Compare local image UIs

Forge vs Fooocus: choose your control budget

Fooocus prepares an SDXL path around prompts and images. Forge exposes more of the model, sampler, conditioning and memory recipe. Choose the decisions you need—not the interface with the stronger slogan.

Evidence checked31 Aug 2026Original Forgedfdcbab · 26 Jun 2025Original Fooocusae05379 · 2 Sep 2025GPU comparisonNot performed
FOOOCUSPromptStyleImagePrepared SDXL route
YOUR JOBWhat must you decide?
FORGEModelSamplerMemoryVisible generation recipe
SHORT ANSWER

Choose Fooocus for guided SDXL. Choose Forge for explicit control.

Start with Fooocus when the job fits SDXL and you want prepared defaults, prompt expansion, styles, image prompts and guided edits. Start with original Forge when you must select model families or files, tune the sampler/conditioning chain, manage memory behavior or use a named Forge extension.

Fooocus is not “a toy with no settings,” and Forge is not automatically faster or better. They are separate applications from the same creator. The current original Fooocus scope is limited LTS around SDXL; the current original Forge snapshot is also dated and must not be confused with its continuation forks.

01 · DECISION ROUTER

Name the decision the UI must carry.

Select the constraint that would make the other tool fail the job. The route is a candidate to test, not a permanent identity.

Choose your main Forge or Fooocus requirement
ROUTE TO TESTTest original Fooocus first
VERIFIEDTest original Fooocus first

Let the application choose more of the starting recipe.

Fooocus is deliberately built around SDXL, prompt expansion, styles and prepared presets. Its official Windows path also downloads the selected default model on first launch.

Acceptance test
Generate from a short prompt, revise it once and use one built-in image task. Note every control you had to open before the result became usable.
VERIFIEDTest original Forge first

Keep the generation and resource decisions visible.

Forge exposes an A1111-shaped form with checkpoint, VAE/text encoder, sampler, scheduler, steps, guidance, dimensions and Forge-specific model-loading controls.

Acceptance test
Record one complete recipe, restart the UI and prove that you can restore the same named settings without relying on remembered defaults.
VERIFIEDTrial Fooocus on that edit

Use the guided image-input path as the test.

Fooocus bundles Vary, Upscale, Image Prompt, Inpaint or Outpaint and Enhance workflows around its SDXL scope. This is more substantial than a one-button generator.

Acceptance test
Import one real image, complete the edit, then repeat it after restart. Count hidden downloads and decisions exposed only after opening Advanced.
VERIFIEDTest original Forge first

Prefer the control surface that exposes the chain.

Forge integrates a ControlNet-style unit workflow with model, preprocessor and guidance controls. Fooocus offers guided Image Prompt types, but that is not the same operating surface.

Acceptance test
Use the exact conditioning model and preprocessor your job requires. Verify weight, start/end guidance, input dimensions and the saved metadata.
VERIFIEDVerify Forge model support

Resolve architecture before interface preference.

Original Fooocus declares an SDXL-only limited-LTS scope and says it has no current plan to add newer model architectures. Original Forge documents SD, XL and FLUX presets plus FLUX-specific formats and controls.

Acceptance test
Name the exact architecture, files and Forge commit. A family label is not enough to prove that every variant, LoRA or ControlNet combination works.
ASSUMPTIONKeep both; run one receipt

Measure the complete session before replacing anything.

Install them separately and use the same supported SDXL checkpoint for a workflow comparison. Keep default behavior as a second Fooocus trial rather than mixing “same settings” and “best defaults.”

Acceptance test
Complete the ten-record receipt below. Choose only if one tool removes a repeated cost without breaking a required model, edit or recovery path.
02 · IDENTITY AND STATE

Same creator. Separate products. Dated scopes.

A commit date is only useful beside what changed. We pin owner, repository and product scope before comparing features.

ORIGINAL FORGElllyasviel/stable-diffusion-webui-forge

WebUI-shaped platform with Forge-specific resource and model work.

ORIGINAL FOOOCUSlllyasviel/Fooocus

SDXL-centered prompt-and-image application in limited LTS.

VERIFIED Fooocus main is dated later than Forge main, but its latest change is CI-only and its README still declares SDXL-only limited LTS. Do not convert chronology into a winner score.

03 · TASK MATRIX

Compare what each surface asks you to decide.

These rows describe inspected original repositories. They do not transfer claims from Fooocus forks or Forge continuations.

