Forge UI Glossary
Decode the labels before changing them. These definitions explain each term’s job, the term it is most often confused with, and the guide that shows it in a working task.
Read a Forge control as one part of a chain.
A model file supplies a component. A generation control changes the run. A memory control changes where the work fits. An editing tool adds another input. The same word—especially “model,” “weight,” or “strength”—can belong to different parts of that chain.
This glossary covers Forge interface and workflow language. It is not a general machine-learning encyclopedia, a compatibility promise for arbitrary files, or documentation for reForge, Forge Classic, or Forge Neo.
Six pairs that look related but do different jobs.
Search the label you see in Forge.
Search accepts terms, aliases, and definition words. Category buttons narrow the same complete, server-rendered list; they do not load definitions after a click.
Showing all 63 terms
Forge & interface
6 termsCommit
A specific recorded state of the repository, identified by a Git hash such as dfdcbab.
A branch can move after an update, while a commit hash names one fixed code state. A release date and commit date are not interchangeable.
Extension
Additional code that adds or changes WebUI behavior, commonly installed under the extensions directory.
An extension can be version-sensitive. Its presence in the interface does not guarantee compatibility with the current Forge commit.
Forge
The project in lllyasviel/stable-diffusion-webui-forge, built from the familiar Stable Diffusion WebUI interface with a resource-management backend and integrated features.
reForge, Forge Classic, and Forge Neo are separate downstream projects; their features and releases do not guarantee Forge behavior.
Integrated extension
Feature code shipped inside the Forge repository under extensions-builtin, such as its ControlNet and LoRA integrations.
Integrated means bundled with this repository; it does not mean that every third-party model file or workflow is included or supported.
UI preset
The top-level Forge selector that changes which model-oriented controls are shown; current choices include sd, xl, flux, and all.
A UI preset does not download or load a checkpoint and is not a model-family converter.
WebUI
The browser interface used to control a locally running Forge process.
Closing the browser tab does not necessarily stop the Forge process; localhost is an address on your machine, not the application itself.
Models & files
14 termsCheckpoint
The primary model selection used to generate an image; in Forge it may be a traditional bundled checkpoint or a supported diffusion-model file.
A checkpoint is not a LoRA, VAE, text encoder, or ControlNet model, even when all use.safetensors files.
CLIP-L
A CLIP text-encoder component used by some supported model layouts, selected through Forge’s VAE / Text Encoder control when supplied separately.
CLIP-L is not CLIP skip, and it is not interchangeable with T5-XXL simply because both process text.
Diffusion in Low Bits
The Forge model-loading control that chooses an automatic or explicit lower-bit representation for supported diffusion weights.
Automatic preserves the checkpoint’s own precision when possible; this control is not the same as selecting a different checkpoint.
Diffusion model
The model component that performs the iterative denoising work; newer layouts can store it separately from the VAE and text encoders.
UNet and DiT describe different architectures. Treat the publisher’s model layout as the contract instead of calling every diffusion model a UNet.
Embedding
A learned prompt token loaded by the textual-inversion system and used with a compatible base family.
An embedding is not a LoRA or checkpoint; its trigger belongs in the prompt rather than the Checkpoint selector.
FP8 / FP16
Floating-point formats used for model weights or computation; lower precision can reduce memory use but support and behavior depend on the component and hardware.
A precision label is not a model family, and changing precision cannot repair an incompatible VAE, encoder, or LoRA.
GGUF
A supported container format commonly used for quantized diffusion-model or text-encoder weights in split model layouts.
GGUF names a file format, not one precision, architecture, or guarantee that an arbitrary file will load in Forge.
LoRA
A smaller learned modification applied to a compatible base model, usually through an Extra Networks card or prompt token.
A LoRA cannot generate by itself and does not automatically cross SD 1.x, SDXL, or FLUX families.
Model family
A compatible architecture group that determines which checkpoints, LoRAs, text encoders, VAEs, and control models can work together.
Matching file extensions do not make assets from different families compatible.
NF4
A 4-bit BitsAndBytes quantization choice exposed by Forge for supported model-loading paths.
NF4 is not a universal mode for every checkpoint, and an NF4 filename or selector does not guarantee a speed improvement on every GPU.
Quantization
Representing model weights with fewer bits to reduce their storage or memory cost, with quality, speed, and compatibility trade-offs that depend on the method and hardware.
