Do not move the same file through random folders. First prove that Forge discovered it, then that the loader recognized it, and finally that one controlled generation used it.
Evidence boundary: product behavior is verified against original Forge commit dfdcbab. We did not run your files on a GPU; compatibility for an untested artifact remains UNKNOWN until its controlled job passes.
A listed file can still be wrong, and a clean load can still fail to prove a LoRA effect. Stop as soon as one acceptance condition is missing.
Resolve the active path, accepted suffix, physical file and owning Refresh action.
Use the first state-dict, recognition, corruption or allocation message—not the later browser disconnect.
Prove the base, exact adapter, console load and fixed-seed A/B result before judging strength.
Choose the earliest statement that matches. The result names one action and one acceptance condition.
Confirm the installation that is actually running, then inspect its effective checkpoint root. Current Forge lists .ckpt, .safetensors and .gguf files recursively and excludes files named as external VAE companions.
VAE and text encoders share one multiselect but live in different default roots. Forge scans models/VAE and models/text_encoder, plus explicit --vae-dir and --text-encoder-dir locations.
The standalone default is models/Lora with a capital L. Current LoRA discovery accepts .pt, .ckpt and .safetensors recursively. A shared checkpoint path does not prove that the separate LoRA path resolved.
Visibility proves only filename discovery. The file can still be incomplete, unreadable, the wrong component, or an architecture Forge does not recognize at this commit.
A card can be visible even when the adapter does not match the loaded family. Use the card-inserted alias, any documented trigger, one compatible base and a fixed-seed A/B pair while watching the LoRA console lines.
A crash during checkpoint load, T5 load or LoRA patching is not evidence that the artifact is incompatible. System RAM, swap and VRAM can fail before the format test completes.
Each gate proves only one claim. Do not skip from a visible filename to “Forge used it.”
Path, suffix, permissions and Refresh passed.
The UI and console name the exact artifact.
No corruption or container-read error.
Required model, VAE, CLIP, T5 or LoRA keys map.
Image completes or the LoRA causes an attributable A/B change.
Defaults apply to a standalone installation. Launch arguments and shared-folder helpers can change the effective location.
| Artifact role | Standalone root | Type override | Listed suffixes | Refresh action |
|---|---|---|---|---|
Checkpoint / diffusion model | models/Stable-diffusion | --ckpt-dir | .ckpt · .safetensors · .gguf | Top Refresh |
VAE | models/VAE | --vae-dir | .ckpt · .pt · .bin · .safetensors · .gguf | Top Refresh |
CLIP / T5 text encoder | models/text_encoder | --text-encoder-dir | .ckpt · .pt · .bin · .safetensors · .gguf | Top Refresh |
LoRA | models/Lora | --lora-dir | .pt · .ckpt · .safetensors | Extra Networks → Lora → Refresh |
--models-dir overrides the models root derived from --data-dir.
--ckpt-dir, --vae-dir, --text-encoder-dir and --lora-dir resolve separately.
The helper maps only existing directories and does not replace an explicit argument already present.
This sequence works for a missing card, a rejected checkpoint and an ignored LoRA because it preserves the first failed stage.
Record the process folder, original repository URL, commit and complete launch arguments. Do not repair a second portable copy by accident.
Decide whether the file is a diffusion checkpoint, VAE, text encoder or LoRA from its publisher page. The .safetensors suffix does not identify the role.
Start with --models-dir and --data-dir, then check the type-specific override. For a shared A1111 setup, verify each mapped directory independently.
Confirm the final filename, byte size, read permission and publisher hash when one exists. Remove a partial browser-download suffix before expecting discovery.
Use the top Refresh for checkpoints, VAE and text encoders. Use the Lora Extra Networks Refresh for LoRAs. A browser reload alone is not the same inventory action.
Generate once with the base package only. Preserve StateDict Keys and the first recognition, missing-component, corruption or allocation message.
After the base passes, add one VAE replacement or one LoRA. Keep prompt, seed, size, sampler, scheduler and every other setting fixed.
Every branch ends when the filename appears. Loading and output are later gates.
--ckpt-dir.models/Stable-diffusion..ckpt, .safetensors or .gguf.Accept when: the exact filename appears.