ORIGINAL REPOSITORIES · CHECKED 31 AUG 2026
BoundaryOriginal ForgeOriginal FooocusYour test
Product thesisExpose a familiar WebUI form plus Forge-specific model and resource controls.Automate a prepared SDXL configuration so the user can focus on prompts and images.Which decisions do you want to own?
Declared model scopeUI presets include sd, xl, flux and all; exact architecture support remains commit-specific.Official README says the project is built entirely on SDXL and has no current plan for newer architectures.Name the model family first.
First model acquisitionThe application package does not give you a general model library; add the required files deliberately.The official launcher downloads preset-specific default models on first run.Do you want a prepared start or explicit files?
Default decision surfaceDense tabs, selectors, sliders, accordions and optional extensions.Prompt and image first; Advanced reveals styles, models, LoRAs and sampling controls.Count decisions before the first useful output.
Prompt treatmentPrompt, negative prompt, styles and weights stay visible in the WebUI-shaped form.Fooocus V2 can expand prompts offline; styles and multi-line prompts shape the prepared recipe.Do you need literal prompt control or guided expansion?
Sampling controlsSampler, scheduler, steps, guidance, dimensions, batch and high-resolution pass controls are exposed.Sampler, scheduler, guidance and sharpness exist under Advanced, while defaults are intentionally automated.Fooocus is reduced-decision, not control-free.
Image-guided workIntegrated ControlNet/IP-Adapter routes and img2img controls; compatibility is model- and commit-specific.Image Prompt provides built-in types including ImagePrompt, PyraCanny, CPDS and FaceSwap in inspected source.Are named components or a guided operation more important?
Inpaint / outpaintForge Canvas and the img2img/inpaint form expose mask and generation controls.Dedicated Inpaint or Outpaint plus directional outpaint and a separate Enhance workflow.Run the exact mask–revise–export loop.
Model files and formatsCheckpoint, VAE/text encoder, LoRA and GGUF/low-bit routes are exposed for supported families.Base model, optional refiner, VAE and LoRAs are selectable within the SDXL-centered design.Visibility does not prove cross-tool parity.
Memory decisionsForge exposes precision, GPU Weights/inference headroom and swap-related choices for supported paths.Runtime flags exist, but the main UI is designed to hide routine tuning behind prepared behavior.Measure peak VRAM and system spill locally.
Third-party capabilityExtensions tab plus built-ins; the official replacement list is dated and not a compatibility guarantee.The original project emphasizes built-in workflows and links users to separate forks for experimentation.Inventory every dependency before moving.
Saved workImages, generation parameters, styles, settings and extension-specific artifacts.Images, metadata/load-parameters behavior, presets and configuration; not a Forge settings import.What must reopen after restart?
Project stateMain feature commit inspected from June 2025; original repo must be separated from newer Forge forks.Limited LTS, SDXL-only stated scope; release v2.5.5 is from August 2024.A later commit date is not a feature score.
04 · CONTROL BUDGET

“Simple” means defaults move upstream.

Fooocus still has advanced controls. The difference is which choices lead the session and which wait behind a prepared route.

FOOOCUS AUTOMATES FIRST

A prepared SDXL recipe

  1. 01Default model download by preset
  2. 02Offline prompt expansion / Fooocus V2 style
  3. 03Prepared Speed or Quality behavior
  4. 04Guided Vary, Upscale and Image Prompt routes

Open Advanced when you need base/refiner selection, LoRAs, negative prompt, sampler, scheduler, guidance or sharpness.

FORGE EXPOSES FIRST

A visible generation recipe

  1. 01Checkpoint and companion model files
  2. 02Sampler, scheduler, steps and guidance
  3. 03Resolution, batch and high-resolution pass
  4. 04Precision, inference headroom and swap choices

Presets reduce clutter, but the operator still owns more of the model and runtime configuration.

05 · 20-MINUTE TRIAL

Test the workflow twice, not the slogan once.

Use one SDXL checkpoint supported by both tools for the control-surface trial. Then run Fooocus with its intended defaults as a separate experience. Do not mix those two questions.

  1. 00–04
    Pin the candidates

    Record both repositories, versions and runtime. Keep installs and outputs separate.

  2. 04–10
    Prompt and revise

    Generate from one short prompt, lock a seed policy, then make one intentional revision.

  3. 10–16
    Edit one image

    Run the real vary, inpaint, outpaint, image-prompt or ControlNet-shaped task you repeat.

  4. 16–20
    Restart and recover

    Reopen the job from recorded artifacts. Count manual reconstruction and hidden dependencies.

TRIAL RECEIPT0 / 10 recorded

Record every field before choosing.

06 · SWITCHING LEDGER

Move files. Rebuild behavior.

