Find the directory used by the running Forge process, identify what the file does, and put it in the folder scanned by that UI control. A familiar filename in a guessed folder is not enough.
Multiple clones, extracted packages, shortcuts, and custom arguments can all make a correct relative path point to the wrong installation.
Use the terminal or launcher that opened the UI. Do not start from an old desktop shortcut, download folder, or another WebUI’s path.
Look for --data-dir, --models-dir, an inventory-specific directory argument, or --forge-ref-a1111-home.
Only then append Stable-diffusion, Lora, VAE, or another inventory directory.
--data-dir or program directory--models-dir or [data]/modelsStable-diffusion, Lora, …The default is [data]/embeddings, beside models. Changing --models-dir does not move them; use --embeddings-dir or change the data root.
Select the launch setup you see. The result describes the base paths; inventory-specific arguments are checked in the next section.
<Forge data directory>/models<Forge data directory>/embeddingsForge uses the active data directory. Without --data-dir, this is normally the WebUI program directory.
VERIFIED Root behavior is traced from paths_internal.py, cmd_args.py, and launch_utils.py ↗
These are default relative paths, not a promise that every installation has already created every optional directory.
<active data directory>/
├── embeddings/
└── models/
├── Stable-diffusion/
├── VAE/
├── text_encoder/
├── Lora/
├── ControlNet/
├── ControlNetPreprocessor/
├── hypernetworks/
├── ESRGAN/ ├── RealESRGAN/
├── BSRGAN/ ├── DAT/
├── HAT/ ├── ScuNET/
└── SwinIR/Case matters on Linux: the current standalone LoRA default is models/Lora. Historical shared A1111 layouts can use lowercase models/lora; use the path actually configured, not a silent spelling variation.
A .safetensors extension does not tell you whether the file is a checkpoint, LoRA, VAE, or ControlNet model. The publisher’s model card must identify the role.
models/Stable-diffusionSubfolders are scanned. Files ending .vae.ckpt or .vae.safetensors are excluded from checkpoint inventory.
models/VAEForge can also find a same-name .vae file beside its checkpoint. Only add one when the model publisher requires it.
models/text_encoderSplit FLUX layouts may require CLIP-L and T5 in addition to the diffusion model and VAE.
models/LoraThe standalone Forge default is capital-L Lora. Linux paths are case-sensitive.
embeddingsThis folder is beside models, not inside it. A discovered file can still be skipped as incompatible.
models/ControlNetPreprocessor downloads use models/ControlNetPreprocessor by default; they are not ControlNet weights.
models/<upscaler family>Common roots include ESRGAN, RealESRGAN, BSRGAN, DAT, HAT, ScuNET, and SwinIR.
models/hypernetworks--hypernetwork-dir replaces this default inventory path.
Use this as a placement checklist. “Visible” and “loads successfully” are separate gates.
| File role | Default path | Owning UI | Proof |
|---|---|---|---|
| Checkpoint / diffusion model | models/Stable-diffusion | Checkpoint selector | The exact file appears in Checkpoint. Then complete a plain generation to prove it loads. |
| Separate VAE | models/VAE | VAE / Clip selector | The VAE appears and the model completes final image decoding. |
| CLIP / T5 text encoder | models/text_encoder | VAE / Clip selector | Every required encoder is selectable and the prompt-encoding stage completes. |
| LoRA | models/Lora | LoRA Extra Networks inventory | The card appears, then a fixed-seed A/B test shows the expected effect. |
| Textual inversion embedding | embeddings | Textual Inversion Extra Networks inventory | A compatible embedding appears and can be applied to its matching model family. |
| ControlNet model | models/ControlNet | ControlNet model selector | The unit lists the model and a minimal ControlNet generation completes. |
| Upscaler | models/<upscaler family> | Extras / upscale selector | The intended upscaler name appears and processes a small known image. |
| Hypernetwork | models/hypernetworks | Hypernetwork inventory | The hypernetwork appears and can be applied to a compatible base model. |
Refreshing answers “can the scanner find it?” A controlled generation answers “can this exact file set load and run?”
Compare the filename, expected size, and publisher hash when available. Keep every required companion file from the same documented version.
Checkpoint, VAE, Text Encoder, LoRA/Textual Inversion, and ControlNet expose separate discovery paths. Refreshing one does not certify the others.
If only part of a split model appears, stop. A selectable diffusion model does not prove its VAE and encoders are present.
Use the matching preset and model-publisher settings. Keep optional LoRAs, ControlNet, hires passes, and extensions off unless the file under test requires them.
If a file appears but fails, stop moving it between folders. Record the earliest console error and verify architecture, version, required components, and hash. Inspect the checkpoint scanner ↗
Path arguments do not all behave as one universal “extra models folder” setting.
--data-dirChanges the broad data base used for default models, embeddings, outputs, extensions, and other user data.
--models-dirReplaces the models root from which many model-family defaults are derived. Embeddings remain outside it.
--ckpt-dirAdds a recursively scanned checkpoint directory. --ckpt can add one specific checkpoint file.
--vae-dir · --text-encoder-dirAdd their configured directories to the corresponding module inventories in the inspected code.
--lora-dir · --embeddings-dirSet the effective root for those inventories. Do not expect the standalone default to remain active automatically.
--controlnet-dirSets the ControlNet weights root. The separate preprocessor-cache argument controls a different directory.
