FORGEFIELD GUIDE
SD 1.X · SDXL · FLUX · NO UNIVERSAL WINNER

Choose the model family that fits your workflow

Start with what cannot change: the add-on or workflow you must use, the exact model license, and the complete file set. Use hardware as a measured acceptance test—not a guessed VRAM cutoff.

Open the decision path Compare the families
Evidence checked24 Aug 2026Code snapshotdfdcbab · 26 Jun 2025ScopeOriginal Forge only
01 · Direct answer

Constraints choose before preferences do

The family decision becomes small once you remove candidates that cannot run the required asset, workflow, or license.

  1. 01

    Required asset

    A LoRA, embedding, ControlNet model, adapter, or fine-tune already names its compatible base family. That contract comes first.

    FIXES FAMILY
  2. 02

    Required workflow

    Confirm that the current original Forge build and every extension in the workflow support the exact family and variant.

    REMOVES CANDIDATES
  3. 03

    Exact license

    Check weights, derivatives, outputs, redistribution, commercial purpose, and access conditions on the publisher’s page.

    PASS / FAIL
  4. 04

    Complete layout

    Count the checkpoint or diffusion model plus every required VAE, CLIP, T5, refiner, and configuration dependency.

    LOAD CONTRACT
  5. 05

    Local acceptance test

    Measure the smallest valid job on your machine, then add only the components the real workflow needs.

    KEEP / REJECT
If you already need one family-specific add-on, you do not have a three-way model contest.

You have a compatibility check inside that family. Find or train an equivalent asset before considering a switch.

See the SD 1.x ↔ SDXL LoRA boundary in Forge Q&A ↗
02 · Constraint router

Which statement describes the first real job?

Select the strongest constraint. The result is a shortlist or a gate—not an algorithm pretending to know your taste and hardware.

Choose your strongest model-family constraint
DECISIONThe add-on chooses the shortlistMatch its declared base family
NEXT ACTION

Read the LoRA, embedding, ControlNet, IP-Adapter, or extension documentation. If it says SD 1.5, SDXL, or FLUX, start with that family and exact supported version.

STOP SIGNAL

Do not try to make the file cross families because its extension looks compatible.

03 · Family matrix

Three architecture contracts, not three quality scores

These cards describe the decision boundary. Exact fine-tunes and conversions can change the package, settings, terms, and measured load.

01 · 512-era latent diffusion

SD 1.x

FORGE UI · sd

Choose it when an existing SD 1.x checkpoint, LoRA, embedding, ControlNet model, or established 512-era workflow is the non-negotiable part of the job.

Typical file set
Usually a self-contained checkpoint; a model card may require a separate VAE.
Prompt controls
Uses the classic CFG and negative-prompt workflow exposed by the sd preset.
Starting scale
The SD v1.5 base model was trained and evaluated at 512 × 512. Follow the exact fine-tune’s documented aspect ratios and baseline.
Add-on boundary
Use SD 1.x-trained add-ons. SDXL and FLUX add-ons do not become compatible because they share .safetensors.
First risk to test
A mature-looking workflow can still depend on an old extension or model version that the current Forge build handles differently.
Open the base model card ↗Run the SD 1.5 baseline in Forge →
02 · High-resolution two-encoder family

SDXL

FORGE UI · xl

Choose it when the exact SDXL checkpoint and its add-on ecosystem fit the task, and you want the familiar Forge checkpoint workflow without a FLUX-specific dependency.

Typical file set
The SDXL base is usable alone; the original pipeline also documents an optional refiner. Fine-tunes often ship as one checkpoint.
Prompt controls
Uses the conventional prompt and negative-prompt controls; the xl preset hides Clip skip and exposes GPU Weights.
Starting scale
Use the model publisher’s native-size guidance. Forge’s inspected xl defaults are a starting interface state, not a rule for every fine-tune.
Add-on boundary
Use SDXL-trained LoRAs, embeddings, and control models. A Pony or other derivative still needs its own documented compatibility contract.
First risk to test
Assuming every SDXL download needs the original refiner or uses identical VAE, resolution, and prompt syntax.
Open the base model card ↗Run the SDXL baseline in Forge →
03 · 12B rectified-flow variants

FLUX

FORGE UI · flux

Choose it when a specific FLUX variant’s prompt behavior, workflow, file layout, and license meet the job—and you can validate its full memory path locally.