The applications can see some of the same assets, but their defaults, prompt treatment and extension surfaces do not become equivalent.

Plan shared model storage
01

SDXL checkpoints

Often reusable as files

Both applications must support the exact checkpoint and companion VAE; compare hashes, not filenames.

02

LoRAs

Potentially reusable

Match SDXL base, trigger words, strength and loader behavior. Retest each one.

03

Prompts

Copyable text

Fooocus prompt expansion, styles and defaults can change the effective conditioning. Text alone is not parity.

04

Styles and presets

Rebuild deliberately

Fooocus presets and Forge style rows are different configuration artifacts. Do not rename one as the other.

05

Input and output images

Keep originals

Preserve source images, masks and unedited outputs outside either application folder.

06

Fooocus configuration

Does not import into Forge

Record preset, config values and Advanced selections; map only the settings required for the task.

07

Forge extensions

Do not transfer

Find a built-in Fooocus route or accept that the task remains in Forge. A fork is a separate product decision.

08

Output folders

Keep separate during trial

Separate trees prevent mistaken comparisons and make rollback auditable.

09

Working installation

Do not overwrite

A migration is complete only after the candidate survives restart and repeats the real job.

07 · CLAIM FILTER

Six shortcuts that break the comparison.

We publish a claim only at the level the evidence supports. Preferences remain preferences; dated code remains dated code.

STALE SNAPSHOT

“Fooocus is abandoned.”

The current official description is limited LTS with bug fixes only. That is a narrow scope, not the same claim as an archived repository.

UNKNOWN

“Fooocus makes better images.”

The interface changes defaults, styles and prompt treatment. Without a controlled chain, a preferred output does not establish universal quality.

UNKNOWN

“Forge is faster and uses less VRAM.”

Hardware, model, precision, dimensions, batch, versions and memory settings must be held constant and reported.

VERIFIED

“Fooocus has no advanced controls.”

Rejected. Its inspected UI includes models, LoRAs, sampler, scheduler, guidance, sharpness, image-prompt weights and developer options.

VERIFIED

“They are versions of the same app.”

Rejected. The same creator owns both repositories, but they are separate applications, install trees, scopes and update histories.

ASSUMPTION

“More visible controls are always better.”

Extra choices help only when the task needs them and the operator can restore them. Otherwise they are maintenance cost.

09 · QUESTIONS BEFORE MOVING

Direct answers to the comparison queries.

The answer changes when the model family, edit or dependency changes. Start with that noun.

Is Forge better than Fooocus?

Not universally. Forge is the stronger candidate when you need visible model, sampling, ControlNet and memory decisions. Fooocus is the stronger candidate when a prepared SDXL prompt-and-image workflow removes decisions you do not want to manage.

Should a beginner use Forge or Fooocus?

Start with Fooocus if the goal is a first SDXL image and guided built-in edits with minimal setup choices. Start with Forge if learning the model files and generation parameters is part of the goal, or if a required non-SDXL family rules Fooocus out.

Is Fooocus easier to use than Forge?

For its intended SDXL path, usually fewer decisions are visible before generation. That does not make every task easier: Advanced settings, first-use downloads and unsupported model families can move the complexity elsewhere.

Does Forge or Fooocus make better images?

The UI name alone cannot answer that. Compare model and VAE hashes, effective prompt after expansion/styles, sampler, scheduler, steps, guidance, seed, dimensions and post-processing. We did not run a controlled GPU quality test for this page.

Which is faster, Forge or Fooocus?

No universal winner is claimed. Measure cold start, model load, first generation and warm repeats on the same machine with the same supported SDXL model, dimensions, batch and runtime. Include setup and editing time if workflow speed is the real question.

Which uses less VRAM, Forge or Fooocus?

That depends on hardware, runtime, model, refiner, precision, offload, dimensions and the chosen task. Forge exposes more memory decisions; exposure is not proof of a lower peak. Record VRAM and system-memory spill locally.

Which is better for SDXL?

Both are credible SDXL candidates with different operating models. Fooocus is entirely centered on an opinionated SDXL experience. Forge supports SDXL inside a broader WebUI-shaped model and control surface. Run the same checkpoint in both, then run Fooocus defaults as a separate trial.

Can Fooocus run FLUX?

The original lllyasviel/Fooocus README says the project is built entirely on SDXL, is in limited LTS and has no current plans for newer model architectures. It recommends alternatives including WebUI Forge for FLUX. Do not use a Fooocus fork as evidence for the original.

Can Fooocus run Stable Diffusion 1.5 models?

