps splits the picture once into its light, its
reflections and its surface shapes, on your own computer (two models in Settings › Helpers: Marigold IID Lighting
and MoGe-2). On a Deep paint layer (Layer menu) what you paint lands under the picture's own light and keeps its
highlights; the eyedropper picks the surface's own colour, Fill and the bucket work on it, Keep reflections sets
how much shine stays, and a selection, a mask or the layer's Limit keep it in place. Where nothing is painted the
picture stays exactly as it was. - Light brush: Layer › New Light Brush Layer paints light where the brush goes; Relight layers on a picture with
surface maps light the picture's own colour and keep its reflections (the Surface maps switch). Without Depth
Anything downloaded, a picture with surface maps takes its depth map from them, so the light brush and Relight need
no second download. In deep shadows, where the surface maps' light is too dark to trust, Relight goes over smoothly
to its way without them, so shadows stay clean instead of speckled or coloured. - Wrap onto Surface (Layer menu): a logo, a pattern or a word laid round a curved surface such as a cup or a
bottle, under the picture's light, in perspective; drag the flat picture with the Transform tool and the wrap follows.
A large picture placed small keeps smooth edges.
The layer list's sparkle button holds the Deep paint layer, the light brush and Wrap. - Agents:
surface_maps,add_surface_layer,wrap_layer, andrelight'ssurface(122 tools). - Seedream 5.5 runs on OpenRouter too (one key for many models; 1K, 1.5K and 2K, a transparent background on request,
up to ten pictures). Not run live yet. - Gemini 4 Hydrogen for prompt upsampling, on your Google key and on your OpenRouter key. It comes after Gemini 3.8 Flash in the list, so the model you have picked stays as it is.
Fixes from a review of the whole code (54 defects, two of them able to lose or change your pixels):
- A layer copied, or an undo step taken, while the autosave stored a large layer keeps its pixels. A Ctrl+J, or a reorder, delete or group followed by Ctrl+Z, during that upload left a layer that reopened with the picture from before your last stroke.
- Exports and runs now match the screen for a limited filter layer in a blend mode. A filter layer with a Limit in Multiply, Screen, Overlay and the like darkened or lightened the picture where the Limit is 0 in a plain export, in the picture sent to a run and in frequency separation; a depth Limit without a depth map did so over the whole layer.
- Paid API runs on fal.ai, Replicate and WaveSpeedAI keep waiting through a network drop or a sleep of up to two minutes instead of giving up on a finished, charged picture, and no longer ask the provider to stop it. Cancel during a WaveSpeedAI download says the job had already finished and is charged.
- Generate new keeps its new tab while the render still runs on your ComfyUI, so the picture is not lost; a New that fails in a tab of its own says why in the tab you were on.
- Requests through curl on Windows check certificate revocation the way a browser does, so they work behind company proxies and antivirus HTTPS scanning. Leftover upload files from a crash are cleaned from the temp folder at start.
- Split into layers works on large pictures: a picture larger than 4096 x 4096 in area goes scaled down, and the layers land at your picture's size.
- Settings changed while settings.json was briefly unreadable no longer wipe your stored per-recipe values, LoRA rows, presets or plugin data; a damaged settings file that cannot be moved aside no longer slows the app, and you are told about it. Imported brush tips survive a damaged brushes.json, and saving tips quickly one after another no longer loses one.
- Blur and the effects pack's Glass frost no longer make half-transparent pixels more opaque (a 50 % area stays 50 %; opaque pictures are unchanged).
- Extend canvas keeps hidden layers, moved with the picture, and its noise border is much faster on large pictures (only the new border gets noise).
- Opening a picture into a tab that is still restoring its document now waits instead of mixing the two. A tab you edit right after opening or restoring it keeps its unsaved-changes marker.
- Flatten, Merge into base, a bucket fill, the magic wand, grow, shrink, feather, invert and frequency separation no longer land an answer over a picture or selection that changed while they ran: they say so, and you run them again. Ctrl+Z pressed during a large bucket fill takes back the fill.
- A click that moves nothing no longer adds an undo step and clears redo; undoing a text edit no longer also takes back the style changes and moves made before it. A second touch or pen contact no longer throws away the stroke in progress.
- Apply mask on a text layer turns it into pixels, so a later text edit can no longer bring the hidden parts back.
- A box description's
@imgreference keeps naming its picture when references are moved, hidden or deleted; Copy rows names From pictures by their@imgnumber. - After an undo or redo while it ran, a relight through a provider hides the right relight layer, and a background removal or a pasted mask lands on its layer. Key screen on Auto asks for the screen colour when a layer's fill and its screen would disagree.
- Picking another recipe while a ComfyUI run uploads no longer sends the other recipe's model files into the run. Save as new recipe never writes over a recipe file of yours with a similar name. The ComfyUI status no longer shows an error after an earlier, slower connect fails.
- Krea 3 Edit on Comfy Cloud fits a slightly off-shape answer back onto the selection. A Split into layers run whose window reloaded no longer brings back only its base picture.
- Importing a damaged or crafted .abr brush pack no longer freezes the window. A second question opened while a dialog is open no longer leaves the first one stuck over the editor. Closing a busy tab twice no longer puts it on Reopen Closed Tab twice. During a compare with only one result chosen, the middle of the view takes clicks again.
