DominikDoom/a1111-sd-webui-tagcomplete
1
0
Fork 0
mirror of https://github.com/DominikDoom/a1111-sd-webui-tagcomplete.git synced 2026-02-10 02:00:09 +00:00
Code Issues Packages Projects Releases Wiki Activity

Proper support for SDXL embeddings

Browse Source 
Now in their own category, other embeddings don't get mislabeled anymore if an XL model is loaded
This commit is contained in:
DominikDoom
2023-09-26 14:14:20 +02:00
parent 94ec8884c3
commit 998514bebb
2 changed files with 18 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
javascript/ext_embeddings.js
View File

@@ -11,12 +11,15 @@ class EmbeddingParser extends BaseTagParser {
if (searchTerm.startsWith("v1") || searchTerm.startsWith("v2")) {
versionString = searchTerm.slice(0, 2);
searchTerm = searchTerm.slice(2);
} else if (searchTerm.startsWith("vxl")) {
versionString = searchTerm.slice(0, 3);
searchTerm = searchTerm.slice(3);
}
let filterCondition = x => x[0].toLowerCase().includes(searchTerm) || x[0].toLowerCase().replaceAll(" ", "_").includes(searchTerm);
if (versionString)
tempResults = embeddings.filter(x => filterCondition(x) && x[2] && x[2] === versionString); // Filter by tagword
tempResults = embeddings.filter(x => filterCondition(x) && x[2] && x[2].toLowerCase() === versionString.toLowerCase()); // Filter by tagword
else
tempResults = embeddings.filter(x => filterCondition(x)); // Filter by tagword
} else {

42
scripts/tag_autocomplete_helper.py
View File

@@ -156,44 +156,30 @@ def get_embeddings(sd_model):
# Version constants
V1_SHAPE = 768
V2_SHAPE = 1024
VXL_SHAPE = 2048
emb_v1 = []
emb_v2 = []
emb_vXL = []
results = []
try:
# Get embedding dict from sd_hijack to separate v1/v2 embeddings
emb_type_a = sd_hijack.model_hijack.embedding_db.word_embeddings
emb_type_b = sd_hijack.model_hijack.embedding_db.skipped_embeddings
# Get the shape of the first item in the dict
emb_a_shape = -1
emb_b_shape = -1
if (len(emb_type_a) > 0):
emb_a_shape = next(iter(emb_type_a.items()))[1].shape
if (len(emb_type_b) > 0):
emb_b_shape = next(iter(emb_type_b.items()))[1].shape
loaded = sd_hijack.model_hijack.embedding_db.word_embeddings
skipped = sd_hijack.model_hijack.embedding_db.skipped_embeddings
# Add embeddings to the correct list
if (emb_a_shape == V1_SHAPE):
emb_v1 = [(Path(v.filename), k, "v1") for (k,v) in emb_type_a.items() if v.filename is not None]
elif (emb_a_shape == V2_SHAPE):
emb_v2 = [(Path(v.filename), k, "v2") for (k,v) in emb_type_a.items() if v.filename is not None]
for key, emb in (loaded | skipped).items():
if (emb.filename is None):
continue
if (emb_b_shape == V1_SHAPE):
emb_v1 = [(Path(v.filename), k, "v1") for (k,v) in emb_type_b.items() if v.filename is not None]
elif (emb_b_shape == V2_SHAPE):
emb_v2 = [(Path(v.filename), k, "v2") for (k,v) in emb_type_b.items() if v.filename is not None]
if emb.shape == V1_SHAPE:
emb_v1.append((Path(emb.filename), key, "v1"))
elif emb.shape == V2_SHAPE:
emb_v2.append((Path(emb.filename), key, "v2"))
elif emb.shape == VXL_SHAPE:
emb_vXL.append((Path(emb.filename), key, "vXL"))
# Get shape of current model
#vec = sd_model.cond_stage_model.encode_embedding_init_text(",", 1)
#model_shape = vec.shape[1]
# Show relevant entries at the top
#if (model_shape == V1_SHAPE):
# results = [e + ",v1" for e in emb_v1] + [e + ",v2" for e in emb_v2]
#elif (model_shape == V2_SHAPE):
# results = [e + ",v2" for e in emb_v2] + [e + ",v1" for e in emb_v1]
#else:
# raise AttributeError # Fallback to old method
results = sort_models(emb_v1) + sort_models(emb_v2)
results = sort_models(emb_v1) + sort_models(emb_v2) + sort_models(emb_vXL)
except AttributeError:
print("tag_autocomplete_helper: Old webui version or unrecognized model shape, using fallback for embedding completion.")
# Get a list of all embeddings in the folder