How to work with Image Bench.

Opening the app

1. Open /image-bench/app on your host, for example http://127.0.0.1:7860/image-bench/app, and bookmark the address. This is the usual way to start Image Bench.
2. Image Bench opens in its own browser tab and loads the host UI into one of its own tabs: the one with the host icon, after Settings.
3. If Image Bench is already open, opening the address again brings the open copy to the front and closes the new tab. Where the browser does not allow a page to take focus, the open copy flashes its title instead.

Image Bench tabs

The tab bar follows the usual course of work from left to right. You set up a generation in Generate, ControlNet or Explore, and the results arrive in Generated. From there you move the images worth keeping to the trays, compare candidates in Compare, and upscale the best ones into Upscaled. Library, the project chip, Settings and the host tab sit at the right end of the bar.

The numbers on the tabs count the images each one holds. A tray tab flashes when an image is added to it, and you can drag an image card onto a tray tab to put the image there.

WebUI and ComfyUI

Image Bench looks and works the same on both hosts. Controls that the host or the current workflow cannot use are hidden, so a ComfyUI workflow without a Hires. fix sampler has no Hires. fix group. The cheat sheet lists which features each host supports.

Start from a bundled workflow

At startup, Image Bench copies its own reference workflows into the Image Bench folder of ComfyUI's Workflows sidebar. These workflows are already marked. Choose one from the workflow dropdown in the Generate toolbar and use it as a template for your own.

Marking your own workflow

1. Open the workflow in the host tab.
2. Right-click a node, choose Title, and enter a marker from the table below, for example ib:cp:0 for the checkpoint loader.
3. Save the workflow. The Generate tab now shows the controls that the markers enable.

LoRAs in the prompt

The Image Bench Prompt node reads the <lora:name:weight> tags in its text, applies those LoRAs, and passes the text on with the tags removed. Mark this node ib:pp, and the LoRA list in the Generate tab works the same way it does in WebUI.

The layout

The Generate tab follows the layout of the SD WebUI txt2img and img2img tabs, so the controls are where a WebUI user expects them. The common controls fit in a 1200×800 window without scrolling; the ones you need less often sit behind buttons and switches and appear in the same place when you click them. The numbers in the screenshot match the list below it.

1. The Generate button starts a run and becomes Stop while the run is in progress. Next to it are Skip and the number of images to generate. Variations… generates seed variations (see chapter 6).
2. The × in front of the image count resets it to 1.
3. The count button sets the number of images to the value of Settings › Images set by the count button, 128 in the screenshot. It is highlighted while the count has that value. WebUI WebUI generates at most 100 images per batch unless you raise the Batch count maximum in ui-config.json in the WebUI folder and restart WebUI.
4. The ∞ button generates images until you press Stop. Click it again to go back to the number of images you had before.
5. The prompt tools, from left to right: Snippets, PNG Info, the eraser that clears both prompts and the console drawer. With Settings › Show the prompt lock button, the prompt lock appears before the console drawer button.
6. The mode switch selects txt2img, img2img or Inpaint.
7. The WebUI UI preset, which sets up WebUI for a model family.
8. The checkpoint dropdown. The button in front of it switches the dropdown to the VAE and text encoder files, and the button after it opens the checkpoint's page on Civitai.
9. The image size and the number of steps. The button in front of the size switches between preset sizes and width/height fields.
10. The prompt history buttons and the variants dropdown. While a history entry is shown, a fork button keeps it as a variant. The token count of the prompt is at the right end of the same line.
11. The Copy tab beside the prompts opens a menu that copies both prompts, either one of them, or the infotext: the prompts with all the parameters of the tab, in the text format that WebUI writes into its images. It copies the prompts as shown, so with a history entry or a variant open it copies that one. A middle click or Shift+click on the tab copies both prompts at once.
12. The LoRA and References tabs. The field next to them filters the list.
13. The parameter groups. Sampling and Seed are always shown; the third group holds the Hires. fix settings in txt2img and the img2img or inpaint settings in the other modes.
14. The vertical tabs switch this area between the parameter groups and the session filmstrip.
15. The result cell has three views: Result shows the latest image, Input the input image and Mask the mask. Use the Image and Mask buttons to upload an input image or a mask. The Input view draws the mask and the inpaint crop over the image, and the Mask view shows the mask by itself together with its size (see chapter 5).