Typical file set
Packaged NF4/FP8 checkpoints or a separate diffusion model plus VAE, CLIP-L, and T5, depending on the distribution.
Prompt controls
Forge’s flux preset exposes Distilled CFG, swap controls, and GPU Weights; official dev guidance uses a different guidance contract from classic SD.
Starting scale
Use the exact variant documentation. BFL’s reference examples use 1024 × 1024, but a converted Forge checkpoint may publish another tested baseline.
Add-on boundary
Use FLUX-compatible LoRAs and controls for the exact variant/format. Low-bit loading and LoRA patching can change RAM as well as VRAM behavior.
First risk to test
Choosing “FLUX” without deciding dev versus schnell, license, packaged versus split layout, and required encoders.
Open the base model card ↗

VERIFIED Current Forge presents sd, xl, flux, and all UI presets; its loader separately recognizes SD 1, SDXL, and FLUX structures from weight keys. Inspect the presets · Inspect model detection ↗

04 · Compatibility

The family name is an interface between every component

Each part must speak to the same architecture. Folder placement and a visible card do not repair a shape or encoder mismatch.

BASECheckpoint or diffusion model
CONDITIONINGCLIP · T5 · embedding
MODIFIERSLoRA · ControlNet · adapters
WORKFLOWPreset · extension · settings
VISIBLE ≠ COMPATIBLE

The scanner found a file

A .safetensors file appearing only proves discovery. The internal keys and tensor shapes still have to match the loaded architecture.

PRESET ≠ CONVERTER

The UI shows suitable controls

The preset changes visibility and starting values. It does not transform an SDXL model into FLUX or infer the license and add-ons.

FAMILY ≠ EXACT VERSION

Derivative rules can narrow further

A family-compatible asset can still target one base variant, fine-tune, encoder, prediction type, or precision path.

MODEL ≠ WHOLE WORKFLOW

Test the dependency set

A base image can work while the required LoRA, ControlNet, refiner, or high-resolution pass exceeds memory or fails to load.

05 · License gate

Read terms at checkpoint level

Family names are technical categories. They do not give one shared permission set.

BASE EXAMPLE

Stable Diffusion v1.5

The current mirror reproduces the CreativeML OpenRAIL-M model card and warns that the original RunwayML repository is deprecated. A derivative may add its own terms.

Model card ↗
BASE EXAMPLE

SDXL 1.0 base

Stability AI lists CreativeML Open RAIL++-M. The base card does not grant blanket clearance for every SDXL fine-tune or dataset.

Model card ↗
VARIANT SPLIT

FLUX.1-dev

Black Forest Labs lists its dev non-commercial license. Its card discusses permitted generated-output use, but the weights and derivatives remain governed by the full terms.

Model card ↗
VARIANT SPLIT

FLUX.1-schnell

The publisher currently lists Apache-2.0 and describes personal, scientific, and commercial use. Verify any third-party conversion or derivative separately.

Model card ↗
License receipt for the exact download
RECORDED TERMS0 / 5 recorded

This is an evidence checklist, not legal advice. When the interpretation affects a real business or distribution decision, obtain qualified legal review.

06 · Hardware decision

Measure the complete path, not the VRAM label

Two machines with the same advertised VRAM can differ by system RAM, shared-memory behavior, driver, precision path, thermal limits, and workload.

  1. 01

    Choose one exact candidate

    Record publisher, model page, filename, hash, family, variant, license, and all required companion files.

  2. 02

    Use the matching Forge preset

    Select sd, xl, or flux. Start from the model publisher’s baseline, not settings copied from another family.

  3. 03

    Make the workload minimal but valid

    One image, batch size 1, documented resolution, required modules only, no LoRA, ControlNet, refiner, high-resolution pass, or unrelated extension.

  4. 04

    Record cold and warm behavior

    Keep model-load result and time separate from the next generation. Capture dedicated VRAM, shared GPU memory, system RAM, elapsed time, and first console warning.