- Selection by text, Selection → box and the sample colour probe use far less memory on large pictures; an Object intersect keeps the selection's true bounds.
- 3D object: placing or editing an object is one undo step, and its preview no longer piles up GPU memory while you drag.
- Assistant: the cost line counts cached tokens once and includes Gemini's thinking tokens, and "last turn" is that turn's cost; Undo this turn shows only under the turn it takes back; a chat deleted while its turn runs stays deleted; Stop during a question is no longer recorded as a decline; long chats keep less in memory and on disk.
- A security update of a library the MCP server ships with (proxy-addr 2.0.8).
Scumble 0.1.55: new work in a tab of its own, recipes that remember, and faster ORA files
Get Scumble
| Package | Download | Size | Notes |
|---|---|---|---|
| Microsoft Store (Windows) | Scumble in the Store or winget install scumble |
– | Signed by Microsoft and updated by the Store; a new version can reach the Store a few days after GitHub (Microsoft's certification). |
| Windows installer | Scumble-Setup-0.1.55.exe | 124 MB | Updates itself; not code-signed yet, so SmartScreen asks once. |
| Windows portable | Scumble-0.1.55-portable-win-x64.zip | 175 MB | No install: unpack and start Scumble.exe; keeps its data in a folder beside it and does not update itself. |
| Linux AppImage | scumble-0.1.55.AppImage | 152 MB | Early build made by CI, not tested by the author yet. |
| Linux .deb | scumble-0.1.55.deb | 120 MB | Early build made by CI, not tested by the author yet. |
| Claude Desktop extension | scumble-0.1.55.mcpb | 81 KB | Adds Scumble's MCP tools to Claude Desktop with one click; needs Scumble installed (any of the above). |
The other files of this release (latest.yml, latest-linux.yml and the .blockmap) are for the app's updater; you do not need to download them.
- Generate new and New open a tab of their own when the current one holds work (issue #12): the document stays as it is, and the new tab takes its reference layers, prompt, settings and boxes along. A tab with nothing in it yet still takes the picture in place, and a New you cancel closes its tab again. For agents,
generate_newandnew_canvaswork on the tab they name, as before. - The Settings you pick for a recipe stay with it (issue #12): a new tab, and the recipe picked again after another one, start with the values you last chose in its Settings rows instead of the recipe's own. A preset you pick counts too; Recipe's own values (in the Preset row, for an API model a button of its own) puts the recipe's back and forgets what you set.
- Large ORA files open about three times as fast, and the window stays responsive while their layers are read (a 15,000 x 10,000 ORA in about 1.8 s instead of 5.4). Their layers decode in the background workers straight from the file.
- Requests now go through your system's curl (Settings › Network, on by default): API providers, Comfy Cloud and Comfy Router, prompt upsampling, model downloads and a remote ComfyUI. Company networks often let curl through where they block other programs. Your keys go to curl on its input, never on its command line; switched off, or with no curl found, Scumble sends its requests itself as before. A ComfyUI on this computer is always reached directly.
- Seedream 5.5 now also runs on WaveSpeedAI, Comfy Cloud, ToAPIs, Comfy Router and Oxen.ai, and Split into layers runs on the provider you picked for it, so WaveSpeedAI and Oxen.ai can split too. A layer that comes back without a position keeps its own shape in the middle of the picture instead of being stretched, and the status line says so. Written from the providers' docs: none of these routes has run against the live API yet.
- A new skin, Lighthouse Academy (Settings › Appearance): near-black panels, pale yellow type and a warm orange, with "Supported by" and Lighthouse Academy's logo in the tab row.
- Zoomed into a large picture over the tile memory budget, the view stays on the GPU instead of falling back to the slower drawing path on every frame. A tile memory setting of 2048 MB or more now takes effect (it used to drop to 16 MB).
- Adding to, subtracting from or intersecting a large selection is faster where the change leaves a part of it as it was, and that part is no longer copied into the undo history a second time (a 15,000 x 10,000 picture: adding nothing to half of it 171 -> 67 ms).
- Haze by depth without the GPU now matches the screen and the export. Its depth edge used to sit more than a map pixel off at a zoomed-out view. A depth map is no longer uploaded to the GPU again on every pass of a haze, lens blur, limit or relight filter.
- A
.cubeLUT whose DOMAIN is not 0..1 now reads that range as its input range. Before, it scaled the table's colours instead, so such a LUT came out too dark or too bright. A LUT with the usual 0..1 range gives the same colours as before. - A failed stitch no longer uses a newer selection. If placing a provider's answer had to fall back after a stitch worker error, the answer was cut with whatever was selected at that moment; it now uses the selection the run was sent with, or says the stitch failed.
- Split into layers on BytePlus ModelArk now gets the number of layers you ask for: the request counted the base as one of them, so asking for one layer was refused.
- The assistant's Split into layers takes
source: "picture"as the visible picture again, even when a layer's name contains the word "picture". - A PNG exported from a local MiniMax H3 Image run now carries a workflow that loads back into ComfyUI: the filter that keeps API keys out of a PNG took the upsampler's
keyinput, the name of a JSON field, for a key.