To give more room to the prompts or to the parameter groups, drag the handles between them.

Two-way sync

While the Generate tab is open, it reads the host's controls and writes your edits back to them. If you change the steps in Image Bench, the host's slider moves; if you change the sampler in the host tab, Image Bench updates.

txt2img, img2img and Inpaint WebUI

The switch at the right end of the toolbar selects the mode and switches the WebUI tab to match the selected one. Because img2img and Inpaint share the same WebUI controls, Image Bench remembers the parameters of each mode and restores them when you switch back.

Checkpoint and VAE / text encoders

By default, the model dropdown shows the selected checkpoint . In Forge Neo, the button in front of it switches the dropdown to the VAE and text encoder files and back, and Image Bench remembers your choice. The info button after the dropdown opens the checkpoint's page on Civitai.

The VAE / text encoder dropdown is a checklist with separate VAE and text encoder sections. It sets the VAE / Text Encoder field of Forge Neo, and each change reaches Forge right away. You can filter both dropdowns: start typing while a dropdown is closed, and it opens with only the entries that contain every word you typed.

Size

The button in front of the size toggles between a dropdown of preset sizes, shown with the list icon , and width/height fields, shown with the ruler icon . The fields round to multiples of 16. Image Bench remembers your choice for each mode. Inpaint starts with the fields, which set Resize to in that mode.

The status pill

The free space in the middle of the toolbar shows one status at a time, as a pill. These are the statuses you may see:

Stop and Skip buttons

While a generation runs, the Generate button turns into Stop, and Skip next to it becomes active. Skip ends the image that is being generated and continues with the rest of the batch. Stop ends the whole run. To stop and immediately generate again with the current parameters, Ctrl+click, Shift+click or middle-click Stop; while you hold Ctrl, the button reads Restart.

The same commands have keys that work in every tab, also while you type in a text field: Ctrl+Enter generates, or stops a running generation; Ctrl+Shift+Enter restarts it; Ctrl+Shift+Backspace skips the current image. They do nothing while a dialog is open. They take effect once the Generate tab has been opened in the window, and they don't work in the host tab, where WebUI and ComfyUI have their own Ctrl+Enter.

History, diff and variants

The buttons around the prompt fields give you a prompt history, a diff of any earlier prompt against the current one, named variants, snippets and PNG Info.

1. Each generation you start from the tab adds its prompt pair to the history, which keeps the last 20.
2. The history buttons above the prompt step through the entries. The pencil button returns to the current prompt, which is the one you can edit.
3. A history entry is displayed as a read-only diff against the current prompt, with removed text struck through and new text highlighted.
4. The fork button saves the displayed history entry as a named variant. Variants are listed in a dropdown of their own.

Copying the prompts

Click the Copy tab at the top of the vertical tab rail, next to the prompts, to open the copy menu in the top right corner of the prompt. When you point at a menu item, the fields it copies are outlined. The menu copies what the prompt fields show, so while a history entry or a variant is open, you get that one and not the current prompt.

• Prompt and negative prompt copies both prompts, each after a Prompt: or Negative prompt: line.
• Prompt and Negative prompt copy a single prompt.
• Infotext copies the prompts and the tab's parameters as WebUI infotext, the text WebUI saves in its images. To write it into a PNG, use Paste infotext… in the image context menu.

To copy both prompts without opening the menu, middle-click or Shift+click the tab.

LoRAs

The LoRA tab lists the LoRAs installed on the host, grouped by their folders. A LoRA is checked when the prompt contains its <lora:…> tag, so the checkboxes always match the prompt, including tags you typed or pasted. The number on the tab is the count of LoRAs in the prompt.