  5. 05

    Add the real dependency

    If the task needs a LoRA or ControlNet, add exactly one and repeat with the same seed. A base-only pass is not final acceptance.

  6. 06

    Decide against written limits

    Keep the family only if the exact workflow meets your time, memory, stability, output, and license requirements.

More GPU Weights is not always safer or faster.

Current Forge warns that allocating nearly all GPU memory to model weights can leave no headroom for computation, cause fallback, or trigger OOM. Larger images need their own headroom.

Inspect the current warning ↗
07 · Decision record

Keep a result another person can repeat

A family choice is useful when the evidence names the exact model and the conditions under which it passed.

IDENTITYForge commit · model URL · version · hash
CONTRACTFamily · variant · license · complete file set
BASELINEPreset · prompt · seed · sampler · steps · dimensions
MEMORYDedicated VRAM · shared · RAM · GPU Weights
TIMECold load · warm generation · second run
OUTCOMEImage · console · required add-on · accept/reject reason
SAME USER JOB

Use the same target scene and acceptance criteria.

FAMILY-VALID BASELINE

Let each exact model use its documented settings.

USEFUL COMPARISON

Compare successful workflows, not sabotaged defaults.

08 · Rejected shortcuts

Six rules that produce the wrong family

Each shortcut removes one of the constraints that actually determines whether the workflow succeeds.

“The newest family is automatically the best”

Newer does not preserve a required LoRA, ControlNet model, extension, license, prompt syntax, or acceptable runtime. Choose against the job.

“My GPU has 8 GB, so one family is guaranteed”

VRAM alone omits model format, full file set, resolution, batch, shared memory, RAM, driver, Forge commit, and add-ons. It defines a test environment, not a verdict.

“Every .safetensors file works together”

The extension is a container. The architecture, internal keys, tensor shapes, encoder contract, and trained base determine compatibility.

“The Forge preset changes the model family”

The preset changes controls and defaults. The loader still has to recognize the selected weights as SD 1, SDXL, FLUX, or another supported structure.

“A smaller download always needs less runtime memory”

Runtime adds text encoders, VAE, sampling tensors, temporary conversions, LoRA patching, and offload buffers. Measure the complete workload.

“Forge’s license covers the model”

Forge software and model weights are separate works. Save and review the exact model terms before the file loses its page context.

09 · Selection FAQ

Questions to answer before downloading a second model

The answers keep family selection separate from later choices about fine-tunes, quantization, folders, and performance tuning.

What is the best model family for Stable Diffusion WebUI Forge?

There is no universal winner. The correct family is the one that satisfies the required workflow and add-ons, passes the exact model license, includes a complete supported file set, and completes your measured baseline on your machine.

Should a Forge beginner start with SD 1.5, SDXL, or FLUX?

Start from your first real task. If nothing fixes the family, SDXL is a reasonable candidate for a familiar checkpoint workflow, but it is still a shortlist—not a guarantee. Test one documented model before collecting add-ons.

What is the difference between SD 1.5, SDXL, and FLUX in Forge?

They are different model architectures with different text encoders, native training scales, guidance behavior, file layouts, and compatible add-ons. Forge can expose each through a related interface, but it does not make their components interchangeable.

Which Forge UI preset should I use for SD 1.5, SDXL, and FLUX?

Use sd for SD 1.x, xl for SDXL, and flux for FLUX. The preset changes visible controls and starting values; it does not convert or identify a mislabeled model for you.

Does selecting the flux preset convert an SDXL checkpoint to FLUX?

No. The preset only changes the Forge control surface and defaults. The loader still has to identify a compatible architecture inside the selected weights.

Can I use an SD 1.5 LoRA with SDXL in Forge?

No. Treat LoRAs as base-family-specific. Use an SD 1.x LoRA with its documented SD 1.x base and an SDXL LoRA with its documented SDXL base.

Can I use an SDXL LoRA with FLUX?

No unless the publisher explicitly provides a compatible conversion for that exact FLUX architecture. A shared .safetensors extension is only a container format.

