FORGEFIELD GUIDE
LOCAL PROCESS · ONE MODEL · CLEAN BASELINE

Your first run in Stable Diffusion WebUI Forge

If the console stays open, prints a local URL, and the Forge interface loads, installation has succeeded. Add one compatible model, select its matching UI preset, keep optional tools off, and generate one baseline image before changing anything else.

Start the first-run check Learn the interfaceGenerate the first image
Evidence checked24 Aug 2026Code snapshotdfdcbab · 26 Jun 2025
01 · Baseline gate

Four checks before Generate

This checklist does not configure Forge. It keeps the first test small enough that a failure has one traceable cause.

First-run readiness checklist
BASELINE READINESS0 of 4 ready
02 · Console and browser

The browser is a client. The console is the process.

Forge runs on your machine and serves its interface to a local browser address. Closing the tab does not uninstall anything; closing the running console normally stops the local server.

Inspect the current launch code ↗
FIRST-RUN SIGNALS
  1. 01
    Environment preparesChecking and loading dependencies…
  2. 02
    WebUI startsRunning on local URL: http://127.0.0.1:[port]
  3. 03
    Model loads on demandModel selected → Loading Model → Model loaded
  4. 04
    Generation completesProgress reaches 100% without a traceback

Illustrative wording only. Use the exact address and messages printed by your build.

IF THE TAB DID NOT OPEN

Leave the console running and open the exact local URL it printed. Browser autolaunch and server startup are separate events.

IF LOCALHOST REFUSES

Return to the console. A closed process, startup traceback, different port, or blocked local process must be resolved before browser troubleshooting.

IF THE BROWSER CLOSES

Reopen the printed URL while the console is still alive. Do not add --listen or --share merely to repair a local tab.

03 · Model boundary

Forge is the workbench, not the model

A correct Forge installation can open with no usable checkpoint. Model files come from their own publisher, carry their own license, and must match the workflow you intend to run.

FORGE PROVIDES

The interface and runtime

  • txt2img, img2img, settings, output and integrated controls.
  • Model discovery, loading and memory-management paths.
  • A local Gradio-based browser interface.
YOU PROVIDE

Compatible learned weights

  • A checkpoint or supported model layout from a source you trust.
  • Any required VAE or text encoders for that model family.
  • Enough storage, RAM and GPU memory for that exact workload.

  1. Choose one model family

    For a first baseline, a single-file SD or SDXL checkpoint is simpler than a split model with separate text encoders. If you choose FLUX, follow the file requirements published for that exact variant rather than treating it like an SDXL checkpoint.

  2. Check the source and license

    Do not download executables presented as models. Record the model page, filename, version, hash when provided, and license. Forge’s AGPL license does not replace the model’s terms.

  3. Place the file in the matching directory

    The paths below are relative to Forge’s data directory. The Windows package normally puts that data inside its extracted webui folder; Git installs normally use the repository folder.

  4. Refresh, then select

    Use the refresh control beside the top model selectors or restart Forge. Seeing the filename in the Checkpoint list proves discovery; a successful generation proves the complete model setup.

Default relative paths in inspected original Forge code
WhatDefault relative pathFirst-run useSource
Checkpoint / primary modelmodels/Stable-diffusionRecognizes supported .safetensors, .ckpt, and .gguf files.sd_models.py ↗
Separate VAEmodels/VAEUse only when the model instructions require a separate VAE.sd_vae.py ↗
Text encodermodels/text_encoderUsed by model layouts that require separate CLIP or T5 components.main_entry.py ↗
txt2img outputoutputs/txt2img-imagesDefault location for automatically saved txt2img samples.shared_options.py ↗

Custom --data-dir, --models-dir, model-directory arguments, shared A1111 paths, or saving settings override these defaults. Inspect your launch arguments and Settings before assuming a file is missing.

04 · First image

Read the interface from model to output

The exact values depend on the model. The safe first test is the matching preset’s defaults, one image, no high-resolution pass, and no optional control pipeline.

Ready to generate?