Scumble 0.1.54: large files open in a fraction of the time, and MiniMax-H3-Image
Get Scumble
| Package | Download | Size | Notes |
|---|---|---|---|
| Microsoft Store (Windows) | Scumble in the Store or winget install scumble |
– | Signed by Microsoft and updated by the Store; a new version can reach the Store a few days after GitHub (Microsoft's certification). |
| Windows installer | Scumble-Setup-0.1.54.exe | 124 MB | Updates itself; not code-signed yet, so SmartScreen asks once. |
| Windows portable | Scumble-0.1.54-portable-win-x64.zip | 175 MB | No install: unpack and start Scumble.exe; keeps its data in a folder beside it and does not update itself. |
| Linux AppImage | scumble-0.1.54.AppImage | 152 MB | Early build made by CI, not tested by the author yet. |
| Linux .deb | scumble-0.1.54.deb | 120 MB | Early build made by CI, not tested by the author yet. |
| Claude Desktop extension | scumble-0.1.54.mcpb | 81 KB | Adds Scumble's MCP tools to Claude Desktop with one click; needs Scumble installed (any of the above). |
The other files of this release (latest.yml, latest-linux.yml and the .blockmap) are for the app's updater; you do not need to download them.
- Generate new asks before it replaces your work. When the tab holds layers, a picture or edits, it names what goes, since the answer replaces them and the undo history, and you can cancel and save first.
- MiniMax-H3-Image on MiniMax's own API, a new provider (Size 1K / 2K / 4K, a Highspeed choice, International or China region). Written from MiniMax's docs, not run against the live API yet.
- MiniMax-H3-Image on your own ComfyUI (0.45 or newer): edits with up to 15 more pictures, Generate new and Comfy's prompt expander. Written from Comfy's templates and an open pull request, not run yet.
- Large TIFF files open faster, and small palette TIFFs with transparency open correctly. Multi-strip and tiled TIFFs decode across worker threads: a 15,000 x 10,000 LZW TIFF opens in about 3.5 s instead of 13.7, a PackBits one in 3.3 s instead of 7.4. Opening another file stops the one still being read at once. Palette TIFFs of 1, 2 or 4 bits with an alpha channel no longer open nearly transparent.
- Large PSD files open about twice as fast, and the window stays responsive while their layers land.
- The assistant's render questions show the settings. Before a render or upscale its question lists the document's settings and marks what the chat changed, and it can set crop options, use the magic wand and add boxes without failing on the active layer.
- The assistant asks before merging into your own layer. Merging down from the assistant now asks when the target is a layer you made or the base image, and unanswered question cards are switched off when a turn stops, so an old card cannot answer a later question.
- Split into layers (Image > Split into Layers..., MCP
split_layers): Seedream's layer decomposition on BytePlus ModelArk (5.5 and 5.0 Pro) lands a picture's elements as named layers in one new group above your layers, as one undo step. Not run against the live API yet. - A paid run is not lost when the window reloads or crashes. Its picture is kept and opens in a tab of its own, and View > Reload asks first while a run is going. Save to recipe in the ComfyUI window refuses when the page shows another graph. A failed export over an existing file keeps the finished temporary file and names it. Long commands over MCP get their full time, and parallel connections no longer fail.
- Fibo 2.0 Edit Turbo on your own ComfyUI: edits by instruction with up to three more pictures, with Comfy's Prompt Enhancer in front (a Settings row). Not run live yet. Bria's weights are CC BY-NC 4.0, non-commercial; commercial use needs an agreement with Bria.
- GPT Image 2.6 Flare and Sunburst are also on Oxen.ai, with the Complexity row. Oxen renamed some fields, so GPT Image 2 and 2.5 on Oxen.ai send the selection under its new name and Qwen its pictures. Written from Oxen's model list, not run against the live API yet.
- Relight's shadows land about 1.5 times as fast after a light moves (0.5 s instead of 0.75 at 15,000 x 10,000). A sun's shadows now come from the same exact calculation as the other lights, so a few pixels on a shadow's edge can differ by one visibility step from earlier versions.
- Saving over a read-only
.scumblefile stops at once with a clear message. Opening two documents at the same time no longer trips over shared files, or removes a file that still exists from Open Recent. Temporary save files on a locked or unplugged drive are cleaned up on a later start. A headless MCP session now finishes saving even when that takes longer than five seconds. - Layered PSD exports, dissolve layers and the magic wand over the whole picture are faster on large pictures, with the same pixels. A 15,000 x 10,000 PSD exports in about 1.5 s instead of 2.0, a dissolve layer's export in 1.4 s instead of 1.9, and a whole-picture wand takes about 1.2 s instead of 1.4.
- Scumble's own PNGs, also the layers inside
.scumbledocuments, open faster: their parts decode in parallel across the worker pool, and an image starts decoding while it is still uploading instead of after it. - Grow and Shrink Selection are faster when they have to run on the window instead of the background workers (about 1.3 to 1.8 times), and large temporary memory is given back right after.
- Colour match without WebGL2 is faster. The fallback that runs on the CPU, without WebGL2 or above the GPU's largest texture, goes through the Rust module in 4 MP pieces, about 1.7 times as fast at 15,000 x 10,000, and now rounds as the GPU path does.
- The bucket fill on large pictures no longer stalls the window for its readbacks. On tile layers the flooded area is composited straight into the tiles; at full opacity the pixels are the same as before, at lower opacity within one level.
