FORGEFIELD GUIDE
FAMILY · FILE SET · COMPATIBILITY · LICENSE

Models in Stable Diffusion WebUI Forge

Forge is the interface and loader; the model supplies the learned image-generation weights. Identify the family, exact version, complete file set, license, and published baseline before testing one clean generation.

Choose a model family Understand VAE and text encodersOpen the exact folder map
Evidence checked24 Aug 2026Code snapshotdfdcbab · 26 Jun 2025
01 · Model boundary

A filename is not a complete model specification

The extension tells Forge what kind of container it may try to inspect. It does not tell you the base family, companion files, license, memory cost or whether the loader supports that exact architecture.

01Base familySD 1.x · SDXL · FLUX
02File layoutBundled checkpoint or split modules
03Matching add-onsVAE · encoder · LoRA · controls
04Loadable setupProved by one baseline image
PRIMARY WEIGHTS

Checkpoint or diffusion model

The main learned weights. Forge labels the top selector Checkpoint even when a newer layout uses a diffusion-model file plus separate modules.

IMAGE CODEC

VAE

Encodes images into latent space and decodes latents back to pixels. It may be embedded, managed automatically or supplied as a separate file.

PROMPT CONDITIONING

Text encoder

Turns text into conditioning the diffusion model can use. Split FLUX layouts can require CLIP-L and T5 selections together.

BASE MODIFIER

LoRA

A smaller set of learned changes applied to a compatible base. It is not a standalone generator and does not cross model families automatically.

Install and validate one LoRA →
TOKEN MODIFIER

Embedding

A textual-inversion file activated through prompt vocabulary. Its vector shape and training family must match the loaded model.

STRUCTURE CONTROL

ControlNet model

A separate control model used inside the integrated ControlNet workflow. It must match both the intended control type and base family.

02 · Family selector

Choose by workflow and published compatibility

This overview introduces the family boundary. Use the complete SD 1.x vs SDXL vs FLUX decision guide for add-on, license, layout, and local hardware gates.

NO UNIVERSAL WINNER

The smaller download, newer family or lower-bit file is not automatically the right model for your workflow.

01SD 1.xUI: sd
Typical layout
Usually one checkpoint; a separate VAE may be requested.
Choose when
A model card explicitly identifies Stable Diffusion 1.x compatibility and the workflow depends on its ecosystem.
Compatibility boundary
Use SD 1.x LoRAs, embeddings and ControlNet models. An SDXL or FLUX add-on is not interchangeable.
Run SD 1.5 in Forge →
02SDXLUI: xl
Typical layout
Usually one checkpoint; some workflows specify a VAE or refiner separately.
Choose when
The publisher identifies SDXL as the base family and documents its intended resolution and generation settings.
Compatibility boundary
Use SDXL-trained add-ons. A filename containing “XL” is a clue, not sufficient evidence by itself.
Run SDXL in Forge →
03FLUXUI: flux
Typical layout
Either a packaged checkpoint or a split diffusion model plus VAE, CLIP-L and T5 files.
Choose when
The exact variant has Forge-specific instructions and your machine can hold its full file set and runtime workload.
Compatibility boundary
NF4, FP8, GGUF and full weights are different layouts or storage choices, not quality tiers Forge can choose for you.
Run FLUX in Forge →
FLUX needs a layout decision, not just the flux preset.

Official Forge guidance distinguishes packaged NF4/FP8 checkpoints from raw or GGUF diffusion models that use separate VAE, CLIP-L and T5 files. The preset changes the interface; it does not assemble missing components.

Inspect the official split-file guide ↗
03 · Before download

Record the model card before the file leaves it

A local filename often loses the information needed to reproduce, license or troubleshoot the model later.

  1. ENTITY

    Publisher and original model page

    Save the repository or model-card URL. A repost with the same name may be a conversion, merge or different version.

  2. FAMILY

    Base architecture and variant

    Record SD 1.x, SDXL or FLUX, plus the exact base or fine-tune identity. Do not infer it from artwork, tags or file extension.

  3. LAYOUT

    Required companion files

    Check whether VAE, CLIP, T5, refiner or another module is embedded, optional or mandatory. Download the set as one versioned unit.

  4. FILE

    Exact filename, size and hash

    A partial download can have a plausible name. Compare the published size and checksum when one is provided before diagnosing Forge.

  5. TERMS

    License and access conditions

    Read the terms for use, derivatives, redistribution and commercial work. Forge’s AGPL-3.0 license does not replace a model’s separate license.

  6. RECIPE

    Publisher’s starting settings

    Keep the intended resolution, sampler, scheduler, steps, guidance behavior, VAE, triggers and negative-prompt notes with the file.

A model can be technically loadable and still be unsuitable for your intended use or prohibited by its terms. This page explains Forge file handling, not legal clearance.