If the screen still feels unfamiliar, use the code-derived interface map first. Then follow the dedicated first-image procedure for the exact control order, model-family decision, output paths, seed record, and failure checks.

Learn the six interface zones →
Generate your first Forge image →
VERIFIED CONTROL MAPORIGINAL REPOSITORY CODE · dfdcbab
UIsdxlfluxall
CHECKPOINTModel selection⌄
VAE / TEXT ENCODERAutomatic⌄
DIFFUSION IN LOW BITSAutomatic⌄
PROMPTDescribe the image to generate…
GPU WEIGHTS (MB)AUTO
SWAP METHODQueue / AsyncSWAP LOCATIONCPU / Shared
LOCAL OUTPUT
PRODUCT CAPTURE
NOT YET AVAILABLE
Explanatory map built from verified labels in original Forge source ↗. It is not a product screenshot. A current, rights-cleared capture is not available in this workspace.
  1. 01UI preset

    Choose sd, xl, or flux to match the model family. The preset changes visible controls and starting values; it does not convert the model.

  2. 02Checkpoint

    Select the one model you added. If loading is deferred, the first Generate click may take longer while Forge loads its weights.

  3. 03VAE / Text Encoder

    Leave the built-in or automatic choice unless the model publisher names separate required files.

  4. 04txt2img prompt

    Use a plain test such as a ceramic teapot on a wooden table, soft window light. Leave model-specific trigger words and LoRAs out of the baseline.

  5. 05Settings

    Keep the preset defaults, batch size 1, and optional accordions off. Use a fixed seed after the first result when you need a controlled comparison.

  6. 06Generate

    Watch both browser progress and console output. A preview alone is not enough if the console ends with a traceback.

Starting values returned by preset logic in inspected commit dfdcbab
PresetIntended starting familytxt2img sizeGuidance valuesSampler · scheduler
sdSD 1.x-style checkpoint512 × 640CFG 7Euler a · Automatic
xlSDXL-style checkpoint896 × 1152CFG 5Euler a · Automatic
fluxCompatible FLUX layout896 × 1152CFG 1 · Distilled CFG 3.5Euler · Simple
These are code defaults, not universal quality settings.

A model publisher may specify a different resolution, sampler, scheduler, CFG behavior, VAE, or text encoder. Follow that model’s documentation after the clean preset baseline.

Inspect the preset logic ↗
05 · Done state

A useful baseline is more than a picture

The first result becomes diagnostic evidence only when you can identify how it was produced and repeat the same test after one change.

01 · PROCESS

Console stayed alive

No fatal traceback followed model loading or generation.

02 · MODEL

One model loaded

The console and result metadata identify the selected checkpoint and modules.

03 · OUTPUT

An image was saved

The gallery result exists in the configured txt2img output directory.

04 · RECORD

Parameters are recoverable

Prompt, seed, size, sampler, scheduler, model hash, and code state are recorded.

KEEP WITH THE TEST

Your minimal reproduction record

Forge
Repository + commit
Environment
OS + GPU + driver
Model
Filename + hash + VAE/TE
Generation
Prompt + seed + size + sampler + scheduler + steps + CFG
State
Preset + launch arguments + extensions
Evidence
PNG metadata + complete console log

The inspected code saves generated samples as PNG by default and stores generation parameters with the result. Output paths and metadata settings are configurable. A fixed seed alone does not guarantee identical output across different model hashes, VAEs, schedulers, commits, or environments.

See why update-era reproduction needs the full state ↗
06 · Failure router

Match the symptom before changing settings

Start from the earliest failed state. Random package upgrades, command-line flags, and extension changes can replace the original error with a harder one.

The UI opens, but the Checkpoint list is empty

Confirm the file is under the active models/Stable-diffusion path or a configured --ckpt-dir. Verify the extension, click the model refresh control, and restart once. If the model belongs to a split layout, confirm you did not place its text encoder or VAE in the checkpoint folder.