• To add a LoRA, check it or click its name. Image Bench appends <lora:name:weight> to the end of the prompt. Unchecking the LoRA removes its tag from anywhere in the prompt.
• The weight in the tag is the LoRA's preferred weight, and the LoRA's activation text is added after the tag. You set both in the LoRA's metadata in WebUI's LoRA card, which stores them in a .json file next to the LoRA. Values set for the LoRA in the Explore tab override the metadata. If neither is set, the weight is 1.
• To change the weight later, edit the number in the tag. The LoRA stays checked.
• The filter field above the list matches LoRA names and folder names. The refresh button next to it makes the host rescan its checkpoints and LoRAs, so you can use a newly downloaded LoRA without restarting the host.

ComfyUI The LoRA list requires the positive prompt to be an Image Bench Prompt node marked ib:pp. This node loads the LoRAs named in the <lora:…> tags of its text and passes the text on with the tags removed.

Taking parameters from an image: PNG Info

1. Drop an image on the prompt, or click the PNG Info button to the left of the eraser that clears the prompts.
2. Check the groups you want to apply: prompts, model, sampling, size, Hires. fix and so on.
3. Click Apply. Only the checked groups change.

Img2img and Inpaint

Both modes start from an existing image, and you choose between them with the mode switch at the right of the Generate tab's toolbar. Img2img takes only an input image and regenerates all of it, with the denoising strength setting how far the result may move away from the input. Inpaint takes an input image and a mask, and regenerates only the masked area while the rest of the image stays as it is.

Adding an input image

The result cell of the Generate tab has three views: Result, Input and Mask. You can set the input image in any of these ways:

• Upload a file, paste one with Ctrl+V, or drop one on the Input view.
• Drag an image card from any gallery onto the Input view.
• Use the image actions Send to img2img (M) or Send to inpaint (B).
• Send the image from Photoshop, Krita or GIMP (see chapter 11).

Inpainting with a mask from your editor

Image Bench has no mask brush. Make the mask in Photoshop, Krita or GIMP as a selection, and send it together with the image with Input / Inpaint: Both in the editor plugin (see chapter 11). A feathered selection gives the mask soft edges. You can also upload a mask file with the Mask button of the result cell.

WebUI Image Bench inpaints through WebUI's img2img › Inpaint upload tab, where the image and the mask are separate uploads. It does not use the brush canvas of the Inpaint tab.

References for edit models

Edit models (Klein, Kontext, Qwen Edit, Anima and Krea 2) take reference images. You add, reorder and remove them in the References panel, in the tab next to LoRA: drop, paste or browse for an image, or drag one from Generated or the trays. The sends to list at the bottom chooses where images from your editor go, the input image or the references. In ComfyUI, each reference slot is a Load Image node marked ib:ref:N.

Batches and the session filmstrip

When you generate a batch with more than one image, the session filmstrip replaces the parameter groups. It holds the last 265 images of the session, grouped by run. Click an image to view it large, or press the right button on it to open its action rim at the pointer. The arc below the ring shows the image in the result pane in place of the latest image; on the image already shown there, the same arc brings back the latest one.

Model and LoRA exploration

To see what several checkpoints or LoRAs make of the same prompt, select them in the Explore tab and start the run. Image Bench generates the current prompt with each of them in turn and shows the results side by side. A LoRA exploration also generates the images without the LoRA for reference, unless you turn off the No LoRA button. When the run ends, your model and parameters are restored.

Image Bench — Explore

When you select both checkpoints and LoRAs, you can switch the run from By model to Matrix. A matrix run generates one image for every combination of a checkpoint and a LoRA and lays them out as a grid, with the checkpoints in rows and the LoRAs in columns. With No LoRA on, the first column shows each checkpoint without a LoRA. All the images of a matrix use the same random seed, so the differences you see come from the models alone. Save results saves the grid as a single HTML page.

