AUTOMATIC1111/stable-diffusion-webui · error · Exception

Couldn't identify {filename} as neither textual inversion em

Error message

Couldn't identify {filename} as neither textual inversion embedding nor diffuser concept.

What it means

textual_inversion.py's embedding loader inspects the torch-loaded object to decide what it is: SD1.x embeddings (string->tensor dict), SD2/klg, clip_g/clip_l pairs, or single-tensor diffuser concepts. If the data matches none of these shapes (not a dict of tensors, not the expected keys), it raises this 'couldn't identify' error before constructing the Embedding.

Source

Thrown at modules/textual_inversion/textual_inversion.py:310

        emb = next(iter(param_dict.items()))[1]
        vec = emb.detach().to(devices.device, dtype=torch.float32)
        shape = vec.shape[-1]
        vectors = vec.shape[0]
    elif type(data) == dict and 'clip_g' in data and 'clip_l' in data:  # SDXL embedding
        vec = {k: v.detach().to(devices.device, dtype=torch.float32) for k, v in data.items()}
        shape = data['clip_g'].shape[-1] + data['clip_l'].shape[-1]
        vectors = data['clip_g'].shape[0]
    elif type(data) == dict and type(next(iter(data.values()))) == torch.Tensor:  # diffuser concepts
        assert len(data.keys()) == 1, 'embedding file has multiple terms in it'

        emb = next(iter(data.values()))
        if len(emb.shape) == 1:
            emb = emb.unsqueeze(0)
        vec = emb.detach().to(devices.device, dtype=torch.float32)
        shape = vec.shape[-1]
        vectors = vec.shape[0]
    else:
        raise Exception(f"Couldn't identify {filename} as neither textual inversion embedding nor diffuser concept.")

    embedding = Embedding(vec, name)
    embedding.step = data.get('step', None)
    embedding.sd_checkpoint = data.get('sd_checkpoint', None)
    embedding.sd_checkpoint_name = data.get('sd_checkpoint_name', None)
    embedding.vectors = vectors
    embedding.shape = shape

    if filepath:
        embedding.filename = filepath
        embedding.set_hash(hashes.sha256(filepath, "textual_inversion/" + name) or '')

    return embedding


def write_loss(log_directory, filename, step, epoch_len, values):
    if shared.opts.training_write_csv_every == 0:
        return

View on GitHub (pinned to 82a973c043)

Solutions

1. Verify the file is genuinely a textual-inversion embedding (small, usually < 1 MB, contains '<concept>' string keys mapping to tensors)
2. Re-download the embedding from its original source; compare file size/hash against the publisher
3. If it is a Kohya-style file, convert it first or use a trainer/tool that emits the dict-of-tensors format
4. Inspect locally: d = torch.load(f, map_location='cpu'); print(type(d), list(d)[:5]) — the loader needs dict values of torch.Tensor

Example fix

# before: passing a LoRA/unknown .pt into the embedding loader
# after: check the shape before loading
import torch
d = torch.load(path, map_location='cpu')
if not (isinstance(d, dict) and any(isinstance(v, torch.Tensor) for v in d.values())):
    raise SystemExit(f'{path} is not a textual-inversion embedding')

import torch
def looks_like_embedding(path):
    d = torch.load(path, map_location='cpu')
    if not isinstance(d, dict):
        return False
    vals = list(d.values())
    return len(vals) > 0 and all(isinstance(v, torch.Tensor) for v in vals[:3]) or 'clip_g' in d

def is_ti_embedding(data) -> bool:
    if not isinstance(data, dict):
        return False
    if 'clip_g' in data and 'clip_l' in data:
        return True
    vals = list(data.values())
    return bool(vals) and isinstance(vals[0], torch.Tensor)

try:
    ti_manager.load_from_file(path)
except Exception as e:
    if 'Could not identify' in str(e) or "Couldn't identify" in str(e):
        log.warning('skipping non-embedding file %s', path)
    else:
        raise

When it happens

Trigger: Loading a .pt/.bin file through the train/embedding tab whose pickled content is e.g. a raw state dict with unexpected keys, a whole model object, an empty dict, or a numpy array instead of torch.Tensor; also truncated downloads that unpickle to garbage.

Common situations: Downloading a LoRA or full checkpoint and renaming it .pt as if it were an embedding; embeddings saved by incompatible forks (Kohya, old A1111); partial/interrupted downloads.

Related errors

• Invalid learning rate schedule. It should be a number or, fo
• Unknown checkpoint: {x}
• Lora layer {self.network_key} matched a layer with unsupport
• Could not find a module type (out of {', '.join([x.__class__
• Sampler not found

AI-assisted analysis of AUTOMATIC1111/stable-diffusion-webui@82a973c043 (2026-08-14). Data as JSON: /api/errors/20cc3df00b113b95. Report an issue: GitHub.