--forge-ref-a1111-home [path] maps existing A1111 inventories, while explicit directory arguments take precedence. Verify every mapped folder independently; partial sharing is possible.
Do not collapse a split model set into one folder because the files share a model name.
models/Stable-diffusionPackaged checkpoint, raw diffusion model, or GGUF according to the exact guide.
models/VAEThe VAE named for that exact distribution.
models/text_encoderRequired CLIP-L and/or T5 encoders.
DATED OFFICIAL GUIDANCE The official FLUX split-file instructions document this three-inventory layout. Packaged NF4/FP8 variants can differ. Open the exact guide ↗
Work from root to inventory to file before changing performance flags or reinstalling Forge.
Confirm the running installation, active data/model root, models/Stable-diffusion path, supported extension, finished download, and Checkpoint refresh. If --ckpt-dir or --ckpt is present, resolve it exactly.
Create models/Lora under the active model root if needed. Preserve capitalization on Linux, verify --lora-dir, refresh Extra Networks, and inspect the first loader message.
Check file role, accepted extension, full companion set, and the specific module refresh. A checkpoint appearing does not prove that a separate VAE, CLIP-L, or T5 was routed correctly.
Verify that model weights are in the effective models/ControlNet path—not only a preprocessor in ControlNetPreprocessor. Refresh ControlNet and read the first file/state-dict error.
Check [data]/embeddings, not [models]/embeddings. Then verify --data-dir and --embeddings-dir. The loader can skip an embedding that does not match the loaded model family.
Check every mapping separately. A1111 reference mapping only uses paths that exist, explicit arguments take precedence, and the standalone/current versus historical shared LoRA directory can differ in capitalization.
Folder placement is complete; compatibility diagnosis starts now. Match architecture and version, confirm companion files and hash, then preserve the earliest console error. Do not treat discovery as support.
Choose and verify the model family →Each answer starts with the current original Forge defaults and names the condition that can change them.
By default, model inventories start under models in the active Forge data directory. First identify the running installation and any --data-dir or --models-dir argument; an absolute path copied from another guide is not evidence of your active root.
Put checkpoint or primary diffusion-model files in models/Stable-diffusion, unless --ckpt-dir, --ckpt, an A1111 reference home, or another configured path changes the checkpoint inventory.
The standalone default is models/Lora with a capital L. --lora-dir replaces that inventory path. Subfolders are scanned.
Create models/Lora inside the active model root, preserving the capital L. Confirm that you are editing the installation that actually launched and that --models-dir or --lora-dir did not point elsewhere.
Use models/VAE for separate VAE files. Forge may also discover a same-name .vae file beside its checkpoint. Use only the VAE specified for the exact model.
For the split-file layout in official Forge guidance, put them in models/text_encoder. Packaged checkpoints may use a different layout, so follow the documentation for the exact distribution.
Use embeddings at the data-directory level—not models/embeddings. --models-dir does not move it; --data-dir or --embeddings-dir can.
Use models/ControlNet by default. Do not confuse this with models/ControlNetPreprocessor, which is the separate cache for preprocessor models.
Use the directory for the upscaler family, such as models/ESRGAN, models/RealESRGAN, models/BSRGAN, models/DAT, models/HAT, models/ScuNET, or models/SwinIR.
Yes for the principal checkpoint, LoRA, VAE, text-encoder, embedding, and ControlNet inventories inspected here. Keep roles separated even when recursive scanning works.
Use the refresh control owned by that inventory after copying a file. Restart Forge when you change launch arguments, a base directory, or when the inventory has no working refresh action.
Verify the active data and model roots, exact filename extension, completed download, permissions, and refresh action. Then inspect the first console message about the file.
Check models/Lora capitalization, the effective --lora-dir, supported extension, subfolder permissions, and the LoRA inventory refresh. A shared checkpoint path does not prove the LoRA path is shared.
Confirm that the file role, folder, and extension match, then refresh the specific module. For split models, verify that every companion file named by the publisher was downloaded.
Confirm that weights—not only a preprocessor or config—are in the effective ControlNet directory. Refresh ControlNet and read the first file or state-dict error in the console.
It replaces the general models root used to derive many default model subdirectories. It does not move the separate embeddings directory.
--data-dir changes the broader user-data base, including default models, embeddings, outputs, and extensions. --models-dir changes the model root only.
Yes. The official --forge-ref-a1111-home helper conditionally maps existing A1111 checkpoint, VAE, hypernetwork, embedding, LoRA, ControlNet, and preprocessor paths. Explicit directory arguments win.
Do not assume one root maps correctly because the folder layouts differ. Configure the relevant Forge inventory paths individually or use filesystem links only if you can maintain and verify every target.
Discovery only proves that the scanner found a candidate file. Verify architecture, exact version, companion files, hash, and the earliest console error; visibility is not compatibility.
We separate current path behavior from dated guidance and reported symptoms.
Data/model roots, inventory defaults, supported discovery extensions, refresh controls, and path-argument behavior.
FLUX split-file and shared A1111 setup explain exact published layouts, not a timeless compatibility promise.
Empty checkpoint lists, absent LoRA cards, ControlNet None, and skipped embeddings define diagnostic questions—not universal causes.
Architecture compatibility, file integrity, licensing, adequate hardware, and output quality still require their own checks.
Current original Forge code and primary project discussions were cross-checked against five locally catalogued video transcripts covering checkpoints, ControlNet, embeddings, split text encoders, LoRAs, and shared storage. Videos informed user language; code controls every path claim.