FORGEFIELD GUIDE
TXT2IMG · ONE MODEL · ONE REPEATABLE RESULT

Generate your first image in Forge

Open txt2img, match the UI preset to your model family, select the checkpoint, then use the settings documented for that exact model. Generate one image with optional tools off and keep the result’s seed and metadata.

Start with the 7-step path Choose your model family
Evidence checked24 Aug 2026Code snapshotdfdcbab · 26 Jun 2025ScopeOriginal Forge only
01 · Direct path

Seven controls from blank UI to saved PNG

Keep the first run deliberately plain. Its job is to prove that the model, interface, generation pipeline, and output path work together.

  1. 01
    Open txt2img

    Use the text-to-image tab. Do not start in img2img, Extras, or an extension panel.

  2. 02
    Match UI preset → model family → checkpoint

    Choose sd, xl, or flux, then select the compatible model. Add separate VAE or text encoders only when its publisher requires them.

  3. 03
    Write one literal prompt

    Try: a ceramic teapot on a wooden table, soft window light. Leave Negative prompt empty unless the model documentation says otherwise.

  4. 04
    Copy the model’s documented baseline

    Use its recommended resolution, sampler, scheduler, steps, and guidance. Set Batch count and Batch size to 1. Keep Hires. fix, LoRAs, ControlNet, scripts, and extensions out of the test.

  5. 05
    Start with seed -1

    This requests a random seed. After the result appears, keep the concrete seed stored in its generation record. Verify the seed controls ↗

  6. 06
    Select Generate once

    Watch the browser and console. The first request may load the model before sampling, so it can take longer than the next run.

  7. 07
    Prove and preserve the result

    Confirm the image is in the gallery, the console has no fatal traceback, the PNG exists in the effective output folder, and its metadata records the run.

FORGE SHORTCUTS

Ctrl + Enter generates, Alt + Enter skips, and Esc interrupts in the inspected UI code.

Verify in ui_toprow.py ↗
02 · Model contract

Choose the family before the numbers

Forge provides controls, not one universal recipe. Select what your model page says it is; this router tells you what to verify next.

FORGE UI PRESETsd
GENERATION CONTRACT

Select the SD-family checkpoint. Use the resolution and generation settings documented for that exact model.

FILE CHECK

Usually one checkpoint; add a separate VAE only when the model publisher requires it.

VERIFIED The current original Forge UI exposes sd, xl, flux, and all presets plus Checkpoint, VAE, and Text Encoder selectors. Inspect the source ↗

03 · Before Generate

Build a baseline, not a showcase

A first result is useful when it changes one question—“does this documented model setup work?”—into evidence.

Baseline receipt
READY SIGNALS0 / 6 recorded
Do not copy “best settings” across model families.

The visible controls are shared, but their useful ranges are not. Even two fine-tunes from the same family can publish different baselines.

Inspect sampler controls ↗
04 · Read the UI

What each first-run control proves

The interface is easier to debug when you read it as a signal chain rather than a wall of settings.

01 · PRESET

Shows the right control surface

Match it to the family. It is not a model downloader or converter.

02 · CHECKPOINT

Chooses the learned weights

The filename appearing proves discovery. A completed image proves much more.

03 · PROMPTS

Describe the test image

Plain language makes failures easier to separate from special syntax.

04 · SAMPLING

Defines the model-specific run

Resolution, sampler, scheduler, steps, and guidance belong to the model contract.

05 · SEED

Controls the starting noise

-1 requests random. Reuse the recorded concrete seed for the next comparison.

06 · OUTPUT

Holds image and run record

Gallery, automatic file, and embedded infotext form the evidence set. Inspect PNG saving ↗

05 · Output map

Gallery, automatic file, and Save are different

The browser gallery is the visible result. The filesystem location is controlled by Forge’s saving settings.

AUTO-SAVEoutputs/txt2img-images/[date]/

The inspected defaults save generated samples and organize them into date subdirectories. Custom data directories and Settings can change this path.

shared_options.py ↗
FOLDER BUTTONOpen effective txt2img folder

The folder control beside the gallery asks Forge to open the configured sample directory. Use it instead of guessing where this installation writes.

ui_common.py ↗
GALLERY SAVElog/images

Save copies the selected gallery result to the dedicated save directory; Save zip handles selected files as an archive. This is separate from automatic sample saving.

ui_common.py ↗
PNG INFORecover generation parameters

With PNG infotext enabled, the file can carry the prompt and generation record. PNG Info reads it and can send settings into another workflow.

infotext_utils.py ↗

VERIFIED against original Forge code at dfdcbab. Defaults describe a clean inspected configuration, not every existing user profile.

06 · Done state

One image, four pieces of proof

A picture in the gallery is encouraging. A repeatable baseline also survives outside the browser session.

1

Gallery

The final image appears, not only an intermediate preview.

2

Console

Progress completes without a fatal traceback after model load.

3

File

The PNG exists under the effective txt2img output setting.

4

Record

Model, prompt, concrete seed, size, sampler, scheduler, steps, guidance, and Forge state are recoverable.

