Generate Panel Reference

This page describes each section of the Generate panel in detail. For a first walkthrough — choosing a model, generating a first image, and sending results to the Gallery or the Canvas — see Generating and modifying still images.

Setting the image size

The Size section sets the width and height of the image.

• Width and Height can be set by dragging, by typing a value, or by dragging the corner of the small size preview. Sizes run from 64 to 4096 pixels and snap to a multiple the model requires, usually 8, 16 or 32.

• Aspect ratio opens a picker with common shapes: 21:9, 16:9, 3:2, 4:3, 1:1, 3:4, 2:3, 9:16 and 9:21, or Free. Choosing a ratio locks it and keeps the current pixel count, so changing the width changes the height to match. The lock button between the two fields locks or unlocks the current ratio. Arrow keys move between the shapes.

• Your ratios in the same picker holds ratios you name yourself, such as Cinema at 2.39:1. Add ratio starts from the locked ratio when the lock is on, otherwise from the current size’s ratio; each saved ratio shows the size picking it would set, and can be edited or deleted (with Undo). You can keep up to 50. Saved ratios are only shapes, never sizes, so they work with every model, and they belong to your account, shared by every project and by the canvas frame tool. A locked ratio you have not named shows as its terms, such as 12:5.

• Swap width and height turns a landscape image into a portrait one and back.

• Each model has an optimal image area, typically about one megapixel for newer models and less for SD 1.5. The line under the fields shows the image’s size in megapixels, marks the recommended size, and estimates the run time from your recent runs.

• Reset size, the button beside Swap width and height, opens a menu. Each item shows the size it sets, and the item matching the current size is checked; choosing it again changes nothing:

  • Optimal size for the current ratio — the model’s optimal area at the current aspect ratio, named by its saved name when it has one. It is absent at 1:1, where it would be the native square.
  • Model defaults — the size set in the model’s default settings. It appears only when the model sets one. A locked ratio stays locked and takes the default size’s ratio; an unlocked one stays unlocked.
  • Square at native size — a 1:1 square at the model’s native size, such as 1024×1024 for SDXL and FLUX or 512×512 for SD 1.5, with the ratio locked.

  Right-clicking the width and height and choosing Reset to model default also resets the size this way.

When the Canvas is where you generate, the section says Linked to the canvas frame: the width, height and aspect lock are the generation frame’s, and the frame’s aspect lock is this one. Any change here, typed, dragged or reset, resizes the frame around its center, and while the canvas is open that resize is a canvas step, so undoing it on the canvas puts the frame and the size back. Resizing the frame on the canvas changes these fields; a frame you resize freely leaves the ratio Free even when it happens to match a preset, so only a ratio you choose or lock holds.

Models produce their best results near their optimal size. Much larger images can show repeated subjects or distorted anatomy. To make a large image, generate at the optimal size and then upscale.

Render settings

The Render section controls how the image is generated. Each slider has a mark at the model’s recommended value. A field you have changed shows a small dot to its left; click the dot to reset the field to the model default. You can also double-click the field, or open its menu by right-clicking it (or with the menu key or Shift+F10 while it has focus) and choose Reset to model default.

Steps

Steps is the number of denoising passes the model makes. More steps add detail and refine the image, up to a point where further steps change little and only take longer. Distilled “Turbo” and “Lightning” models are designed for a handful of steps (often 4 to 9), while standard models usually need 25 to 50. The model default is a good place to start.

CFG and Guidance

CFG (classifier-free guidance) sets how strictly the model follows your prompt. Higher values follow the prompt more literally but can look harsh or oversaturated; lower values are looser and more natural. Typical values are 3 to 7 for most models, and 1 for distilled models, which are trained to work without CFG. Some models, including FLUX, show Guidance instead: a value built into the model that plays the same role, typically around 3.5. The model’s page under Local Models gives recommended values.

Scheduler

The Scheduler is the algorithm that decides how much noise to remove at each step. Schedulers differ in speed, in how many steps they need, and in the character of the result. Euler, the DPM++ family and UniPC are good general-purpose choices; LCM is only for LCM and similar distilled models. The list of schedulers depends on the model, and some models use a fixed schedule and do not show this field. When you have picked another scheduler, the dot to the left of the field switches back to the model’s own choice.

Seed

The Seed sets the starting noise for the image. The same model, prompt, seed and settings reproduce the same image, which lets you change one setting at a time and compare the results.