Quantization does not shrink the image dimensions or guarantee faster generation.
T5-XXL
A large text-encoder component required by some FLUX model layouts and selected separately when it is not bundled.
A quantized T5 file changes storage and memory characteristics; it does not become the base diffusion model.
Text encoder
A model component that converts prompt text into conditioning the diffusion model can use.
A text encoder is not the diffusion model or tokenizer. Split layouts may require more than one encoder.
VAE
The component that converts between image pixels and the model’s latent representation, including the final decode into a viewable image.
A VAE does not replace the checkpoint. A missing or wrong VAE can fail at the decode stage even when sampling appears to run.
Generation controls
16 termsBatch count
How many batches Forge runs one after another for the same Generate action.
Batch count multiplies total outputs and time; it is not the number processed simultaneously.
Batch size
How many images Forge asks the model to process together in one batch.
Batch size usually affects peak memory more directly than Batch count and is not the total number of repeated batches.
CFG Scale
The guidance strength applied between positive and unconditional conditioning in the standard CFG path.
CFG Scale is separate from Distilled CFG Scale; setting CFG to 1 disables the negative-conditioning branch in current processing code.
CLIP skip
A compatibility control that chooses how early to stop through supported CLIP text-encoder layers.
CLIP skip is not CLIP-L, and it is not meaningful for every model family.
Denoising strength
In img2img, inpainting, or a high-resolution second pass, the fraction of the denoising process used to transform the starting latent.
It is not ControlNet weight or image opacity; higher values generally permit more departure from the source.
Distilled CFG Scale
A separate guidance value passed to models that expose distilled guidance, including documented FLUX workflows.
It does not mean the same thing as standard CFG Scale. Do not copy one number into both fields by assumption.
Hires. fix
A two-stage txt2img workflow that first generates at one size, upscales, then denoises a higher-resolution pass.
It is not a simple post-process upscale; it adds another diffusion pass, memory load, and its own optional sampler, scheduler, checkpoint, VAE, encoder, prompt, and denoising controls.
Img2img
The workflow that encodes a source image and denoises it toward a new result, controlled strongly by Denoising strength.
Img2img is not inpainting unless a mask limits where the edit is applied.
Negative prompt
Text used for the unconditional or negative side of classifier-free guidance in workflows that use it.
In current Forge processing, CFG Scale 1 skips unconditional conditioning, so the negative prompt has no effect in that mode.
Prompt
Text that describes or conditions what the model should generate, interpreted through the active model’s encoders and training.
A prompt cannot compensate for an incompatible model component or a ControlNet map that encodes the wrong structure.
Sampler
The numerical method that advances the latent through the denoising process.
A sampler is not the noise schedule. Forge exposes Sampler and Schedule type as separate choices.
Sampling steps
The number of denoising steps requested from the sampler for a generation pass.
More steps are not automatically better; useful ranges are model- and sampler-dependent.
Scheduler
The rule that distributes noise levels or timesteps across the requested sampling steps.
A scheduler does not replace the sampler. Automatic lets Forge resolve the schedule associated with the selected sampler.
Seed
The number used to initialize generation noise; keeping it fixed helps isolate the effect of one changed setting.
The same seed is not a cross-version promise when the model, sampler, scheduler, dimensions, software, or other inputs differ.
Txt2img
The workflow that starts generation from text conditioning and an initial latent rather than from a source image.
ControlNet input images can guide a txt2img run, but that does not turn the tab into img2img.
Upscaler
The algorithm or model used to enlarge an image or latent, either between Hires. fix passes or as a post-processing operation.
An upscaler cannot recover ground-truth detail that was never present, and not every upscaler performs a diffusion pass.
Memory & loading
11 termsAsync
The alternative Forge swap method intended to overlap supported weight movement with other work when the environment can do so reliably.
Async is not guaranteed to be faster and can be less suitable on some systems; prove it with the same workload.
GPU Weights
Forge’s megabyte target for how much model-weight data should remain on the GPU when supported swapping is active.
It is not total VRAM use. Setting it near the card’s full capacity can leave too little space for inference and cause fallback, severe slowdown, or an out-of-memory error.
Inference memory
The GPU memory needed while the model actually computes, beyond the space occupied by persistent weights.
