CHECKPOINT · XL PRESET · NATIVE BASELINE · CLEAN DECODE

Run SDXL in Forge from checkpoint to first valid image

Use one documented SDXL checkpoint, the xl preset, and that model’s own resolution and settings. Prove the base can load, sample, decode, and save before adding a VAE, refiner, LoRA, or Hires fix.

01 · Checkpoint contract

Know the model before Forge sees the file

“SDXL” identifies an architecture family. The exact publisher, base or derivative, checkpoint version, recipe, companion files, hash, and license define your runnable model.

  1. IDENTITY

    Publisher page and SDXL base

    Save the original model-card URL and confirm the file is an SDXL base checkpoint or a fine-tune that declares its SDXL upstream.

  2. FILE

    Exact filename, size, and hash

    Compare all values the publisher provides. A plausible filename does not guarantee a complete or authentic download.

  3. RECIPE

    Resolution and generation settings

    Record dimensions, sampler, scheduler, steps, CFG, negative-prompt guidance, triggers, and any required VAE.

  4. TERMS

    Model and derivative licenses

    Forge’s AGPL-3.0 software license does not grant rights to an SDXL checkpoint, fine-tune, LoRA, or its outputs.

02 · Clean baseline

One checkpoint. One plain prompt. No optional load.

This pass answers a narrow question: can this exact SDXL checkpoint complete the full Forge pipeline on this machine?

  1. 01

    Place the primary checkpoint

    Put the finished file in the active models/Stable-diffusion directory. If your install uses path overrides or shared folders, resolve the active path first.

  2. 02

    Refresh and select it under Checkpoint

    Use the top refresh control or restart Forge. Confirm the displayed name matches the file you confirmed.

  3. 03

    Select xl

    The preset changes the visible controls and starting values; it does not convert or repair a mislabeled checkpoint.

  4. 04

    Restore the model-card recipe

    Use the publisher’s documented dimensions, sampler, scheduler, steps, CFG, and prompt guidance. Set batch size to 1.

  5. 05

    Leave optional components off

    No Refiner, Hires fix, LoRA, embedding, ControlNet, face restoration, styles, or unrelated extensions. Leave VAE / Text Encoder in the model’s documented state.

  6. 06

    Generate and watch the console

    Keep the first error if one appears. A successful run must load, encode the prompt, sample, VAE-decode, and save a normal image.

  7. 07

    Save the test

    Keep the PNG, infotext, seed, model filename/hash, Forge commit, launch arguments, console result, and generation time.

03 · XL preset

A useful starting state—not a model card

These values come from the current Forge code. User-saved XL values can replace them, and an individual checkpoint can require a different recipe.

TXT2IMG SIZE896 × 1152

Portrait-oriented current default.

CFG SCALE5

Current txt2img starting value.

SAMPLEREuler a

Use model-card value when specified.

SCHEDULERAutomatic

Forge resolves the current automatic path.

CLIP SKIPHidden

The XL preset hides the regular control.

INFERENCE MEMORYVisible

Keep the default for first acceptance.

04 · Native baseline

Start where the exact checkpoint was documented

Resolution is part of the model recipe, not a decorative export setting. It changes composition, latent workload, memory, and whether the comparison is meaningful.

SDXL MODEL NOTES

Multiple aspect ratios are part of the architecture story

The SDXL paper reports multiple-aspect-ratio training and a second text encoder. This supports treating aspect ratio as intentional, not forcing every job into one square.

Read the primary paper ↗
FORGE UI

896 × 1152 is the current portrait start

The xl preset supplies a usable UI default. It does not override a fine-tune’s publisher recipe.

YOUR CHECKPOINT

Its model card wins the baseline decision

Stability AI describes SDXL 1.0 at native 1024×1024, while its reference code includes multiple 1024-scale aspect ratios. For a fine-tune, use the exact publisher’s documented baseline.

Read the official SDXL 1.0 announcement ↗
Do not diagnose an SDXL fine-tune from a copied SD 1.5 canvas.