• New seed picks a random seed.
• Seed mode sets how seeds are chosen: Random picks a fresh starting seed each time you press Invoke, Fixed reuses the seed in the field, and Increment and Decrement use successive seeds, counting up or down by one. A line under the field shows the seeds the next batch will use.
• Recent seeds shows thumbnails of recent images. Click one to reuse its seed; the mode switches to Fixed.

Model-specific options

Some families add their own controls to the Render section:

• DyPE (FLUX.1) — position extrapolation for sizes above the 1024px training resolution. Auto turns it on above 1536px, Area by image area and 4K optimized is a fixed tuning; Manual exposes the scale (λs) and exponent (λt).
• Shift (Z-Image, Qwen Image) — an override for the sampling schedule’s timestep shift, from 0.1 to 10. Left on auto, the model derives it from the image size; Qwen Image’s Lightning LoRAs are trained for 3.
• Seed variance (Krea-2, Z-Image) — perturbs the prompt conditioning so nearby seeds give more varied images; Variance strength and Randomized channels set how much.
• Conditioning rebalance (Krea-2) — see Krea-2.
• Low-noise guidance (Wan) — guidance for the low-noise expert; left unset, it reuses the main guidance.
• Skip layers (negative branch) (MOSS-Image) — the transformer blocks the negative-prompt pass leaves out; see MOSS-Image.

Negative prompt

A negative prompt lists what you do not want in the image, such as blurry, watermark, extra fingers. Models that support one show a + Negative prompt… button under the prompt; click it to open the box, and click Turn off negative prompt to hide it again without losing the text.

Not every model uses a negative prompt. SD 1.5, SDXL, SD 3.5 and CogView4 always use it. Several newer models, including Z-Image, Qwen Image, Qwen-Image-2.1, Krea-2, FLUX.3 Klein Base and MOSS-Image, use it only when CFG is above 1, and the box says so. FLUX.1, FLUX.2 and Ideogram 4 do not use a negative prompt, and the box is not shown for them.

Refiner

The Refiner section appears for SDXL models. Pick an SDXL Refiner model to have it finish the last part of the denoising schedule; the base model stops at Refiner start (0.8 by default, so the refiner handles the final 20%) and the refiner continues from there with its own Refiner steps, Refiner CFG and Refiner scheduler. The Positive and Negative aesthetic score condition the refiner on the aesthetic rating it was trained with. A Refiner start of 1 turns the refiner Off and disables its tuning controls. Lower Refiner start to reactivate the selected refiner; the start control and model picker stay available while it is off. The controls are disabled until a refiner is selected, and refiner models are only offered here, never as the main model. The refiner also runs for SDXL generations from the Canvas. It cannot be combined with PiD Decode. See SDXL.

Advanced settings

The Advanced section appears for models with extra options, and its header counts the settings you have changed. Depending on the model it can include:

• VAE and VAE precision — for SD 1.5 and SDXL, choose a different VAE or decode at FP32 precision. Other families set their VAE in the Components section.
• CLIP skip and CFG rescale — SD 1.5 and SD 2.x only.
• HiDiffusion — improves structure at large sizes for SD 1.5 and SDXL; see HiDiffusion.
• Color compensation — SDXL only.
• PiD Decode — decodes the image at up to four times its size; see PiD Super-Resolution Decode.
• Encoder quantization — Qwen Image only; quantizes the Qwen VL text encoder to 8-bit or 4-bit on load to save VRAM. A quantized encoder cannot stay in the model cache, so it is loaded from disk on every generation, twice when CFG is above 1.
• Reference K/V cache — Krea-2 only, and only with active Edit references and an enabled LoRA in Concepts. Turn it on for edit LoRAs trained with isolated references (ai-toolkit’s kv_cache option), such as the thedeoxen Krea-2 pose LoRA; the references then run through the model once per generation, which is faster. Leave it off for other edit LoRAs. See Krea-2.
• Seamless tiling — makes the image tile along the X axis, the Y axis, or both.

Two output filters live in the Generate widget’s settings (the cog in its header, or Settings › Generate) rather than in the panel’s form, because they apply to every image the project’s Generate panel and Canvas produce: the NSFW checker blurs images it flags as explicit, and the Invisible watermark stamps every saved image. Both are off by default. Changing one in an open project changes it for that project only; set your default for every project on the Launchpad under Preferences. Upscale results, workflow runs and images from external models are not filtered.