A model fitting as weights does not guarantee that the requested image size, batch, attention work, VAE decode, or second pass will fit.
Never OOM
A February 2024 Forge memory feature described as splitting and offloading work to avoid some out-of-memory failures.
This is historical terminology, not a promise that every workload fits or that the same checkbox exists unchanged in the current build.
Offload
Keeping some model components or weights outside dedicated VRAM and moving them when needed to reduce peak GPU memory pressure.
Offload trades memory pressure for transfers and latency; it does not make RAM, bandwidth, or compute unlimited.
Queue
The conservative Forge swap-method choice documented for moving model weights in an ordered path.
Queue here is a memory-transfer strategy, not a list of pending image-generation jobs.
Shared GPU memory
System memory that the operating system can expose to GPU workloads when the platform supports it.
It is slower than dedicated VRAM and should not be added to dedicated VRAM as if both had equal performance.
Swap Location
The Forge control that chooses where supported swapped weights wait outside their active GPU allocation; current labels are CPU and Shared.
The choice changes the memory and transfer path, not the model’s creative behavior.
Swap Method
The Forge selector for the strategy used to move supported weights between GPU memory and the chosen swap location.
Queue and Async are strategies, not quality levels. Hardware and driver behavior determine which is useful.
System RAM
The computer’s main memory, used by Forge, Python, loaded files, offloaded weights, and the operating system.
System RAM is not VRAM. Moving weights out of VRAM can raise RAM use and transfer cost.
VRAM
Memory available to the GPU for loaded weights, active computation, intermediate tensors, and other graphics or system use.
Advertised VRAM is not all available to Forge, and GPU Weights is not a live free-VRAM meter.
Editing & guidance
12 termsAlpha channel
The per-pixel channel that records opacity in an image with transparency.
A checkerboard preview is only a display convention; verify that the saved file retains an alpha channel.
Control Weight
The strength assigned to a ControlNet unit’s influence during its active timestep range.
It is not LoRA weight, denoising strength, or GPU weight memory. More control weight is not automatically more accurate.
ControlNet
An integrated workflow that feeds an additional structural signal—such as pose, edges, or depth—into compatible diffusion generation.
The interface does not bundle every control-model weight, and ControlNet does not by itself preserve identity, clothing, or exact pixels.
ControlNet model
A separate model that applies a specific kind of control signal during diffusion and must match the intended map type and base family.
It is not the preprocessor or base checkpoint. Appearing in the dropdown proves discovery, not compatibility.
Forge Canvas
Forge’s Gradio 4 image canvas for tasks such as drawing masks, panning, zooming, and editing image inputs.
Canvas interaction is an interface feature, not a generation model. Tablet and browser behavior in the announcement is older information.
Inpaint area
The choice between processing the whole image or a crop around the masked region during inpainting.
Only masked changes the working crop and effective detail; it is not the same setting as Mask mode.
Inpainting
Generating a replacement for a masked region while using the surrounding image and selected masked-content settings as context.
Inpainting does not automatically preserve every unmasked pixel; crop mode, mask blur, denoising, model, and resolution still matter.
LayerDiffuse
An integrated Forge feature announced for generating or editing images with transparency-related workflows.
It is not ordinary background removal and requires a compatible workflow; the 2024 announcement is an earlier feature version.
Mask
A grayscale or painted selection that tells the inpainting workflow which region is targeted, according to the active mask mode.
White-versus-black meaning depends on Mask mode; inspect the selection instead of assuming from appearance alone.
Mask blur
The control that softens the mask edge so the generated region can blend into nearby pixels.
It does not blur the whole output and cannot repair a mask placed over the wrong area.
Pixel Perfect
A ControlNet option that calculates preprocessor resolution from input and output geometry.
It does not validate model family, map type, control weight, or the quality of the detected map.
Preprocessor
A ControlNet-stage tool that converts a source image into a control map such as edges, pose, or depth.
The preprocessor creates the signal; the ControlNet model consumes it. Use None when the input is already the required map.
Outputs & records
4 termsGeneration checklist
This guide’s name for the minimum details needed to reproduce a result: build, files, prompt, seed, controls, console outcome, and saved output metadata.
It is a practical testing term, not an official Forge button or file format.
Infotext
Forge’s text record of generation inputs such as prompts, seed, steps, sampler, schedule, guidance, dimensions, and selected model details.