04 · Inventory overview

Keep every component in its owning inventory

This compact overview connects model roles to Forge controls. For active-root resolution, all path overrides, upscalers, refresh behavior, and missing-file diagnosis, use the complete model folder map.

Choose a model file role
ACTIVE ROUTECheckpoint / primary model
DEFAULT PATHmodels/Stable-diffusion
Appears in
Checkpoint
Current scanner accepts
.safetensors · .ckpt · .gguf
Done state
The filename appears in Checkpoint; generation then proves the complete layout.
Default routes in inspected original Forge code
File roleDefault relative pathForge UI inventoryPrimary code
Checkpoint / primary modelmodels/Stable-diffusionCheckpointsd_models.py ↗
VAEmodels/VAEVAE / Text Encodermain_entry.py ↗
Text encodermodels/text_encoderVAE / Text Encodermain_entry.py ↗
LoRAmodels/LoraLoRA Extra Networkspreload.py ↗
EmbeddingembeddingsTextual Inversion Extra Networkscmd_args.py ↗
ControlNet modelmodels/ControlNetIntegrated ControlNetshared.py ↗

Folder names matter on case-sensitive filesystems. The standalone LoRA default is models/Lora. Always inspect the launch line when --data-dir, --models-dir, --ckpt-dir, --vae-dir, --text-encoder-dir, --lora-dir or --controlnet-dir is present.

05 · Place and load

Discovery, selection and generation are three separate checks

Do not treat a visible filename as the finish line. Forge must discover the file, assemble the selected modules, and complete generation without a mismatch or traceback.

  1. Finish and verify every download

    Make sure the browser or download client has removed any partial-download suffix. Compare the exact filename, expected size and publisher hash when available.

    EXPECTED · one complete versioned file set

  2. Place by role, not by extension

    A VAE and a checkpoint can both use .safetensors; they still belong in different inventories. Keep related files in documented folders rather than moving everything into Checkpoint.

    EXPECTED · each component has one resolved path

  3. Refresh the inventory that owns the file

    The top Forge refresh control rebuilds Checkpoint and VAE / Text Encoder choices. Extra Networks and integrated controls maintain their own lists; refresh there when available or restart once.

    EXPECTED · the exact filename becomes selectable

  4. Select family before modules

    Choose sd, xl or flux, then select the primary model. Add a separate VAE or text encoders only when the model layout requires them.

    EXPECTED · console prints the selected loading parameters

  5. Generate without modifiers

    Use one image, the model publisher’s baseline settings, no LoRA, no ControlNet and no high-resolution pass. Watch the console through completion.

    EXPECTED · valid output with no fatal traceback

  6. Add one compatible modifier

    Fix the seed, add one LoRA or other module, and compare. If the output does not change, inspect family compatibility, trigger syntax and console mismatch messages.

    EXPECTED · one attributable change
DISCOVEREDListed in the correct inventory
ASSEMBLEDRequired modules selected
PROVEDClean image + clean console
06 · Shared libraries

Reuse weights, not environments

Large model files can be referenced from another location. Forge itself, its Python environment, extensions and configuration should remain independently recoverable.

Shared storage adds a path dependency.

Record the resolved source folder, preserve quoting around paths with spaces, and test each inventory. Never point Forge at another UI’s virtual environment or copy its extension folder merely because model sharing works.

Read the original shared-model announcement ↗
07 · Failure router

Start with the earliest model state that failed

Changing precision, memory controls and extensions cannot fix a file that was never discovered or a LoRA trained for another family.

The Checkpoint list is empty or the new model is missing

Confirm the active data directory and resolved checkpoint path, then check the filename and completed size. Current code scans .ckpt, .safetensors and .gguf in the checkpoint inventory. Use the top refresh control or restart once. A VAE, encoder or LoRA will not appear here.

The checkpoint appears, but Forge says a CLIP, T5 or VAE state dict is missing

You probably selected a split layout without its required modules, or selected the wrong companion file. Reopen the publisher’s file checklist, place VAE in models/VAE and text encoders in models/text_encoder, refresh, then select every required item in VAE / Text Encoder. Do not guess between CLIP-G, CLIP-L and T5.

Diagnose VAE and text encoders →Inspect a current unresolved example ↗
The LoRA folder has files, but the LoRA tab is empty

Check the resolved --lora-dir and standalone default models/Lora. Watch capitalization on Linux and do not use models/Loras unless that exact directory is configured. Refresh Extra Networks or restart, then inspect the console for file-read errors.

Repair model and LoRA discovery →See the reported discovery symptom ↗
The LoRA is listed, but the image does not change

Match the LoRA’s training family to the loaded base, include its documented trigger, and compare one fixed seed with and without the LoRA. The current loader can report version mismatch, unmatched keys or skipped keys; those messages are stronger evidence than the card being visible.