The checkpoint appears, but the first Generate click is slow

This can be expected while Forge loads the selected model. The maintainer explains that loading may be deferred until generation begins because large models can take significant time to load. Watch for continuing console activity instead of repeatedly clicking Generate.

Read the model-loading explanation ↗
Localhost refused the connection

The local server is unavailable at that address. Check whether the console exited, whether startup printed a different port, and whether security software blocked the Python process. Restart from the launcher and use the exact URL printed after a successful launch.

Diagnose the process, port, and browser boundary →
The browser says “Connection errored out”

Treat the browser message as a symptom and read the console. The official dated troubleshooting note groups this with resource pressure, system swap, corrupted models, and environment failures. Preserve the first traceback before changing pagefile settings or reinstalling packages.

Route the exact connection state →Open official troubleshooting #1474 ↗
The output is black, distorted, or unrelated to the prompt

First check the model family against the selected UI preset, then the required VAE and text encoders. Remove LoRAs, ControlNet, Hires. fix, scripts, and extensions. Reproduce with the simple baseline before changing precision or memory flags.

Generation runs out of memory

Return to batch size 1, the preset’s starting resolution, and no optional tools. Record the model layout and complete console log. Do not assume another user’s GPU Weight, swap, or low-bit configuration applies to your model and hardware.

The image finished but disappeared from the browser

Check the configured outputs/txt2img-images directory. Browser delivery and local file saving are separate parts of the run, so a UI interruption does not prove the image was never written.

07 · After the baseline

Add capability one variable at a time

A working generation is your rollback point. Keep it intact while you add models, LoRAs, extensions, shared folders, or advanced memory controls.

  1. 01

    Save the known-good state

    Keep the baseline PNG, metadata, console log, model hash, and current Forge commit.

  2. 02

    Add one asset or tool

    Introduce one LoRA, VAE, model, or extension—not several categories together.

  3. 03

    Repeat the same generation

    Use the same seed and full parameter set. Compare the console and output before moving on.

  4. 04

    Keep third-party ownership visible

    Extensions have separate repositories, update schedules, code permissions, and compatibility boundaries.

Extension compatibility is dated.

The original repository’s replacement list is explicitly temporary and includes community submissions. Confirm the repository, supported Forge state, and recent issue history before installing extension code.

Inspect the temporary extension list ↗
Compare workflow options Check original Forge status
08 · First-run FAQ

Questions that belong to the first run

These answers cover the boundary between a successful installation and a stable first generation.

Does Forge download a Stable Diffusion model automatically?

No model is promised with the Forge application. You obtain a compatible model separately, review its source and license, place its files in the correct directories, then refresh the model selectors.

Does localhost mean my images are generated in the cloud?

No. Localhost points the browser back to a server on your own machine. Initial setup, model tools, extensions, or explicit sharing options can still make network requests, so “local interface” does not mean every installed component is permanently offline.

Can I close the Forge browser tab?

Yes, and you can reopen the exact local URL while the console process remains active. Avoid closing or restarting the browser during the first diagnostic generation because it makes the run harder to observe.

Where do Forge-generated images go?

The inspected default for txt2img samples is outputs/txt2img-images under the active data directory. Settings and launch arguments can change it; Settings → Paths for saving shows the effective configurable destinations.

Why is the first image slower than the next one?

Forge may defer model loading until the first generation request. Loading weights and moving model components into the selected memory path adds work before sampling begins. Compare later runs only after the same model is already loaded.

Should I start with SD, XL, or FLUX?

Start with the model family for which you have a complete, documented file set. A single-file SD or SDXL checkpoint is usually simpler for validating the UI than a split FLUX layout. This is a troubleshooting choice, not a claim that one family is universally better.

Why does the same seed produce a different image?

A seed is only one input. Model and VAE hashes, scheduler, sampler, dimensions, precision, prompt parsing, extensions, launch arguments, Forge commit, and sometimes the environment can all affect reproduction. Compare the complete generation record.

Should I install extensions before making the first image?

No. First prove that original Forge, one model, and the preset defaults can complete a generation. Then add one extension and repeat the same baseline so you can identify what changed.