- Faster PNG and ORA exports on large pictures, with the same files. A 15,000 x 10,000 PNG takes about 0.87 s instead of 1.01, an ORA 2.0 s instead of 2.26: the pieces are now gathered, composited and packed directly in the Rust module's memory.
- A depth map is ready faster, with the same bytes. The guided filter behind it and its preparation run off the window in Rust, for a depth map from Depth Anything V2 Small and from a layer.
Scumble 0.1.53: each area on its own, Krea 3 Edit, and SAM2 off the main thread
Get Scumble
| Package | Download | Size | Notes |
|---|---|---|---|
| Microsoft Store (Windows) | Scumble in the Store or winget install scumble |
– | Signed by Microsoft and updated by the Store; a new version can reach the Store a few days after GitHub (Microsoft's certification). |
| Windows installer | Scumble-Setup-0.1.53.exe | 124 MB | Updates itself; not code-signed yet, so SmartScreen asks once. |
| Windows portable | Scumble-0.1.53-portable-win-x64.zip | 175 MB | No install: unpack and start Scumble.exe; keeps its data in a folder beside it and does not update itself. |
| Linux AppImage | scumble-0.1.53.AppImage | 152 MB | Early build made by CI, not tested by the author yet. |
| Linux .deb | scumble-0.1.53.deb | 120 MB | Early build made by CI, not tested by the author yet. |
| Claude Desktop extension | scumble-0.1.53.mcpb | 81 KB | Adds Scumble's MCP tools to Claude Desktop with one click; needs Scumble installed (any of the above). |
The other files of this release (latest.yml, latest-linux.yml and the .blockmap) are for the app's updater; you do not need to download them.
- Krea 3 Edit, on fal.ai, WaveSpeedAI, Replicate, Comfy Cloud, OpenRouter, Comfy Router and Oxen.ai, with up to eight pictures, a Planning setting and Generate new (not on OpenRouter). It repaints the whole crop and your selection limits what lands. Built from the providers' docs, not run against the live APIs yet.
- Cancel on fal.ai and Replicate now asks the provider to stop the job. Before, Scumble only stopped waiting and the job ran on and was charged; a run that times out asks the same. If the provider does not answer within two seconds the line says the job may still finish.
- Importing a recipe no longer overwrites another one without saying so. A workflow or a recipe without an id of its own gets a free name, and a recipe that does carry an id says whether it replaced one of yours or shadows a shipped one, in the status line when you import from the File menu too.
- Long API errors stay readable. An error answer no longer carries a whole picture back into the status line and the log, and a key inside one is removed before the text is cut.
- A fill layer with a depth limit exports almost twice as fast, and with edge snap its soft edge now follows the picture as it does on screen. On a 15,000 x 10,000 picture the fill layer takes about 6.2 s instead of 11.2, a PSD holding one 9.9 s instead of 16.7. As long as the picture below is still the one the depth map was made from, export and screen agree exactly up to 4096 px on the long side; on larger pictures a hard colour edge can differ in a band a few pixels wide.
- A crash or a locked file can no longer wipe your autosave, settings or API keys. The three stores are now flushed to disk before they replace the old file, and a read that fails no longer counts as an empty store. A damaged
settings.jsonorsecrets.jsonis kept under a name of its own (settings.json.damaged, then.damaged.1and so on), the status line and the log say so once, and new settings and keys save normally again. Closing the window while another program holdssettings.jsonno longer skips the save. - Each area of a selection on its own (issue #9): a selection of several separate areas (three people selected by text, say) goes to the model as one crop around all of them, scaled down to the model's size, so each area came back soft. Before such a run the status line now says so when that crop goes out at less than half its size, and Each area, beside Generate, is shown while the selection holds two or more separate areas: it generates each area as its own crop and run with the same prompt and recipe, one after another, each result its own layer, and leaves your selection as it is (a turn, a crop or resize, Generate new and a second Generate on that tab wait until the areas are done). Up to 8 areas; areas whose crops would overlap count as one, and specks (smaller than a 20,000th of the picture, and anything under 256 pixels) outside the areas are left out. On an API model each area is a run of its own and is charged as one, so Scumble asks first and names the runs (the in-app models are free and do not ask). On your own ComfyUI the next area is queued only once the one before has landed. While the areas run the button reads Stop: it stops the run going (an API run as Cancel does; one on your ComfyUI finishes there) and the areas not yet sent. For agents:
generatetakeseach_area(its timeout bounds the whole run),statuscounts the selection's areas, and the in-app assistant's question names the runs. The areas are counted in the background on a coarse copy of the selection: on a 15,000 x 10,000 picture the window spends under a millisecond on it. - The blur and sharpen brushes and frequency separation take about half the time, with the same pixels as before. On a 15,000 x 10,000 picture a 400 px blur stroke takes about 33 ms per move instead of 74, a 1000 px one about 99 instead of 231, the sharpen brush about half as long as well, and frequency separation of a 16-megapixel area about 0.7 s instead of 2 s, while the window stops answering for half as long. The soft edge of a generated result's mask and relight's surface map are worked out a little faster too.
- The pixel kernels really give memory back now. In 0.1.52 the first kernel instance of the window and of each background worker stayed in memory after it was replaced, so the memory 0.1.52 said its workers returned after a large operation (about 290 MB after a 15,000 x 10,000 export) stayed in use until the app was closed; now it is freed.
