StableSwarmUI, A Modular Stable Diffusion Web-User-Interface, with an emphasis on making powertools easily accessible, high performance, and extensibility.
MIT License
Bot releases are hidden (Show)
Well! Hi there, it's been a minute. Welcome to 0.6.2 Beta -- this one got a rather long time before I finalized the release, as I wanted the "Simple" tab ready to go first. So as a side effect... this update has more new features than any other release thus far, and more bugfixes/QoL improvements/etc than I can even be bothered to try to count.
a photo of a <alternate:cat,dog>
, and a prompt-editing-timestep syntax a photo of a <fromto[0.2]:green, blue> cat
-- https://github.com/Stability-AI/StableSwarmUI/discussions/11#discussioncomment-9206140
ModelSpecificEnhancements
under Swarm Internal that you can uncheck to disable SDXL rescond smartness and possibly other enhancements like it in the futureMask Behavior
inside init image, defaulting to Differential
now with the ability to swap back to Simple Latent
. This makes masked image gen (including eg <segment:
usage) a bit cleaner/betteralign_your_steps
scheduler, optimal for 10 steps but can do less, useful especially for coherent SVD outputs without taking a year to processMask Shrink Grow
parameter, for when using image editor / masks, to crop the image to just the mask, then grow the mask by a few pixels, and generate based on that. This is useful for refining small details.Published by mcmonkey4eva 7 months ago
Swarm is now in Beta status! After months of "Alpha" status, I'm ready to say Swarm is one of the best Stable Diffusion UIs out there, and for many use cases simply the best.
?
on every param to tell you how each feature works, and most features you might care about are available out of the box. There's tab completions for prompt syntax. Everything is ready to help you.ShiftedLatentAverageInit
under advanced sampling - empty latents are normally initialized as zeroes, but that's not the correct average value for VAEs. Check this box to use actual channel-average values (based on FFHQ dataset as calculated by Birchlabs https://huggingface.co/datasets/Birchlabs/sdxl-latents-ffhq)<break>
syntax to separate conditioning manually to multiple parts--
ps yes of course when SD3 comes out Swarm will support it right away.
Published by mcmonkey4eva 9 months ago
Swarm is gearing up for a transition out of Alpha status into Beta. Final work to do before it's ready for "Beta" includes documentation, translations, and finishing up the initial image editor tool. The Beta will most likely also be relicensed as the fully FOSS MIT License.
Swarm is already at a point where it's one of the best possible UIs you can have for Stable Diffusion, and there's very few reasons to not use Swarm. Reaching Beta will mark the point where the reasons to use anything other than Swarm are almost nil. I'm incredibly proud of the progress this project has made over the past months in Alpha.
Unsampler Prompt
, use to unsample your input image before resampling. This replaces the random noise with noise based on the input image + a prompt, allowing for more tightly controlled image editingno seed increment
if for example you want to muck with wildcards and not have seed change, or whateverSKIP:
prefixNo Previews
advanced param under Swarm Internal. Implicitly enabled by default for grid generations. May make some generations slightly faster.Refiner Save Before Refine
advanced param (under Refiner)FreeBackendMemory
, buttons on Server Info tab, and server settings for autoclear after specified timePublished by mcmonkey4eva 10 months ago
InitImageResetToNorm
, lets you overdrive InitImaageCreativity(denoising) above 1.0 by mixing the init with latent emptinessLogs
tab under Server
, so you can look through the raw server logs if needed entirely in UI. This also lets you view the Debug logs without having to enable them in the main terminal.restart all backends
quick buttonRemoveBackground
under advanced sampling, to just quickly remove the background from your images<clear:x>
works like the <segment:
node, but removes the content from the image - just like RemoveBackground but targetable with ClipSeg.ControlNet Start
and End
, to control which steps a controlnet applies to (only visible if Advanced params is checked)Refiner HyperTile
param (a technique that makes sampling a bit faster on large images)Generate
and click Generate Previews
, it will now try to use a preset named Preview
, and can even help you generate one if you have an sdxl_lcm_lora
available alreadyPublished by mcmonkey4eva 11 months ago
MultiGPU
selector choose All
, then create a workflow where the output nodes (PreviewImage/SaveImage) have different colors (use 1 unique color per backend). Each output's sub-graph will be split to a different comfy backend - see video here https://github.com/Stability-AI/StableSwarmUI/discussions/11#discussioncomment-7555872
SwarmLoraLoader
node to let the main UI lora selector work, see here for detailsVAE Tile Size
param under Advanced Sampling
for if you need VAE tilingused_embeddings
to metadata when you use TI embeddingsSet As Refiner
quickly?
clickies) between buttons, hoverables, or none (hidden)CLIP Stop At Layer
under Advanced Sampling
for if you want that for old SDv1 NAI style modelPersonal Note
parameter, under advanced->swarm internal, if you want to handwrite a note into the metadataOpen In Folder
button on the hamburger menu for imagesPublished by mcmonkey4eva almost 1 year ago
<segment:face> a beautiful face
on the end of a prompt to trigger automatic inpainting of segments matched by text<object:...>
syntax to complement <region:...>
except object
emits additional generation calls in the middle to more aggressively form the wanted imageFreeU Apply To
param to control whether freeU applies to base, refiner, or bothReVision Zero Prompt
param to force ReVision to zero the prompts given to the base (and leave unzeroed for refiner)Published by mcmonkey4eva about 1 year ago
<embedding:filename>
, <lora:filename:1>
, <random:a, b, c>
, <preset:presetname>
- see https://github.com/Stability-AI/StableSwarmUI/discussions/11#discussioncomment-7057501 for detaila photo of a catdog <region:0,0,1,0.5> a photo of a cat <region:0,0.5,1,0.5> a photo of a dog
(proper UI for it and more advanced tooling TBA)StepSwapNoisy
method if you want to use the official-sample-workflow's way of handling noise, which is a bit better for standard refiner (SDXL Base + SDXL Refiner, No upscale) but much worse for anything else.PostApply
, as that always works, whereas StepSwap only works for some cases. (You can always change your personal defaults at will by just creating a preset and naming it Default
!)Published by mcmonkey4eva about 1 year ago
Published by mcmonkey4eva about 1 year ago
I haven't written release notes thus far because there's just so much but I'm going to try for this one.
The major updates here and other features are announced as they happen in the Feature Announcement Thread, I highly recommend following that thread!
This is the first release containing the in-progress UI rework. (There's more to come on that still). Things look differenty now! Hopefully you like it better!
This also contains ReVision support. When using SDXL, drag images into the prompt box to have them parsed as part of the prompt. It's pretty neat!
This release contains a significant amount of minor UI and QOL improvements, and a few fun new features
That. is not. A complete list. That's just the highlights. There's a lot more. Do you see why I normally don't write release notes? There's a lot.
Published by mcmonkey4eva about 1 year ago
Published by mcmonkey4eva about 1 year ago
Published by mcmonkey4eva about 1 year ago
Initial public Alpha of StableSwarmUI.