If a 512×512 run produces repeated, cropped, or poorly composed subjects, restore the SDXL checkpoint’s documented dimensions before changing the prompt or adding Hires fix.

05 · VAE and Refiner

Bundled first; optional only with details

A valid SDXL base image does not require every control visible in Forge to be filled.

VAE / TEXT ENCODER

Follow the checkpoint package

For a self-contained checkpoint, keep the additional-module selector in its documented automatic or empty state. Select a separate VAE only when the exact model page names it. A random SD 1.5 VAE is not a safe SDXL repair.

Acceptance signalThe final decoded image is normal and saved.
REFINER

Prove the base before the switch

Stability AI documents that SDXL Base 1.0 can be used standalone. Forge’s optional Refiner accordion requests a checkpoint of the same architecture and defaults the switch fraction to 0.8.

Acceptance signalBase succeeds first; refiner is then one measured change.
NOT THE SAME CONTROL

Refiner is not Hires fix

A refiner switches model during sampling. Hires fix adds another generation/upscale pass. Either can add memory, time, and a new failure stage; neither belongs in the clean baseline.

Acceptance signalEach optional pass has its own before/after record.
06 · Baseline checklist

A good first image is reproducible

Check each record as you save it. The checklist distinguishes a proven model from a screenshot that cannot be investigated later.

First SDXL baseline record
RECORDED DETAILS0 / 7 recorded
07 · Failure router

What is the earliest thing that went wrong?

Select the symptom that best matches the run. The route preserves details and avoids changing five settings at once.

Choose the earliest SDXL failure
Discovery failed

Resolve the active checkpoint folder before changing generation settings.

Confirm the finished file is in the active models/Stable-diffusion directory, then use the top checkpoint refresh control or restart Forge. A VAE placed in models/VAE will not appear as a checkpoint.

Resolve the active model folder →
08 · Acceptance

Valid does not mean perfect

Accept the technical pipeline before judging whether the model is creatively suitable for your job.

PASS · LOAD

Correct model identified

The console loads the intended SDXL checkpoint without a family, state-dict, or file-integrity error.

PASS · SAMPLE

Requested settings complete

Batch 1 finishes at the documented baseline with no silent restart or terminal failure.

PASS · DECODE

Normal final pixels

The saved image is not black, grey, tiled, severely stretched, or different from the completed preview because of a decode failure.

PASS · RECORD

Run can be reproduced

The PNG/infotext, seed, checkpoint identity, Forge commit, module state, console result, time, and memory are retained.

Bad hands, illegible text, or a difficult composition are not automatically Forge failures.

The official SDXL Base 1.0 card lists limitations including legible text, compositional tasks, faces, people, and lossy autoencoding. First compare your exact model recipe; then evaluate model capability separately.

Read the publisher’s limitations ↗
09 · SDXL FAQ

Questions behind a first SDXL run

These answers keep checkpoint identity, preset behavior, resolution, decoding, compatibility, memory, and licensing separate.

Can Stable Diffusion WebUI Forge run SDXL?

Yes. The current Forge loader includes separate SDXL base and SDXL refiner engines, and the interface includes an xl preset. That proves architecture support in the updated code version, not compatibility with every conversion or extension.

How do I install an SDXL model in Forge?

Download the exact checkpoint from its publisher page, keep its license and recipe, place the finished primary checkpoint in the active models/Stable-diffusion folder, refresh the Checkpoint list, choose xl, and prove one base-only image.

Where do SDXL checkpoints go in Forge?

The default primary-checkpoint inventory is models/Stable-diffusion. Path arguments and shared-folder settings can change the active directory, so use the folder resolver if the file does not appear.

Which Forge preset should I use for SDXL?

Use xl. In the current commit it exposes SDXL-oriented controls and starts txt2img at 896×1152, CFG 5, Euler a, and Automatic scheduling. Those are interface defaults, not a universal recipe for every SDXL fine-tune.

What resolution should I use for SDXL in Forge?