- GPT Image 2.6 Sunburst and Flare on Comfy Cloud too, through OpenAI's partner node, with a Complexity setting and the selection as the mask (built from the node's own description, not run on Comfy Cloud yet).
- GPT Image 2 and 2.5 on Comfy Cloud: a run with a selection and reference layers or the Original no longer fails. It goes without the mask, which that node takes beside one picture only, and the selection still limits the result; the status line says so as a plain note.
- A filter layer's sizes follow a resize. Resize image, an upscale of the whole picture and the Realism Pass above 1x now scale a filter layer's sizes in pixels (a blur's radius, a Grain layer's size and plate scale, the glass and chromatic shift, halation, glow, tonal contrast and structure radii) by the same factor, kept within each slider's range, in the same undo step, so the effect keeps its strength on the new size. The grain a Film look or a black-and-white film adds of its own keeps its size in pixels, as do layers of a plugin that is off (the status line says how many).
- Seedream 5.5 on BytePlus ModelArk (early access: BytePlus has to let your account in, then you activate it in the ModelArk console under Model activation). It edits with up to ten pictures and makes new pictures with Generate new, at 0.92 to 4.6 megapixels, as PNG. Its Background row set to Transparent asks for a cut-out, and the result layer keeps that transparency inside the selection. Written from BytePlus's preview docs, not run live yet.
- The dodge / burn and sponge brushes are faster, with the same pixels. On a 15,000 x 10,000 picture a 1000 px sponge stroke moves in 67 ms instead of 87 (a 400 px one in 22 instead of 31), and a 1000 px dodge in 64 instead of 69.
- Liquify lands faster, with the same pixels. The bake of a stroke runs in Rust: on a 15,000 x 10,000 picture the pool's work for a 1000 px push is about 1.8 times faster, and a 200 px push at 1:1 lands in about 20 ms instead of 25. The live preview while you drag runs in Rust too: a 1000 px push at 1:1 draws a frame in 10 ms instead of 16, and the brush itself works about twice as fast (a 1000 px push on a fitted 15,000 x 10,000 picture: about 5 ms a move instead of about 14).
- A local run can no longer pick up another picture of the same name on your ComfyUI. Pictures you open, paste or drop now go up under a name made from their content, such as
image~3fa9c1d2.png, so two different pictures namedimage.png(from a second copy of Scumble, the Inpaint Canvas node or an older session) no longer get mixed up. Layer and tab names stay as they were. A document saved earlier still works: if the server holds a different file under its old name, the run uses its own copy. - A filter layer no longer thickens soft edges. A cut-out or other see-through picture under a filter layer looked more solid at its edges on screen and in some exports than in the plain PNG. They now agree. Opaque pictures look the same, apart from a very small shift (a level or two) inside the soft ramp of a filter layer's Limit.
- A paid run on fal, Replicate or WaveSpeed no longer ends on a short connection problem. Scumble now rides out a few failed status reads and retries a failed download of the finished picture, on Comfy Router and the Partner API too. If it does give up, the message names the job's request id, so you can find the picture at the provider.
- Invert selection is about twice as fast. On a 15,000 x 10,000 picture it takes 0.13 s instead of 0.23, with the same selection. The magic wand over a whole 15,000 x 10,000 picture lands in 0.83 s instead of 1.0, and shrinking a selection by 16 px in 0.35 s instead of 0.48.
- A large ORA exports faster. A 15,000 x ...
Scumble 0.1.52: green screen keying on your ComfyUI, and LoRAs
Get Scumble
| Package | Download | Size | Notes |
|---|---|---|---|
| Microsoft Store (Windows) | Scumble in the Store or winget install scumble |
– | Signed by Microsoft and updated by the Store; a new version can reach the Store a few days after GitHub (Microsoft's certification). |
| Windows installer | Scumble-Setup-0.1.52.exe | 124 MB | Updates itself; not code-signed yet, so SmartScreen asks once. |
| Windows portable | Scumble-0.1.52-portable-win-x64.zip | 174 MB | No install: unpack and start Scumble.exe; keeps its data in a folder beside it and does not update itself. |
| Linux AppImage | scumble-0.1.52.AppImage | 152 MB | Early build made by CI, not tested by the author yet. |
| Linux .deb | scumble-0.1.52.deb | 120 MB | Early build made by CI, not tested by the author yet. |
| Claude Desktop extension | scumble-0.1.52.mcpb | 80 KB | Adds Scumble's MCP tools to Claude Desktop with one click; needs Scumble installed (any of the above). |
The other files of this release (latest.yml, latest-linux.yml and the .blockmap) are for the app's updater; you do not need to download them.