SEED -1random request
RESULT SEEDkeep this number
NEXT RUNchange one thing

A fixed seed is necessary for a controlled comparison, but it is not the whole state. Model and VAE hashes, dimensions, sampler, scheduler, precision, extensions, Forge commit, and environment can still affect the result.

07 · Failure router

Fix the earliest failed signal

Do not compensate for a loading or file-path failure by changing prompt settings.

The Checkpoint selector is empty

Verify the active model directory, file type, and any custom --data-dir, --models-dir, or checkpoint-directory argument. Use the refresh control, then restart once. If the distribution is split, make sure a text encoder was not mistaken for the checkpoint.

See the reported empty-list intent ↗
The checkpoint appears but does not load

Read the first console error. Recheck the model family, required VAE and text encoders, available memory, filename integrity, and the publisher’s instructions. Discovery in a dropdown does not prove compatibility.

Generate is slow before sampling begins

The first request can include loading the selected model. Wait for the console to distinguish loading from sampling. Compare generation times only after the same model is already loaded.

The output is black, blank, or corrupted

Stop the baseline. Save the full console log and generation metadata, confirm the exact model components and model-native settings, and remove optional extensions. Black-output reports exist, but there is no universal setting change that safely fixes every cause.

Find the first bad output stage →Community-reported example ↗
The image is in the gallery but not in the folder I expected

Use the gallery folder button and inspect Settings → Paths for saving. Check the active data directory and whether automatic saving of generated images was disabled. Do not confuse the Save destination with the automatic txt2img path.

The same seed does not reproduce the image

Compare the complete infotext and runtime state, not only the seed: model/VAE hashes, prompts, dimensions, sampler, scheduler, steps, guidance, precision, extensions, launch arguments, Forge commit, and environment.

Control the next txt2img run Diagnose unusable outputCheck the model boundary
08 · First-image FAQ

Questions before and after Generate

These answers cover the practical decisions that stop a first image from becoming a repeatable baseline.

Which Forge tab makes the first image?

Use txt2img. It creates an image from a text prompt without requiring an input image.

Which Forge UI preset should I choose?

Choose sd, xl, or flux to match the loaded model family. The preset changes the interface and starting values; it does not convert an incompatible model.

Does Forge include a Stable Diffusion model?

Do not assume that it does. Forge is the interface and runtime. Obtain the model from its publisher, review its license, and provide every file that model requires.

Why is the Checkpoint list empty?

Confirm that the model is in the active model directory, use the refresh control beside the selector, and check whether custom data or checkpoint-directory arguments changed the path.

What prompt should I use for the first image?

Use one plain, literal scene without LoRAs, embeddings, ControlNet, or model-specific trigger words—for example: a ceramic teapot on a wooden table, soft window light.

Do I need a negative prompt?

Not for the baseline unless the model publisher explicitly requires one. Leaving it empty removes another variable from the first test.

Which sampler, steps, CFG, and scheduler should I use?

Use the baseline documented for the exact model. Forge exposes those controls, but no single numeric recipe is correct for SD, SDXL, FLUX, and every fine-tune.

What resolution should I use in Forge?

Use the model publisher’s documented native or recommended resolution. An arbitrary size can waste memory or produce misleading quality problems.

What seed should I use for the first image?

Seed -1 asks Forge for a random seed. After the first successful image, reuse the actual recorded seed for a controlled comparison.

Why does seed -1 show a different number in the result?

The -1 value is the request for a random seed; the result metadata records the concrete seed that was used. Keep that concrete value.

Why is the first generation slower?

The first request can include model loading before sampling. Watch the console: loading time and generation time are different stages.

Where are Forge txt2img images saved?

With the inspected default settings, samples go under outputs/txt2img-images in the active data directory, with date subdirectories enabled. Your saving settings and launch arguments can override this.

Is the Save button the same as automatic saving?

No. Generated samples are automatically saved when that setting is enabled. The gallery Save button copies the selected result to the configured save directory, which defaults to log/images.

How do I recover settings from a generated PNG?

When PNG metadata saving is enabled, use the PNG Info tab to read the embedded generation parameters and send them back to another workflow.

Why did Forge generate a black or blank image?

Treat it as a failed baseline, not a prompt problem. Keep the PNG metadata and console log, confirm the model family and required components, then troubleshoot the earliest warning or error.

How do I know the first generation really succeeded?

The gallery shows the image, the console completes without a fatal traceback, the file exists in the effective output path, and its generation record identifies the model and settings.

09 · Evidence note

What we verified, and what remains model-specific

Labels keep interface facts separate from user reports and model-publisher decisions.

VERIFIED

UI labels, shortcuts, sampling and seed controls, automatic saving, output buttons, and metadata behavior are traced to current original Forge code.

COMMUNITY-REPORTED

Empty checkpoint lists and black outputs are real search intents in original Forge issues; symptoms do not establish one universal cause.

MODEL-SPECIFIC

Resolution, sampler, scheduler, steps, guidance, VAE, text encoders, and prompt conventions must come from the exact model publisher.

Written by the Forge Field Guide editorial team · technically reviewed against commit dfdcbab · evidence checked 24 Aug 2026.