Accept when: the intended component can be selected.
--lora-dir.models/Lora case.Accept when: one traceable card appears.
Later connection messages often describe a process that already failed. Preserve the complete traceback before restarting.
Forge found no selectable checkpoint when generation requested one. Verify the active models/Stable-diffusion and --ckpt-dir roots, an accepted checkpoint suffix, permissions, hidden-folder settings and Refresh.
Confirm it is not a VAE companion named *.vae.ckpt or *.vae.safetensors, which current checkpoint discovery excludes. Check for a broken symlink and whether the effective root differs from the folder you edited.
Use models/VAE and models/text_encoder in the active model root, or their explicit directory overrides. The current scanner is recursive, but duplicate basenames can resolve to only one physical file in the combined list.
Treat the specific artifact as incomplete or unreadable until proven otherwise. Compare its byte size or hash with the publisher source and download it again. Do not replace core packages first when every other model still loads.
The official Forge troubleshooting announcement groups this with corrupted model files. Test a known-good artifact, preserve the failing file URL and hash, and re-download from its original publisher.
The diffusion file was readable enough to reach architecture detection but did not match a supported model guess. Verify that Checkpoint is not a VAE, encoder, LoRA, partial export or a model intended for another Forge fork.
The selected package did not provide usable diffusion weights for the expected component. Recheck the exact checkpoint artifact and package manifest; adding random encoders cannot supply the missing diffusion model.
The package expects VAE weights and none were recognized. Select the publisher-specified VAE, not merely a file whose name contains VAE, and confirm the file appears in VAE / Text Encoder after Refresh.
A required CLIP text encoder was absent or the selected file had the wrong architecture. Current original Forge FLUX code expects CLIP-L plus T5-XXL; CLIP-G and an IP-Adapter CLIP Vision file do not satisfy that role.
The model package expects T5 and Forge did not recognize usable T5 weights. Verify the documented T5-XXL artifact, its folder and its selection; a CLIP file cannot substitute for T5.
At the inspected commit, more than 12 adapter keys remained unmatched, so Forge returns without applying that adapter to the model/CLIP pair. Keep the exact adapter page and direct file URL for a report.
Forge continued because 12 or fewer keys remained unmatched. This is not a clean load certificate; save the key list and use the controlled A/B result to determine whether the intended effect survived.
The timing can implicate decode, precision or memory, but is not proof. Remove LoRA and optional replacements, test the package VAE with the same seed, then continue to the black-image diagnostic.
Do not mix this diagnosis with new model files or extension updates. First reproduce the same artifact and settings on a clean current install, then compare the last good and first bad Forge commits.
Use the dedicated LoRA guide for installation details. Here the task is to attribute one change after a working base has passed.
Compatible checkpoint, fixed prompt, seed, size, sampler, scheduler and steps.
Card-inserted alias, publisher trigger and documented starting multiplier.
A successful [LORA] Loaded line and a repeatable visual change.
Loader boundary: more than 12 unmatched keys triggers LoRA version mismatch and returns the unchanged model/CLIP pair at the inspected commit. Smaller unmatched sets continue, so preserve the key list.
The checklist stays in this tab only. Nothing is uploaded or stored by the page.
Complete the receipt before changing packages, drivers or several paths at once.
These answers cover the audited question set without creating separate pages for every traceback wording.
The usual discovery causes are the wrong running installation, wrong effective folder, unsupported suffix, unfinished file, permission problem, hidden directory or an inventory that has not been refreshed. Prove discovery before testing compatibility.
The standalone default is models/Stable-diffusion. An explicit --ckpt-dir adds another checkpoint root, while --models-dir changes the base models root.
At the inspected commit, checkpoint discovery lists .ckpt, .safetensors and .gguf recursively. It excludes external VAE companions named *.vae.ckpt and *.vae.safetensors.
Yes. Current checkpoint, VAE/text-encoder and LoRA scans are recursive. Hidden directories can be excluded by the list-hidden-files setting.
Start with the top Refresh control, which rebuilds checkpoint and VAE/text-encoder choices at the inspected commit. Restart only after the active path and inventory action are verified.
F5 reloads the browser document. Use Forge’s inventory Refresh control so the server scans model files again.