- LoRAs on your own ComfyUI and on Comfy Cloud (issue #10): the Generate tab has a LoRAs section right after Settings, shown while the recipe runs on your ComfyUI or on Comfy Cloud and its graph has one model loader (FLUX 3 Image Klein, Flux.2 Klein, Qwen Image Edit 2.1 and its Turbo, MOSS-Image, MOSS-Image Edit, P-H3-Image, a workflow of your own, and every Comfy Cloud recipe but Ideogram 4, whose graph has two); LoRAs never go to the API providers. It holds up to five rows per recipe, each a LoRA picked from your ComfyUI's list (grouped by subfolder, with a search; Enter picks, Escape closes), a strength (−2 .. 2 on the arrows, a typed value up to the node's own range) and an on / off tick; + adds a row, × removes one, and a file the server does not have is marked "(not on the server)". The rows are kept per recipe for every document, and a recipe saved under a new name from the ComfyUI window takes the rows of the one it came from. A run of a ComfyUI recipe loads the LoRAs stored for that recipe, in their order and at their strengths, between the recipe's model loader and whatever reads the model, in the copy of the graph that is queued (the recipe file never changes). The status line and the result's line in the history name them ("FLUX 3 Image Klein (ComfyUI) via your ComfyUI + 2 LoRAs: Flux3_Targa 0.8, nyc_style 1.0, 14 s"), an exported PNG's recipe chunk lists them, and a LoRA file your ComfyUI does not have refuses the run before anything is uploaded. MOSS-Image Edit and P-H3-Image keep their own LoRA as a Settings row; your rows go after it. Under each picked LoRA a grey line says which base model it looks made for: "looks like FLUX 3" when Scumble can read the file's tensor names (your ComfyUI runs on this machine and Settings › Helpers points at its models folder), else "metadata says SDXL" from the file's own metadata (which can be wrong) or the folder it sits in. The line turns amber when the file looks made for another model than the recipe runs ("This LoRA looks made for FLUX.1; this recipe runs FLUX.2 Klein ..."), since ComfyUI loads such a LoRA without an error and the picture may come back unchanged; nothing is refused on it. Trigger words the file carries show as chips, and a click on one appends it to the prompt. The picker lists the folders of the recipe's model first and the rest behind Show all. Under a Comfy Cloud recipe the list is Comfy Cloud's: the LoRAs it lists as your own (imports, Creator plan or higher) first under Your imports, then Comfy Cloud's catalog, and a file name typed into the search goes as written; a name Comfy Cloud's lists do not show is marked, never refused, and Generate and Generate new both load the rows there. LoRA rows ran on Comfy Cloud with one of its catalog LoRAs on Qwen Image Edit 2509; an import under Your imports has not been seen yet (the account tried had none). Agents and the assistant set the rows too (
set_generationloras), read what a run would load (statusloras,list_recipesloras), list your ComfyUI's LoRA files with their hints (list_loras, new, with a search and a filter by base model) and get the LoRAs a run loaded back (generateandgenerate_newloras_sent). - Key a green or blue screen on your own ComfyUI (Image › Key Green / Blue Screen...): CorridorKey 2.0 by Corridor Digital keys it through the node pack ComfyUI-CorridorKey2.0 on your ComfyUI (an NVIDIA card of the RTX 30 generation or newer, data-centre cards such as an A100 or L40S included; the pack is installed there and its models, about 5.1 GB, fetched there once, never by Scumble, and a key without them says which files are missing and how to fetch them; the manual's new chapter Key green / blue screen has the steps). It adds one layer Key with the subject's own colour, the screen's spill taken out, and the matte (soft hair, motion blur) as its layer mask, trimmed to the matte; a keyed layer is hidden in the same step, a key of the picture gets a mid-grey layer Grey under it, and one Ctrl+Z takes it all back. The selection is the hint (trusted as Rough, Automatic or Clean), or none; Screen Auto reads the box's border; Source is the visible picture or the active layer. The box is the selection grown by some context, or the whole picture, and goes at its own size up to 16 megapixels; a larger box is sent scaled down and its matte refined against the full-resolution picture. With a key layer active the dialog offers Re-key from its painted mask. The dialog moves by its title row and closes with its ×. The first key after another job on your ComfyUI takes about 10 s more while the pack loads its models, and that job's models load again at its next run; Keep the model on the GPU between keys saves the load for keys in a row. CorridorKey 2.0 is under the Corridor Key Licence 2.0 (CC BY-NC-SA 4.0 with their terms), used with Corridor Digital's permission in this free app, and builds on C-RADIOv4 (NVIDIA), MoGe-2 (Microsoft) and RVM (Google DeepMind); your keyed pictures are yours, also commercially, and organisations with more than US $20M yearly revenue need one CorridorKey seat per user (corridorkey.com). Agents:
key_screenandstatuskey; agents now have 118 tools. - MOSS-Image Edit (issue #11) as a new recipe on your own ComfyUI (0.43 or newer runs it on its own nodes): it edits the selection by instruction with up to three more pictures (the Original and your reference layers), named in the prompt as image 2, image 3 and image 4, with MOSS-Image and the edit LoRA of Comfy's own template (297 MB, from Comfy-Org's MOSS-Image repository on Hugging Face; a ComfyUI without it stops the run with its download link). On an older ComfyUI the recipe is listed and says which node is missing. Run on a local ComfyUI 0.44; name what each picture is in the prompt ("Image 1 is the scene to edit.