Seed variations WebUI

Variations… next to Generate makes a batch of variations of the current seed. Each image keeps the seed and mixes in a random variation seed at the variation strength you enter; small strengths stay close to the original. The batch count sets the number of images. You do not need to set the variation seed and strength in the UI and clear them afterwards; the button does both.

Automated sweeps from one image

Once you have an image you like, you may want to see whether more steps, another denoising strength or other text encoders improve it. By hand, that means restoring the image's parameters and changing one value before each generation; Image Bench runs the whole series at once.

These image actions start from the parameters of an existing image, including its seed, and generate a series in which only one setting changes. You find them on the left side of the action rim: press the right button on an image card, or hold Ctrl over the result in the Generate tab, and point at the icons below (see chapter 8). With the pointer over the card or in the overlay, the keys run them directly.

Image Bench remembers what you enter in these dialogs (step lists, denoising strengths, encoder sets) for the next run.

Setting up a unit

The ControlNet tab shows one unit at a time and keeps it in sync with the host in both directions. WebUI It has the same units, control types, preprocessors, models and parameters as the ControlNet built into Forge Neo, so what you know about Forge's ControlNet applies here. ComfyUI Each unit is a ControlNet apply node of the workflow (see the end of this chapter).

1. Select the unit. In WebUI, also select txt2img or img2img in the toolbar.
2. Add a control image: drop or paste it, use Send to ControlNet (N) on any image, or send it from your editor.
3. Choose the control type, preprocessor and model. Run the preprocessor and check the preview.
4. Set the weight and the timestep range, then generate from the Generate tab.

Editing a pose WebUI

After you run an OpenPose preprocessor, Edit pose opens sd-webui-openpose-editor over the tab; the editor extension must be installed. The edited pose is rendered into the preview, and the preview becomes the control input.

The ControlNet sweep: weight × ending step

Choosing a weight and an ending step usually takes several tries: with too much control the image follows the control map stiffly, and with too little it drifts away from the composition. A sweep renders one unit at a range of weights and ending steps with the same prompt and seed, and lays out the results as a matrix, with the weights in rows and the ending steps in columns. You can then compare the images side by side and take the values of the one you like.

Image Bench — ControlNet sweep

1. Switch the parameters column from Parameters to Sweep.
2. Set From, To and Step for the weights (rows) and the ending steps (columns).
3. Start the sweep. It first generates a reference image with the unit's current values, then the cells.
4. To use a cell's values, apply the cell, or apply a row or column header. The values are copied to the unit.

The action rim

Most of what you do with an image starts from its card: you look at its parameters, send it to img2img or your editor, upscale it, or put it in a tray. Buttons for all of these would cover the thumbnail, so the card has none. The actions are on a ring that you show when you need it, and also in the card's context menu and on single keys that work while the pointer is over the card.

Press the right button on a card, and the ring opens around the pointer. Move to a sector and release the button there: the action runs and the ring closes. If you release the button without moving, the ring stays open and you can click a sector. A click outside the ring, Esc or the mouse wheel closes it.

The rim toggle in the top right corner of a card shows the ring around the center of the image instead. It stays until you click the toggle again or the pointer leaves the card.

The actions sit in fixed places: parameters at the top, send actions and trays on the right, Remove at the bottom and generate actions on the left. The label in the middle of the ring names the action under the pointer. Clicking the image, inside or outside the ring, still opens it. Holding Ctrl over a card hides the ring, the toggle, the badges and the other controls, so you see only the image, and a click opens it.

The image view

Click a card, or press V while the pointer is over it, to open the image view over the whole window. The view has no toolbars, so the image gets the full window area.