Resetting to defaults

Each model has recommended settings, which you can change in the Model Manager. Reset all to model defaults, the ↺ button in the Generate panel’s title bar, applies them: size, steps, CFG or guidance, scheduler and VAE settings. Your prompts and concepts are left alone, and a notification offers Undo to bring your earlier values back. The button is disabled when your settings already match the defaults.

To return to the settings of an earlier image instead, open Use in the Preview’s header, or right-click the image in the Gallery or Preview and open Reuse from image. All settings restores its model, concepts, prompts, seed, size, render and refiner settings; Remix image does the same with a new seed; and Use prompt, Use seed and Use size recall one thing at a time. Recalled concepts replace the ones in the panel at the weights the image used; any that are not installed or do not fit the model are listed in a notice. If the image’s model is not installed, or the image was saved from the canvas, your current concepts are kept.

To keep a combination of settings for later, use Presets in the panel’s title bar to save the current settings under a name and apply them again. When you save, Include prompts decides whether the preset keeps your prompts too; the panel remembers your choice. Applying a preset saved without prompts keeps the prompts you have. Applying a preset shows a notification with Undo, which restores the settings you had before. Deleting a preset, or a saved Krea-2 conditioning rebalance curve, takes effect immediately; a notification with Undo brings it back.

Adding concepts

Concepts, also called LoRAs, are small add-on models that teach the main model a style, a character, an object or a look. They are added in the Guidance section.

1. Click Add concept and choose a LoRA. The list only shows LoRAs made for the selected model’s family.

2. Adjust its Weight, which sets how strongly it acts. The default is the LoRA’s own recommended weight, or 0.75. Values between 0.5 and 1 are typical; higher values can overpower the image, and negative values push the image away from the concept.

3. Add any trigger words the LoRA needs to your prompt. They are shown under the LoRA’s name, and the + button in the prompt toolbar inserts them for you (see Prompt Triggers).

You can add as many concepts as you like and combine them. Use the switch on each row to turn a concept off without removing it, and the trash button to remove it. If you change to a model from a different family, the concepts it cannot use stay in the list, marked Incompatible and switched off; they run again when you switch back to a model they fit. A concept that is no longer installed is marked Not installed. CogView4, MOSS-Image and external API models don’t support concepts, so with one of them selected the section says so instead of offering LoRAs.

Adding reference images

A reference image guides the new image by example rather than by words: its style, its composition, a person or an object in it. Reference images are added at the top of the Guidance section, and only appear for models that support them.

To add one, click Add reference image and pick images from the Gallery or upload new ones, or drag images from the Gallery onto the section. You can also right-click any image and choose Use as reference image. Most models accept up to five. The order of the references matters to some models; use the arrow buttons on each card to reorder them.

Hovering over a reference’s thumbnail offers Crop, to use only part of the image, and Use size, to set the generation’s size to match the reference. The switch on each card turns a reference off without removing it.

In the crop dialog, drag the crop box or its handles, or pick an Aspect ratio (16:9, 3:2, 4:3, 1:1, 3:4, 2:3 or 9:16) to keep the box in proportion while you resize it. Scroll to zoom in on the image, hold Space (or use the middle mouse button) and drag to move the view, and use the Fit image, Actual size, Zoom in and Zoom out buttons. Apply uses the crop for the reference; Save to Uploads also stores the cropped image as an upload in the Gallery without changing the reference.

What a reference does, and what options it has, depends on the model family:

Model family	How references work	Options
SD 1.5, SDXL	An IP-Adapter model carries the reference’s look into the image.	Mode: Style, Composition or Both. Weight. Under Advanced: the IP-Adapter model, a Style variant and the Active steps range.
FLUX.1	An IP-Adapter, or a FLUX Redux model, which transfers the reference’s content.	IP-Adapter options as above, or Image influence (Lowest to Highest) for Redux.
Anima (not Anima-3.8B)	An Anima IP-Adapter carries a character’s appearance (hair, outfit, accessories) into the image. Up to four references (experimental); images of one character combine well, different characters blend.	Strength (0 to 2; 0 turns it off), per reference. Under Advanced: the IP-Adapter model and the Active steps range.
FLUX.1 Kontext	Kontext edits or recombines the reference images according to your prompt.	None beyond the model.
FLUX.2, Qwen Image Edit, Qwen-Image-2.1	The model reads the references directly and uses them as your prompt instructs, for example put the jacket from the first image on the person in the second.	None.
FLUX.3	As FLUX.2, with up to ten references. The prompt names them by the tag on each card, such as <ref_image_0>; click the tag to copy it.	None.
Krea-2	Style: one reference’s look — palette, texture, rendering — is transferred into the image while the prompt keeps driving the content. Edit: up to three references for an edit LoRA, such as a pose skeleton or a depth map, read in card order. See Krea-2 style reference and edit LoRAs.	Mode: Style or Edit. Style strength (0–2; 0 turns the reference off) for Style.