Check the active models/VAE or --vae-dir root, accepted suffix, completed filename and unique basename, then use the top Refresh control.
Check models/text_encoder or --text-encoder-dir in the running installation. Then press top Refresh and select the file in VAE / Text Encoder.
Forge uses one multiselect for additional modules. The files still perform different roles and live in separate default folders.
The expected CLIP component did not receive a recognized state dictionary. The file may be absent, unselected, incomplete or the wrong CLIP architecture.
The inspected original Forge FLUX engine maps CLIP-L and T5-XXL. CLIP-G does not fill that FLUX text-encoder role.
The selected package expects T5 weights and none were recognized. Verify the exact publisher-specified T5-XXL artifact and selection.
The package did not provide a recognized VAE state dictionary. Verify the required VAE artifact rather than renaming another file.
The diffusion artifact did not match a model architecture recognized by that Forge commit. It may be the wrong file role, an unsupported architecture or a malformed export.
Compare its byte size or SHA-256 with the publisher source. Container-read errors such as MetadataIncompleteBuffer or PytorchStreamReader failures are strong reasons to re-download that artifact.
No. Renaming changes the label, not the internal tensor keys, shapes or architecture. Preserve the publisher filename or keep a traceable local inventory name.
Verify the active --lora-dir or standalone models/Lora directory, including capital L on case-sensitive systems. Current LoRA discovery accepts .pt, .ckpt and .safetensors recursively.
Checkpoint and LoRA directories resolve independently. The --forge-ref-a1111-home helper maps only paths that exist, and an explicit type-specific argument is not overwritten.
Open Extra Networks, choose Lora and use its Refresh action. A visible card confirms inventory discovery only.
Use the card-inserted alias, documented trigger and compatible base. Compare the same seed with and without the tag while checking the console for Loaded, mismatch, unmatched or skipped-key messages.
At the inspected commit, it means more than 12 adapter keys remained unmatched. Forge returns the unchanged model and text encoder for that adapter attempt.
No. The LoRA tag requests the adapter; trigger text only steers conditioning when the publisher documents it.
First treat it as a resource-stage failure. Record whether the crash occurs during checkpoint load, T5 load, LoRA patching, sampling or VAE decode, then run the dedicated memory diagnostic.
Include original Forge commit, launch arguments, hardware, publisher page, direct file URL, filename, size/hash, base family, exact selections, fixed job, first loader line and full traceback.
Tutorials and issues help us understand user language. They do not override the inspected original Forge implementation.
Paths, suffixes, refresh actions, component assertions and LoRA thresholds are tied to commit dfdcbab.
Issue examples demonstrate real symptoms but do not prove a universal cause or fix.
Articles and full transcripts supplied vocabulary and task order; old settings were not promoted to current rules.
Compatibility remains unknown until the exact source, hash, package and controlled result are recorded.
Default root, accepted checkpoint suffixes and VAE companion exclusion.
↗Recursive roots, broken-symlink skip and discovery boundary.
↗Case-normalized extension checks and hidden-directory behavior.
↗Checkpoint/component Refresh, roots, suffixes and basename indexing.
↗Data, model and type-specific directory overrides.
↗Per-directory existence checks and explicit-argument precedence.
↗Missing model, VAE, CLIP and T5 state-dict assertions.
↗CLIP-L + T5-XXL + VAE role mapping.
↗Accepted suffixes, alias resolution and mismatch thresholds.
↗Official 2024 report format and patching behavior; dated low-bit recipes not generalized.
↗Official VAE/text-encoder paths and multiselect workflow.
↗Corrupt-model and memory-stage messages.
↗Common folder and missing-component questions; recipes treated as dated.
↗Community setup vocabulary and split-package workflow.
↗Subfolder organization and user confusion between file roles.
↗Checkpoint placement and Refresh workflow in an early Forge build.
↗Model-family and LoRA discovery vocabulary from a 2025 guide.
↗Beginner checkpoint/LoRA placement and refresh sequence.
↗Shared checkpoints worked while the separate LoRA inventory remained empty.
↗2025 report showing CLIP-G/CLIP Vision confusion around a FLUX CLIP error.
↗2026 traceback for architecture-recognition failure.
↗