Use the exact checkpoint publisher’s documented baseline. SDXL was designed and trained with multiple aspect ratios; one family label does not make every fine-tune share one safe canvas. For Stability AI SDXL Base 1.0, 1024-scale generation is the reference context, while Forge’s current xl preset opens at 896×1152.

Should I generate SDXL at 512×512?

Do not use an inherited SD 1.5 canvas as the acceptance test unless the exact SDXL model card requests it. A smaller image may render, but it does not guarantee the model at its intended baseline and can change composition and quality.

Why does Forge open SDXL at 896×1152 instead of 1024×1024?

That is the current xl preset’s portrait-oriented txt2img default in the current code. It is a UI starting value. Square 1024-scale examples and other aspect ratios can also be valid when the exact model documentation specifies them.

Do I need a VAE for SDXL in Forge?

Only when the exact checkpoint publisher requires a separate VAE or identifies a specific variant. Leave VAE / Text Encoder at its documented automatic or empty state for a self-contained checkpoint; do not attach a random VAE as a generic quality upgrade.

What does VAE / Text Encoder mean for an SDXL checkpoint?

It is Forge’s additional-module selector. A normal self-contained SDXL checkpoint supplies the components its package promises; the current SDXL engine uses two text encoders and a VAE internally. Use an external selection only when the checkpoint documentation calls for it.

Does SDXL require a refiner in Forge?

No. Stability AI’s SDXL Base 1.0 card explicitly says the base model can run standalone. Forge offers an optional Refiner accordion that switches to a checkpoint of the same architecture at a chosen fraction of sampling. Prove the base first.

What should “Switch at 0.8” mean in Forge Refiner?

In the current Refiner control, 0.8 is the default fraction of sampling steps at which Forge switches models; 1 means never and 0.5 means halfway. Use a same-architecture refiner and the workflow’s documented value, not 0.8 as a universal SDXL rule.

Why is my SDXL checkpoint not showing in Forge?

The usual discovery checks are active path, completed filename, supported checkpoint extension, and refresh/restart. Also confirm you did not place a VAE, LoRA, or refiner in the wrong inventory and expect it to behave as the base checkpoint.

Why does an SDXL model fail to load in Forge?

Read the first console error. Verify that the file is an SDXL base or declared SDXL fine-tune, not a refiner-only or another-family file; compare the published size/hash; and retest without extensions or forced precision flags.

Why is my SDXL image black in Forge?

If sampling previewed an image but the saved result is black or grey, test the decode contract: restore the documented bundled/automatic VAE or select only the exact required SDXL VAE, remove extensions, and rerun the same seed.

Why do SDXL images look stretched, duplicated, or distorted?

First restore the model publisher’s dimensions, xl preset, batch size 1 and base-only workflow. Wrong-family add-ons, an inherited SD 1.5 resolution, Hires fix, or a mismatched VAE can confound the result. Then distinguish pipeline distortion from normal model limitations.

What sampler, scheduler, steps, and CFG should I use for SDXL?

Use the exact checkpoint card’s recipe. Forge’s current xl preset starts at CFG 5, Euler a, and Automatic, but a fine-tune may document different values. Record settings in the first accepted image instead of treating one internet recipe as universal.

Can I use an SD 1.5 LoRA or ControlNet model with SDXL?

No. Match LoRAs, embeddings and control models to the declared SDXL base family and supported version. The shared.safetensors extension is a container, not cross-family compatibility.

Can I add Hires fix to the first SDXL test?

Keep it off for acceptance. First prove that the checkpoint loads, samples and decodes at its documented baseline. Add Hires fix later as a separate workload with its own memory and quality comparison.

How much VRAM does SDXL need in Forge?

There is no universal threshold supported by this page. GPU, precision, Forge commit, dimensions, batch, VAE, refiner and add-ons all change the result. Test batch 1 on your machine and record the failing stage and measured memory.

Can I use an SDXL model commercially?

The SDXL family name and Forge software license are not legal clearance. Read the exact base model, fine-tune, conversion and add-on licenses for your intended use.