Image Bench — image view

1. Press and hold the right mouse button anywhere on the image. The rim opens around the pointer. Move the pointer to a sector and release the button to run that action. If you release the button outside the rim, nothing happens and the rim closes.
2. If you release the button without moving the pointer, the rim stays open. You can then click a sector with the left button. To close the image view, click in the center of the rim with the left, middle or right button. Because the rim opens with its center under the pointer, two right clicks without moving the mouse close the view: the first opens the rim, the second clicks its close button. To close only the rim, click anywhere outside it or press Esc.
3. The ring has the same actions as the rim of a card. In the trays, it also has Delete. The arc above the ring sets a marking. The arc below the ring has Reuse seed and Reuse all, which copy the seed or all parameters of the image to WebUI or ComfyUI. The parameters sector shows the generation parameters over the image. The parameters stay visible when you move to another image.

Click the image to show it at full size. The point that you clicked stays under the pointer. Drag the image to pan it, and click it again to fit it to the window. To close the image view, click the dark area around a fitted image or press Esc.

To go to the previous or next image, turn the mouse wheel or press ← or →. At full size, the next image also opens at full size and shows the same area. This lets you compare one detail across a series of images. The rim always applies to the image that is currently shown, even if you scroll to another image while the rim is open.

If the image has a marking, the marking is shown in the lower right corner. Click it to remove the marking. A middle click opens the same context menu as on a card, which includes Open in new window. The action keys and the digit keys work in the same way as over a card.

Keys

Each action has a key that acts on the card under the pointer or on the image open in the overlay. Keys are matched by their position on the keyboard, so they work with any keyboard layout. The cheat sheet lists all of them.

Markings

The digit keys 1 to 9 set a marking (👍 👎 🔴 🟠 🟡 🟢 🔵 🟣 🏁), and 0 removes it. Clicking the marking on a card filters the view by that marking; middle-clicking it removes the marking.

Drag and drop

• Drag a card by its image onto a tray tab. Images dragged from one tray to the other are moved; images from any other view are copied. If the card is selected, the whole selection is dragged.
• In the trays and in Upscaled, drop a card onto another card to change the order.
• Drop a card on the Input or Mask view of the Generate tab, or on the ControlNet control image.
• Drop cards on the project chip to copy them into the project folder.

The context menu

Middle-click an image for its context menu. Its first row has the destinations as colored buttons: send to img2img, inpaint, ControlNet and the editor, and add to Compare and the trays. Below the row are the other actions, followed by the commands that are not on the ring: Open in new window, Copy image, Copy to project folder, Copy infotext and Paste infotext, and a markings submenu with Show only.

Two trays

The Generated tab keeps its images in memory only, so move the ones you want to keep into one of the two trays. The trays work the same way, and how you divide images between them is up to you. You can use both as long-term collections, or use one for keepers and the other as a scratch area for images you need only briefly, for example to compare them or run them through img2img once before you delete them.

The trays survive reloads and restarts, and a project saves their contents. One way to use both as collections:

• Concept tray (Y): ideas worth keeping, such as compositions, styles and happy accidents.
• Image tray (T): the best images, the ones you will upscale, pass on or finish.

You can download each tray as a .zip file and import a .zip file into either tray.

Compare

1. To add images, press A, use the action rim (on a card or in the image view), drop files on the tab, or send them from your editor.
2. Zoom with Ctrl+wheel and pan by dragging. All images follow and show the same area, even if their resolutions differ.
3. In select mode, drag to draw a region; it applies to every image. The crop grid then shows only these regions, side by side.
4. Export the crops as a single PNG.

Setting up

You do not need a project to keep your work between sessions: Image Bench keeps the trays and the Compare tab across reloads and restarts on its own. Projects are for when you work on several things in turn. A project saves the trays and Compare together with the generation parameters into a folder of its own, so you can put one piece of work aside, open another, and come back to the first one later in the state you left it.

1. Set the project root in Settings › Library. It can be a local or network path, written as the server sees it.
2. Create a project with the project chip at the right end of the tab bar. Each project is a folder with an ib_project/ subfolder.
3. Work as usual, and save with the chip or with Ctrl+S.