Run the generic LoRA proof →Inspect the current LoRA loader ↗
The preview looked correct, but the saved result is black or grey

Return to the exact model layout and VAE requirement. A wrong or missing decoder can fail at the final image stage. Remove LoRAs and extensions, use the matching preset, and retest the documented VAE before changing precision flags.

Forge runs out of memory while loading or generating

First confirm that the selected layout is the one you intended. Then reduce one workload variable at a time: optional modules, image dimensions, batch, or model-weight allocation. Do not treat a model’s download size as its full VRAM/RAM requirement, and do not copy hardware promises from a different format.

Read the dated official memory guidance ↗
A shared A1111 path exposes checkpoints but not LoRAs or VAE files

Partial success means each mapping must be checked independently. Inspect the launch line, expected subfolder names and case. A specific directory argument can take precedence, and current mapping code skips a target path that does not exist.

The same model name gives a different result after replacement or update

Compare hashes, not display names. Also record the selected VAE/text encoders, preset, precision choice, sampler, scheduler, seed, dimensions, Forge commit and extensions. A replaced conversion or companion module can change output while the visible model name looks familiar.

08 · Model FAQ

Questions that reveal a missing model decision

These answers separate file discovery from architecture compatibility, loading and usage rights.

What models work with Stable Diffusion WebUI Forge?

There is no timeless “all models” list. Current checkpoint discovery accepts .safetensors, .ckpt and .gguf files, but the loader must also recognize the architecture and receive every required companion module. Check the model publisher’s family, layout and Forge-specific instructions before downloading.

Where do I put Stable Diffusion models in Forge?

Use the dedicated Forge model-folder map to resolve the active data root, path overrides, exact inventory directory and refresh control. Primary checkpoints normally use models/Stable-diffusion, but --data-dir, --models-dir and inventory-specific arguments can change the effective location.

Why is my checkpoint not showing in Forge?

Confirm the active data/model path, finished filename and extension, then use the top model refresh control or restart. If the file is a VAE, text encoder, LoRA, embedding or ControlNet model, it belongs in a different inventory and will not appear as a checkpoint.

What is the difference between a checkpoint and a LoRA?

A checkpoint or primary diffusion model supplies the base generation system. A LoRA is a smaller modification applied to a compatible base family. A LoRA cannot generate by itself and may load with mismatched or skipped keys when paired with the wrong architecture.

Do I need a separate VAE in Forge?

Only when the model publisher requires one or when the chosen layout stores it separately. Many checkpoints already contain or define the necessary components. Do not add a random VAE to repair an unidentified model mismatch.

Can I use an SDXL LoRA with FLUX in Forge?

No. Treat LoRAs as base-family-specific unless their publisher explicitly documents another compatible architecture. Matching the .safetensors extension does not make the internal keys compatible.

Can Forge share model folders with AUTOMATIC1111 or ComfyUI?

Forge has an official --forge-ref-a1111-home argument and individual directory arguments. ComfyUI uses a different folder layout, so there is no universal one-root mapping for every component. Validate each resolved path and keep repositories, extensions and Python environments separate.

Does .safetensors, .ckpt or .gguf tell me which model family it is?

No. These are file containers or storage formats, not the SD 1.x, SDXL or FLUX family label. Record the publisher, model page, base architecture, version and required files; do not classify a model from the extension alone.

Why does a LoRA appear in Forge but have no effect?

Use the correct base family and trigger syntax, check the console for version mismatch or skipped keys, and compare the same prompt and seed with only that LoRA changed. A visible card proves discovery, not successful application.

Can I use a downloaded model commercially?

Forge’s software license does not grant rights to a separate model. Read the exact model and version license for use, redistribution, derivatives and commercial restrictions. This guide cannot replace legal advice.

09 · Evidence scope

What this page proves—and what it does not

File paths and UI labels come from current original Forge code. Model support and performance remain exact-version questions.

VERIFIED

Code-verified

Default base paths, accepted discovery extensions, selector labels, refresh behavior, LoRA path and ControlNet path.

STALE SNAPSHOT

Dated official guidance

NF4/FP8, FLUX LoRA and split GGUF/FP16 instructions were published in August 2024 and are not a live compatibility matrix.

COMMUNITY-REPORTED

Observed user symptoms

Missing LoRA cards, partial shared paths and missing state-dict errors define diagnostic questions, not universal fixes.

UNKNOWN

Not certified here

No universal GPU threshold, quality ranking, complete model catalogue or commercial-use clearance is claimed.

Author
Forge Field Guide editorial team
Technical review
Original code + primary discussions
Inspected snapshot
dfdcbab · 26 Jun 2025
Refresh trigger
Loader, path, preset or model-support change