Skip to content

Scumble 0.1.52: green screen keying on your ComfyUI, and LoRAs

Choose a tag to compare

@github-actions github-actions released this 07 Nov 04:07
· 635 commits to main since this release

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_generation loras), read what a run would load (status loras, list_recipes loras), 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 (generate and generate_new loras_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_screen and status key; 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. @img1 is 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 background and come in while the window stays usable: for a 15,000 x 10,000 PSD with a layer above the background, the window used to stop answering for 9 to 11 seconds and now stops for under 2 (while the file itself is read in), and the open takes about a third less time.
  • Selections on large pictures land sooner and hold the window less. Joining a new region to the selection now runs in the fast pixel kernels: on a 15,000 x 10,000 picture, the magic wand across the picture stops the window for about 0.1 s instead of 0.3 to 0.4 s and finishes about a sixth sooner, select by depth takes about 1.2 s instead of 1.7 s and select by brightness about 1.7 s instead of 2.7 s.
  • PNG files are written faster, byte for byte the same. Choosing how each row of a PNG is packed now takes 16 bytes at a time: on a 15,000 x 10,000 picture a PNG export takes about 1.4 s instead of 1.9 s and an ORA about 3.5 s instead of 5.1 s, and opening a large TIFF is about a second shorter, since the app keeps every layer it stores as a PNG too.
  • Relight's shadows land much sooner after a light moves, the same shadows as before. Each light's shadows are now worked out by all the background workers together instead of one: on a 15,000 x 10,000 picture they land about 0.65 s after one light is moved instead of 3.5 s, about 1 s after three lights are moved instead of 3.7 s, and panning at full size while they are worked out stays smooth (its slowest frame about 10 ms instead of 90).
  • The pixel kernels hold less memory and recover on their own. After an operation, a worker that grew past 16 MB returns to its starting size within a few seconds (after a 15000 x 10000 export the workers keep about 12 MB instead of about 290 MB), and a kernel that fails is replaced and the job finishes the slower way instead of failing.
  • Helper model downloads are checked. Helper models (Settings › Helpers) are now checked against their published checksum after the download, where Hugging Face lists one (all but RMBG-2.0, which is gated); a download that does not match is deleted and can be started again, and a broken download picks up where it stopped. Your Hugging Face token is sent to Hugging Face only. While a model of two files downloads (SAM2's encoder and decoder), the row says which file it is at and how much of the whole download is in.
  • SAM2 large downloads again: its decoder is 24 bytes longer on Hugging Face than Scumble expected, so the download always stopped at the size check. A download that stopped there finishes at the next click without fetching the file again.