Reference images require an extra model for SD 1.5, SDXL, FLUX.1 and Anima (an IP-Adapter or FLUX Redux model). These are available as starter models in the Model Manager, and an installed IP-Adapter is selected automatically.

Qwen-Image-2.1 reads its references through the Qwen3-VL vision tower, so it needs a Diffusers model, a non-GGUF Qwen3-VL 8B encoder, or a GGUF 8B encoder together with its Vision Tower (the mmproj file) under Components. Without one (a single-file model with no encoder selected, or a GGUF encoder with no Vision Tower) the Guidance section says so instead of offering references. See Qwen-Image-2.1 reference images.

A new Krea-2 reference starts as Style when it is the only active one; added beside other active references, or several at once, it starts as Edit. A Style reference at strength 0 is not active. Adding an image beside a lone Style reference still at its default strength turns that card into Edit too; one whose strength was changed keeps its mode. Krea-2 runs either one Style reference or up to three Edit references, never both; Invoke stays disabled and says why until the active references fit. The Mode can be changed while a card is switched off. Switching a card back to Style resets its strength to 1.

Prompt tools

The toolbar above the prompt box holds tools that help you write prompts. From left to right: Templates, Dynamic, Add prompt trigger, Expand prompt, Image to prompt and Prompt history.

Prompt enhancement

Expand prompt turns a short prompt into a detailed description using a local language model. Write a few words, open the tool, pick a Text LLM and a System prompt (the instructions that tell the language model how to rewrite), and click Expand. The expanded text replaces your prompt, and you can edit it before pressing Invoke. Several system prompts ship with InvokeAI, some tuned for particular model families, and you can write your own.

You need a Text LLM model installed; if there is none, the tool offers to open the Model Manager. See Prompt Tools for suitable models and for managing system prompts.

Image to prompt

Image to prompt writes a prompt that describes an existing image, using a vision-language model (LLaVA OneVision). Select an image in the Gallery or Preview, or drag one onto the prompt box, then open the tool, choose the model and click Generate Prompt. The description replaces your prompt. This is useful for recreating an image you like in another model, or for starting from a close description and editing it. See Prompt Tools.

Prompt history

Every prompt you submit with Invoke is kept in the project’s Prompt history, along with its negative prompt, up to the last 100. Open it from the clock button to search your earlier prompts, click Use prompt to bring one back, or delete entries you no longer need. Clear empties the history at once, without asking.

To step through the history without opening it, press Alt+Up and Alt+Down while the prompt box has focus. Stepping past the newest entry brings back what you were typing. The history belongs to the project and is shared by the Generate, Video and Upscale panels.

More prompting features

These features change what the prompt says or how the model reads it. Prompt triggers are described on the Prompt Tools page, and the others in the Prompting Guide.

Feature	What it does
Prompt templates	Wrap your prompt in a reusable style, such as a photographic or concept-art look, without retyping it. Includes 18 built-in templates, and you can create, import and export your own.
Dynamic prompts	Write alternatives such as a {red|green|blue} car to generate one image per combination, or a random sample of them. Wildcards let you reuse named lists of values. Switch off Use dynamic prompts in the popover to use the prompt exactly as typed, for example to keep a # that would otherwise start a comment.
Attention weighting	Make parts of the prompt count for more or less, for example (red car)++ or (red car)1.3, and blend or combine prompts. Select text and press Ctrl+Up or Ctrl+Down to adjust its weight. Applies to SD 1.5, SD 2.x and SDXL.
Prompt triggers	Insert the trigger phrases of your model and concepts, compatible embeddings and wildcards from a searchable menu.