Infotext improves reproducibility but cannot embed every external file, extension version, environment detail, or hidden default.
Latent
The internal compressed representation that the diffusion process denoises before a VAE decodes it into pixels.
A latent preview is not necessarily the final decoded image; a VAE-stage problem can still produce a black or broken output.
PNG Info
The WebUI tab that reads embedded generation information from a supported image and can send recovered settings to another workflow.
A PNG can have metadata stripped by another application, and reading settings does not supply missing checkpoints or extensions.
Try the label from the interface, a shorter word, or All categories. If the term comes from a fork or extension, check that project’s exact repository and version.
A definition tells you what a control is. The task guide tells you what to do next.
Use the glossary to remove ambiguity, then move to the canonical workflow. That page owns the sequence, safe baseline, failure checks, and details checklist.
The confusing parts, answered together.
These answers connect terms that are commonly searched as pairs. They stay inside the behavior confirmed for the current Forge code or carry an earlier-details warning.
What is a checkpoint in Stable Diffusion WebUI Forge?
The checkpoint is the primary model selected for generation. It may bundle several components or, in a supported split layout, work with separately selected VAE and text encoders. It is not a LoRA or ControlNet model.
What is the difference between a checkpoint and a LoRA?
A checkpoint supplies the base generation model. A LoRA is a smaller learned change applied to a compatible base and cannot generate alone. Match its family and documented trigger, then prove its effect with the same seed.
What is the difference between a sampler and scheduler in Forge?
The sampler is the numerical denoising method. The scheduler distributes noise levels or timesteps across its steps. Forge exposes them separately, so a reproducible record includes both.
Why are CFG Scale and Distilled CFG Scale separate?
Standard CFG compares positive and unconditional conditioning. Distilled guidance is a separate value exposed for models trained to accept it. In current Forge, CFG 1 skips unconditional conditioning, so the negative prompt is ignored in that standard path.
What does GPU Weights mean in Forge?
GPU Weights is a megabyte target for supported model weights kept on the GPU during swapping. It is not total VRAM use. The run still needs inference headroom for active computation, image size, VAE decode, and other allocations.
Should GPU Weights equal my graphics card VRAM?
No. Using the full advertised VRAM as a weight target can leave no room for inference and cause fallback, severe slowdown, or an out-of-memory failure. Start from the memory guide’s conservative baseline and change one variable at a time.
What is the difference between Batch size and Batch count?
Batch size is how many images are processed together and usually raises peak memory more directly. Batch count is how many batches run sequentially. Total requested images are batch size multiplied by batch count.
Why does the negative prompt do nothing at CFG 1?
Current Forge processing skips the unconditional-conditioning branch when CFG Scale is 1. Because the negative prompt belongs to that branch, it has no effect in that mode. This does not mean every model should use a higher CFG.
What is the difference between a VAE and text encoder?
A text encoder turns prompt text into conditioning. A VAE converts between image pixels and the latent representation. They perform different jobs and neither replaces the base diffusion model.
What is the difference between a ControlNet preprocessor and model?
The preprocessor derives a map such as pose, edges, or depth from an image. The ControlNet model applies that map during diffusion. The map type, control model, and base-model family must agree.
Does Pixel Perfect make ControlNet compatible automatically?
No. Pixel Perfect calculates preprocessor resolution from geometry. It does not validate the control-model family, map type, active unit, control weight, or detected-map quality.
What does denoising strength change in img2img?
It controls how much of the denoising process is used from the starting image latent. Lower values usually preserve more source structure; higher values permit more change. Compare a short ladder with one fixed seed instead of treating a number as universal.
What is the difference between Hires. fix and an upscaler?
Hires. fix is a two-stage txt2img workflow with an upscale between diffusion passes. A post-process upscaler enlarges an existing image without necessarily running another diffusion pass. The correct route depends on whether you need regenerated detail or only a larger file.
Can PNG Info reproduce an image exactly?
It can recover embedded generation parameters when metadata is present, but exact reproduction can still require the same Forge commit, model files, hashes, extensions, hardware-sensitive paths, and defaults. Metadata may also be stripped after export.
Is Never OOM still a current Forge setting?
Never OOM is documented in a February 2024 announcement. Treat it as historical terminology rather than a guarantee about the current interface. Current memory diagnosis should use the actual controls and console output in your exact commit.