@img1is a picture of the bag to place."), since a bare "the bag from@img1" makes it redraw the scene. The same template also runs on Comfy Cloud as MOSS-Image Edit (Comfy Cloud): an edit of the selection with one picture on your Comfy Cloud key, and Generate new with MOSS-Image's text-to-image template and its prompt enhancer; the edit ran on Comfy Cloud in about half a minute, Generate new there has not run yet. The edit LoRA is chinoll's Moss-Image-Edit, rehosted by Comfy-Org, whose repository card says Apache-2.0; chinoll's own repository names no licence for the weights (its code is GPL-3.0). MOSS-Image's weights are Apache-2.0. - P-H3-Image, Pruna AI's 8-step image model on MiniMax H3: on your own ComfyUI (0.44 or newer runs it on its own nodes, with MiniMax H3's FL2VA model and Pruna AI's P-H3-Image LoRA, text encoder and VAE), on Comfy Cloud (Comfy's own templates) and on WaveSpeedAI. It edits the selection with more pictures (up to seven on your ComfyUI, eight on WaveSpeedAI; Comfy Cloud's template edits the crop alone), named in the prompt as <Picture 2>, <Picture 3> ..., and Generate new makes a picture from the prompt alone; on your ComfyUI Pruna AI's prompt upsampler writes the prompt out first; a run with an empty prompt is refused, since the model would redraw the crop or the upsampler invent an edit. Run on a local ComfyUI; the Comfy Cloud and WaveSpeedAI routes have not run yet. Powered by MiniMax H3; P-H3-Image is under the MiniMax H3 Community License, which does not cover use in the EU, the UK, South Korea or the US without a separate licence from MiniMax (platform.minimax.io/h3-license).
- GPT Image 2.6 Sunburst and Flare (OpenAI's models of 2026-11-04) as two new recipes, on OpenAI's own API and on fal.ai, Replicate, WaveSpeedAI and OpenRouter, Sunburst on ToAPIs too: the same sizes as 2.5 (up to 3840 px an edge), quality up to max, and a new Complexity row in the Settings panel, how much the model plans before it draws: low is faster for simple edits, high (the default) suits busy scenes and lettering. Written from the providers' documentation, not run against the live services yet. GPT Image 2.5 Flare and Sunburst run on fal.ai and Replicate now too, and on WaveSpeedAI they inpaint with your selection as the mask instead of an instruction edit of the whole crop.
- Opening a large PSD no longer freezes the window. Its layers are read in the backgrou...
Scumble 0.1.51: large exports save again, and two new skins
Get Scumble
| Package | Download | Size | Notes |
|---|---|---|---|
| Microsoft Store (Windows) | Scumble in the Store or winget install scumble |
– | Signed by Microsoft and updated by the Store; a new version can reach the Store a few days after GitHub (Microsoft's certification). |
| Windows installer | Scumble-Setup-0.1.51.exe | 124 MB | Updates itself; not code-signed yet, so SmartScreen asks once. |
| Windows portable | Scumble-0.1.51-portable-win-x64.zip | 174 MB | No install: unpack and start Scumble.exe; keeps its data in a folder beside it and does not update itself. |
| Linux AppImage | scumble-0.1.51.AppImage | 152 MB | Early build made by CI, not tested by the author yet. |
| Linux .deb | scumble-0.1.51.deb | 119 MB | Early build made by CI, not tested by the author yet. |
| Claude Desktop extension | scumble-0.1.51.mcpb | 78 KB | Adds Scumble's MCP tools to Claude Desktop with one click; needs Scumble installed (any of the above). |
The other files of this release (latest.yml, latest-linux.yml and the .blockmap) are for the app's updater; you do not need to download them.
- Large exports save again. Saving an export of more than about 500 MB (a PSD of a 15,000 x 10,000 picture with a layer above it is about 0.9 GB, and a large ORA or TIFF can pass it too) ended Scumble's window: it reloaded and brought the documents back from the autosave, but the file was never written. An export now goes to the disk in pieces, whatever its size.
- Two new skins under Settings › Appearance (and View › Skin): Comfy, a deep aubergine night with neon-yellow pills, the side panel's sections as node cards with a yellow socket, and Black Forest, forest-night greens with warm cream type, sage accents and calm rounded panels.
- An export is written to a temporary file beside the target and put in its place only once every byte is written, as a
.scumbledocument is: a save that fails, or a disk that is too full for it (now said before anything is written), leaves an older file of the same name as it was. When another program holds that file open (a viewer), or the file has other names (a hard link), the export is written into it as before; a symbolic link keeps pointing at the new picture. A read-only target is refused before the export is written.
Scumble 0.1.50: MOSS-Image with boxes and the FLUX 3 Klein upsampler
Get Scumble
| Package | Download | Size | Notes |
|---|---|---|---|
| Microsoft Store (Windows) | Scumble in the Store or winget install scumble |
– | Signed by Microsoft and updated by the Store; a new version can reach the Store a few days after GitHub (Microsoft's certification). |
| Windows installer | Scumble-Setup-0.1.50.exe | 124 MB | Updates itself; not code-signed yet, so SmartScreen asks once. |
| Windows portable | Scumble-0.1.50-portable-win-x64.zip | 174 MB | No install: unpack and start Scumble.exe; keeps its data in a folder beside it and does not update itself. |
| Linux AppImage | scumble-0.1.50.AppImage | 152 MB | Early build made by CI, not tested by the author yet. |
| Linux .deb | scumble-0.1.50.deb | 119 MB | Early build made by CI, not tested by the author yet. |
| Claude Desktop extension | scumble-0.1.50.mcpb | 78 KB | Adds Scumble's MCP tools to Claude Desktop with one click; needs Scumble installed (any of the above). |
The other files of this release (latest.yml, latest-linux.yml and the .blockmap) are for the app's updater; you do not need to download them.