What a project contains

• The trays and Compare, as they are when you save.
• A snapshot of the host parameters, the enabled ControlNet units with their images and, in ComfyUI, the workflow.
• The prompt history, tags, notes and the final image.
• Any other files in the folder, shown in tabs with their generation parameters.

The project reminder

While a project is open, a cube in the project color is attached to the Generate button. After you clear the prompts, apply PNG Info or reopen the window, the next Generate asks whether you are still working on this project. Choose Continue project (or press Enter), or New workspace, which starts a blank workspace that keeps the prompts and seed.

Setting up

Some work on a generated image is easier by hand: a mask for inpainting, a fix painted over a flawed area, a collage for img2img. The plugins for Photoshop, Krita and GIMP move images between the editor and Image Bench, so you do not have to save a file in one program and load it in the other. From the editor you send the document or the selection to the input image, the inpaint mask, a ControlNet unit or the Compare tab, and from Image Bench you send any image to the editor, where it opens as a document or a layer.

1. Download the plugin from Library › Manual › Editors and install it.
2. In the plugin panel, enter the host address (the one you open Image Bench from) and connect.
3. The send actions in Image Bench now show the editor's name.

From the editor to the host

The plugin sends the image, and the selection as a mask, to one of the targets below.

From the host to the editor, and keeping the result

• Send to <editor> (P) opens an image in the editor. Shift+P adds it as a new layer to the open document.
• Keep › File in the plugin saves the layered document into the current project folder. Keep › Final does the same and also makes the merged image the project's final image. Both require an open project.

Docs

The Docs section lists the Markdown files under the docs root as a tree and displays them read-only, including their images and the links between them. You set the docs root in Settings › Library; when it is empty, Image Bench shows the model docs that ship with the extension. Point the root to your own notes folder to read your notes here.

Tag lists, the .csv files used for tag autocompletion, appear in the tree as well and can be searched. Clicking a tag copies it with spaces in place of underscores.

Snippets

Until you save snippets of your own, the list contains a default set of styles taken from comfyui-styles-all and Fooocus. Each snippet has a prompt and a negative prompt, and you can organize them in folders. In a snippet, {} or {prompt} marks where the selected text goes when the snippet wraps it.

Civitai model info

The info action of a model looks the model up on Civitai by its hash and shows the result in a dialog. Results are cached in store/civitai/; to refresh one, delete its file there.

The plug icon

When the backend cannot be reached, the plug icon appears after the Library tab. Until the connection is back:

• You can still edit parameters in the Generate and ControlNet tabs.
• Actions that need the server, such as generating, saving or loading lists, are refused, and a single notice tells you why.
• Tray saves that failed are retried after the reconnect.

When the connection returns, Image Bench reloads the WebUI page and sends it your current state again. ComfyUI is not reloaded, but a run that was in progress is ended.

The console drawer

Press Ctrl+` in any tab, or middle-click or Shift+click the host tab, to show the host's terminal output in the lower part of the window. Drag the top edge of the drawer to resize it. Progress bars and colors appear as they do in the terminal.

Restarting a hung host

A hung server cannot restart itself. The optional Host Restart addon runs a small separate process that can restart it. You install the addon by hand; it is not shipped with the extension, because security software may flag its process control.

1. Copy image-bench-host-restart next to the extension. The same folder works for WebUI and ComfyUI.
2. In the addon's store/host_restart.json, enter the command that starts your host, for example {"command": "D:/ai/sd/webui/run.bat"}. This file can only be edited by hand: a setting in the app would let anyone who can reach a --listen host run any command.
3. Restart the host once. The power button of the console drawer and the plug icon now offer a restart.

A host on another computer

Start the host with --listen and open Image Bench at that computer's address. To copy images to the clipboard, run the host over https with --tls-keyfile and --tls-certfile. A self-signed certificate works once you have accepted it in the browser.

Image keys (on a hovered card or in the overlay)

Anywhere in the app

Mouse

Settings that change behavior

WebUI and ComfyUI