Original Fooocus is officially scoped to SDXL, so SD 1.5 is not the supported route described by its current README. Original Forge exposes an sd preset, but the exact checkpoint and workflow still need testing on the pinned commit.

Does Fooocus hide all sampler and guidance controls?

No. The inspected original Fooocus UI exposes sampler, scheduler, guidance, sharpness, performance and model controls under Advanced. Its design automates defaults first; it does not remove every advanced setting.

What is Fooocus V2 prompt expansion?

The official documentation describes an offline GPT-2-based prompt-processing engine used by the Fooocus V2 style. Treat the expanded prompt and selected styles as part of the recipe when comparing or reproducing an output.

Can I use negative prompts in Fooocus?

Yes. Negative Prompt is available under Advanced in the documented interface. A preset or performance mode can still alter the effective recipe, so record those choices too.

Which is better for inpainting and outpainting?

Fooocus offers a guided Inpaint or Outpaint route, directional expansion and Enhance around its own SDXL methods. Forge offers Forge Canvas and a denser img2img/inpaint control surface. Test the complete mask, context, revision and export task.

Which is better for ControlNet?

Choose Forge when you need named ControlNet models, preprocessors, weights and guidance ranges in a WebUI-style unit. Choose Fooocus when its built-in Image Prompt types complete the task without exposing that chain. They are not equivalent interfaces.

Does Fooocus support image prompts and FaceSwap?

The inspected original Fooocus source includes ImagePrompt, PyraCanny, CPDS and FaceSwap types with adjustable stop and weight values in Advanced. Verify required model downloads and licenses before treating the route as portable.

Can the same SDXL LoRA work in Forge and Fooocus?

The file may be reusable, but support is not guaranteed by visibility. Match the SDXL base model, trigger text, weight, loader behavior and any style or prompt expansion, then compare a controlled baseline.

Can Forge and Fooocus share model files?

They can reference deliberately shared storage, but folder mapping and runtime support must be configured. Start with copies or read-only source storage, verify hashes and keep each application environment separate.

Can I import Fooocus settings into Forge?

No native settings importer was verified. Preserve the Fooocus preset, config, styles, expanded prompt and image metadata, then map only the task-relevant settings into Forge. Some prepared behavior has no one-to-one control.

Are Forge and Fooocus made by the same developer?

Both official repositories are under lllyasviel, and the Fooocus README points users to WebUI Forge for newer models. They remain separate applications with different code, licenses, scopes, downloads and update histories.

Is Fooocus still maintained?

Its current README states limited long-term support with bug fixes only. The latest inspected main commit is 2 September 2025 and changes CI configuration; the latest published release is v2.5.5 from 12 August 2024. Read those facts together.

Which project is more active?

A raw latest-commit date is misleading here. Fooocus main is later because of a CI dependency bump, while its product scope is explicitly limited LTS. Forge main last changed in June 2025. Neither fact decides whether your exact workflow works.

Can Forge and Fooocus run offline?

Both are local applications, and Fooocus explicitly describes itself as offline. Initial packages, models, inpaint files, extensions or optional components can still require downloads. Verify first-run traffic and the completed local environment yourself.

Should I replace Fooocus with Forge?

Only if Forge solves a named limit—such as a required model family, explicit conditioning chain or resource control—and repeats the whole job reliably. Keep Fooocus intact until Forge passes the receipt. Keeping both for different jobs is a valid result.

10 · EVIDENCE LEDGER

Primary product facts. Dated community context.

Official repositories establish scope and controls. A community tutorial can reveal a real question, but it cannot establish current quality, speed or maintenance by itself.

Local research catalog used · 6 records
O01Original Forge repository and README

Primary product identity, model routes and control terminology.

O13Forge Canvas announcement #853

Direct-edit and inpaint interaction evidence; dated July 2024.

O18Performance reporting #1181

Unsafe universal benchmark claims rejected.

O20Forge extension list #1754

Third-party compatibility is a dated ledger, not a blanket promise.

DossierOriginal Fooocus README competitor audit

Simplicity, provenance warning and prepared-workflow thesis.

V06Wossi86 Forge UI walkthrough

Community positioning and beginner questions; 2025 opinion, not current product proof.

NO GPU GENERATION

We did not run a matched Forge/Fooocus speed, VRAM or image-quality benchmark for this page.

NO FORK TRANSFER

Fooocus forks and Forge continuations were not used to extend either original product’s feature list.

EDITORIAL ASSUMPTION

“Control budget” is our decision model, not terminology or a recommendation published by either project.

AuthorForge Field Guide editorial teamReviewerTechnical editorial reviewUpdated31 Aug 2026Snapshotsdfdcbab / ae05379