- MOSS-Image on your own ComfyUI. A local recipe for OpenMOSS's open MOSS-Image model (ComfyUI 0.42 or newer runs it on its own nodes; the weights are Apache-2.0): it makes new pictures with Generate new, and with the Boxes switch on your New and Text boxes go as its JSON layout, each box placed in the picture's pixels and a Text box's words rendered there (issue #6). It takes no picture to edit, so Generate on a selection says to use Generate new.
- FLUX 3 Image Klein with Black Forest Labs' prompt upsampler. On your own ComfyUI the recipe now writes your prompt out first, as Black Forest Labs' own pipeline does: their 9B upsampler model looks at the crop and the pictures after it and turns your words into the detailed instruction the model was trained on (a caption for the canvas's shape on Generate new). It runs when its file (qwen3.5_9b_flux3_prompt_upsampler_w4a8, 6.7 GB, in the same Hugging Face repository) is on your ComfyUI; the Prompt upsampler row switches it, and the status line and the log show what it wrote. Without the file, or while boxes go, the prompt goes as typed, with a note.
- The Realism Pass recognises workstation cards: an RTX PRO 6000 Blackwell counts as the RTX 50 generation, an RTX A6000 as RTX 30 and an RTX 6000 Ada as RTX 40, instead of being refused because their names carry no two-digit generation after "RTX" (issue #8); a Quadro RTX 8000, a card too old for it, is no longer let through. The ComfyUI-DLSS5-Enhancer node pack checks the card the same way by name: until it decides by the card's compute capability, it refuses those cards itself, and the status line says so.
- Selection by text works again with the current comfyui-rmbg on your ComfyUI: its SAM3 node now requires a model name, which Scumble did not send, so every selection by text failed (issue #7). Scumble now fills in every input a helper node requires and Scumble does not set itself with that node's own default (for SAM3 the model
sam3.1_multiplex_fp16), for the selection by text, the cutout and the prompt from the picture; an older node pack without the input keeps working. - A local Flux.2 Klein or FLUX 3 Image Klein run that fails in the sampler with "mat1 and mat2 shapes cannot be multiplied" now says why: the status line names the model and text encoder files the run sent, which encoder fits (Klein 9B takes qwen_3_8b, Klein 4B qwen_3_4b, FLUX 3 Image Klein qwen3vl_4b) and the Settings row to change; when the names already fit, it names the other causes (a ComfyUI too old for the model, or a file that holds another model than its name says). The two recipes' descriptions name the fitting encoder too.
- FLUX 3 Image Klein's recipe names the files as Black Forest Labs published them (
flux3-image-klein-int8-convrot.safetensorsand its base, bf16 and nvfp4 siblings), so its download links work again, and says that the Hugging Face repository asks you to log in and accept the licence first. - The Realism Pass says in red when it changed nothing: when the answer is the picture that went (compared pixel for pixel on pieces spread over the picture), the layer still lands and the status line says the pass changed nothing on that card. Below Strength 0.25 nothing is said. Every pass writes its difference to the log.
- An export whose Size or Canvas row would pass what one canvas holds (65,535 px a side, 268 megapixels) is refused before anything is drawn, instead of writing no file or an empty one; a WebP wider or taller than 16,383 px is refused too (it was written cut to that width). A typed export size past 32,768 px a side is scaled down with the aspect kept, and the status line says so (each side used to be clamped on its own, which stretched the picture).
- A document saved by a newer version that holds layers this version cannot show no longer loses them without a word: the status line after the open says in red how many were left out and that the next save drops them. A filter layer that cannot be read is counted too.
- A local run whose prompt names a reference (
@img1) now stops with a plain error (Inpaint Canvas failed: reference 2 (@img2) could not be loaded ...) when the Inpaint Canvas node on your ComfyUI cannot load one of the reference pictures; before, it left that picture out and the later references went under the wrong names. This takes effect once the node on your ComfyUI is updated (git pullin its folder). - The manual names the nodes behind Selection by text's menu: SAM3 is comfyui-rmbg's SAM3 node (weights in
models/sam3; on Windows it also needs the Python packagetriton-windows), GroundingDINO + SAM is comfyui_segment_anything, and SAM3 (core, experimental) is ComfyUI's own SAM3 nodes with a sam3 checkpoint. - Agents and the assistant:
add_textandset_texttakeline_heightandletter_spacing(the text panel's Line and Spacing);reopen_closed_tabbrings back the tab closed last (File › Reopen Closed Tab);relighttakes the model values the Relight panel offers for a generative route (Magnific's strength, style, preserve details and interpolate, the IC-Light route's high-res fix, Bria's light type, LoRA strength and steps on your ComfyUI and Comfy Cloud, where a strength in (0, 1] is read as a fraction, and detail from the original), and values the layer already holds make no undo step;boxes_checksays what the Boxes panel warns of before a paid run (boxes outside the crop, cut by its edge, outside the selection, too small, without a description, or of a kind the recipe does not place);open_documentandreopen_closed_tabcarry the note about left-out layers, andrealism_passanswersunchanged. Agents now have 116 tools.