Are ControlNet models compatible across SD 1.5, SDXL, and FLUX?

Do not assume they are. Match the ControlNet weights to the declared base family, control type, Forge version, and workflow documentation.

Which model family uses the least VRAM in Forge?

No fixed number safely answers this for every checkpoint and workflow. SD 1.x is generally the smaller architecture candidate, but precision, encoders, VAE, resolution, batch, add-ons, offload settings, driver, and Forge commit all affect the measured result.

Can I run FLUX in Forge with 4 GB, 6 GB, or 8 GB VRAM?

Community demonstrations and dated official tests show some low-memory FLUX configurations, but they are not a universal support guarantee. Test one exact format at batch size 1 and record dedicated VRAM, shared memory, system RAM, time, and errors.

Is a smaller GGUF or NF4 file automatically better for low VRAM?

No. Quantization format affects storage and loading behavior, but the complete workflow still includes encoders, VAE, sampling headroom, resolution, and add-ons. Compare exact distributions under a controlled baseline.

Does model download size equal required VRAM?

No. Download size is not peak runtime memory. Loading precision, temporary conversion, text encoders, VAE, sampling tensors, resolution, batch, LoRA patching, and offload location also matter.

Does SDXL require the refiner in Forge?

Not necessarily. The official SDXL base model card states that the base can be used standalone. Use a refiner only when the exact workflow documents it and your baseline justifies the added load.

Does FLUX require separate T5, CLIP-L, and VAE files?

Some layouts do and some packaged checkpoints do not. Official Forge guidance separates raw/GGUF diffusion models from packaged checkpoints. Follow the exact distribution’s file list.

Can I use negative prompts with FLUX like SD 1.5 or SDXL?

Do not copy the classic SD guidance contract. Forge’s official FLUX dev guidance uses CFG 1 and Distilled CFG, with the negative prompt disabled at CFG 1. Follow the exact variant documentation.

Which family is best for existing embeddings and extensions?

The family declared by those assets is the starting constraint. Then verify that the current original Forge build and the exact extension version still support the workflow.

Which model family can I use commercially?

A family name is not legal clearance. The SD v1.5, SDXL, FLUX.1-dev, FLUX.1-schnell, and third-party fine-tunes can carry different terms. Review the exact model and upstream licenses; this page is not legal advice.

What is the license difference between FLUX.1-dev and FLUX.1-schnell?

Their current publisher model cards list FLUX.1-dev under the FLUX.1 dev non-commercial license and FLUX.1-schnell under Apache-2.0. Read the complete current terms and any derivative-model license before use.

How should I compare SD 1.5, SDXL, and FLUX on my PC?

Give each candidate the same user task, but use that exact model’s documented baseline rather than forcing identical incompatible settings. Record files, hash, preset, resolution, load result, warm generation time, dedicated VRAM, shared memory, RAM, and errors.

When should I stop testing a model family?

Stop when the license fails your use, a required add-on has no compatible version, the complete file set cannot load, or the smallest valid baseline exceeds your acceptable time or memory limits.

10 · Evidence scope

Architecture facts are verified; “best” remains your decision

We use current code and model cards for contracts, and community material only to map questions and failure modes.

VERIFIED

Current Forge code

Preset controls, model-type detection, GPU Weights behavior, and visible setting changes at inspected commit dfdcbab.

VERIFIED

Publisher model cards

Base architecture descriptions, example pipelines, variant access, and the displayed licenses for SD v1.5, SDXL, FLUX dev, and schnell.

STALE SNAPSHOT

Dated Forge guidance

The 2024 NF4/FP8 and split-FLUX posts document supported layouts and measured examples, not permanent rankings or hardware guarantees.

COMMUNITY-REPORTED

User constraints

Five catalogued video transcripts and issue/discussion reports expose VRAM, LoRA, prompt, and setup questions; their numeric claims are not universalized.

Transcript setV01 ↗V03 ↗V04 ↗V08 ↗V33 ↗
Author
Forge Field Guide editorial team
Technical review
Forge code + publisher model cards
Inspected snapshot
dfdcbab · 26 Jun 2025
Refresh trigger
Loader, model card, license, or workflow change