Skip to content

Fixing Kohya loras loading: Flux.1-dev loras with TE ("lora_te1_" prefix) + Flux.2-dev loras #13118

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
christopher5106 wants to merge 3 commits into
base: main
Choose a base branch
from
+71 −22
Open

Fixing Kohya loras loading: Flux.1-dev loras with TE ("lora_te1_" prefix) + Flux.2-dev loras #13118

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
8d0c1ec
fixing text encoder lora loading
christopher5106 Feb 10, 2026
cf3c65c
following Cursor's review
christopher5106 Feb 10, 2026
5518d07
fix for loading of Kohya's Flux.2 dev lora
christopher5106 Feb 11, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
83 changes: 61 additions & 22 deletions src/diffusers/loaders/lora_conversion_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -358,7 +358,10 @@ def _get_alpha_name(lora_name_alpha, diffusers_name, alpha):

# The utilities under `_convert_kohya_flux_lora_to_diffusers()`
# are adapted from https://github.com/kohya-ss/sd-scripts/blob/a61cf73a5cb5209c3f4d1a3688dd276a4dfd1ecb/networks/convert_flux_lora.py
def _convert_kohya_flux_lora_to_diffusers(state_dict):
def _convert_kohya_flux_lora_to_diffusers(
state_dict,
version_flux2 = False,
):
def _convert_to_ai_toolkit(sds_sd, ait_sd, sds_key, ait_key):
if sds_key + ".lora_down.weight" not in sds_sd:
return
Expand Down Expand Up @@ -449,7 +452,15 @@ def _convert_to_ai_toolkit_cat(sds_sd, ait_sd, sds_key, ait_keys, dims=None):

def _convert_sd_scripts_to_ai_toolkit(sds_sd):
ait_sd = {}
for i in range(19):

max_num_double_blocks, max_num_single_blocks = -1, -1
for key in list(sds_sd.keys()):
if key.startswith("lora_unet_double_blocks_"):
max_num_double_blocks = max(max_num_double_blocks, int(key.split("_")[4]))
if key.startswith("lora_unet_single_blocks_"):
max_num_single_blocks = max(max_num_single_blocks, int(key.split("_")[4]))

for i in range(max_num_double_blocks+1):
_convert_to_ai_toolkit(
sds_sd,
ait_sd,
Expand All @@ -470,13 +481,21 @@ def _convert_sd_scripts_to_ai_toolkit(sds_sd):
sds_sd,
ait_sd,
f"lora_unet_double_blocks_{i}_img_mlp_0",
f"transformer.transformer_blocks.{i}.ff.net.0.proj",
(
f"transformer.transformer_blocks.{i}.ff.linear_in"
if version_flux2 else
f"transformer.transformer_blocks.{i}.ff.net.0.proj"
),
)
_convert_to_ai_toolkit(
sds_sd,
ait_sd,
f"lora_unet_double_blocks_{i}_img_mlp_2",
f"transformer.transformer_blocks.{i}.ff.net.2",
(
f"transformer.transformer_blocks.{i}.ff.linear_out"
if version_flux2 else
f"transformer.transformer_blocks.{i}.ff.net.2"
),
)
_convert_to_ai_toolkit(
sds_sd,
Expand Down Expand Up @@ -504,13 +523,21 @@ def _convert_sd_scripts_to_ai_toolkit(sds_sd):
sds_sd,
ait_sd,
f"lora_unet_double_blocks_{i}_txt_mlp_0",
f"transformer.transformer_blocks.{i}.ff_context.net.0.proj",
(
f"transformer.transformer_blocks.{i}.ff_context.linear_in"
if version_flux2 else
f"transformer.transformer_blocks.{i}.ff_context.net.0.proj"
),
)
_convert_to_ai_toolkit(
sds_sd,
ait_sd,
f"lora_unet_double_blocks_{i}_txt_mlp_2",
f"transformer.transformer_blocks.{i}.ff_context.net.2",
(
f"transformer.transformer_blocks.{i}.ff_context.linear_out"
if version_flux2 else
f"transformer.transformer_blocks.{i}.ff_context.net.2"
),
)
_convert_to_ai_toolkit(
sds_sd,
Expand All @@ -519,24 +546,36 @@ def _convert_sd_scripts_to_ai_toolkit(sds_sd):
f"transformer.transformer_blocks.{i}.norm1_context.linear",
)

for i in range(38):
_convert_to_ai_toolkit_cat(
sds_sd,
ait_sd,
f"lora_unet_single_blocks_{i}_linear1",
[
f"transformer.single_transformer_blocks.{i}.attn.to_q",
f"transformer.single_transformer_blocks.{i}.attn.to_k",
f"transformer.single_transformer_blocks.{i}.attn.to_v",
f"transformer.single_transformer_blocks.{i}.proj_mlp",
],
dims=[3072, 3072, 3072, 12288],
)
for i in range(max_num_single_blocks+1):
if version_flux2:
_convert_to_ai_toolkit(
sds_sd,
ait_sd,
f"lora_unet_single_blocks_{i}_linear1",
f"transformer.single_transformer_blocks.{i}.attn.to_qkv_mlp_proj",
)
else:
_convert_to_ai_toolkit_cat(
sds_sd,
ait_sd,
f"lora_unet_single_blocks_{i}_linear1",
[
f"transformer.single_transformer_blocks.{i}.attn.to_q",
f"transformer.single_transformer_blocks.{i}.attn.to_k",
f"transformer.single_transformer_blocks.{i}.attn.to_v",
f"transformer.single_transformer_blocks.{i}.proj_mlp",
],
dims=[3072, 3072, 3072, 12288],
)
_convert_to_ai_toolkit(
sds_sd,
ait_sd,
f"lora_unet_single_blocks_{i}_linear2",
f"transformer.single_transformer_blocks.{i}.proj_out",
(
f"transformer.single_transformer_blocks.{i}.attn.to_out"
if version_flux2 else
f"transformer.single_transformer_blocks.{i}.proj_out"
),
)
_convert_to_ai_toolkit(
sds_sd,
Expand Down Expand Up @@ -857,7 +896,7 @@ def _convert(original_key, diffusers_key, state_dict, new_state_dict):
)
state_dict = {k: v for k, v in state_dict.items() if not k.startswith("text_encoders.t5xxl.transformer.")}

has_diffb = any("diff_b" in k and k.startswith(("lora_unet_", "lora_te_")) for k in state_dict)
has_diffb = any("diff_b" in k and k.startswith(("lora_unet_", "lora_te_", "lora_te1_")) for k in state_dict)
if has_diffb:
zero_status_diff_b = state_dict_all_zero(state_dict, ".diff_b")
if zero_status_diff_b:
Expand Down Expand Up @@ -896,7 +935,7 @@ def _convert(original_key, diffusers_key, state_dict, new_state_dict):
state_dict = {
_custom_replace(k, limit_substrings): v
for k, v in state_dict.items()
if k.startswith(("lora_unet_", "lora_te_"))
if k.startswith(("lora_unet_", "lora_te_", "lora_te1_"))
}
Comment on lines 935 to 939
Copy link

Copilot AI Feb 10, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This change adds support for keeping lora_te1_ keys in the Flux/Kohya conversion path, but there doesn't appear to be a regression test covering loading a non-Diffusers LoRA state dict with lora_te1_-prefixed text-encoder weights (existing LoRA tests don’t mention lora_te1_). Adding a small unit/integration test would help prevent future regressions where TE1 keys get filtered out again.

Copilot uses AI. Check for mistakes.

if any("text_projection" in k for k in state_dict):
Expand Down
10 changes: 10 additions & 0 deletions src/diffusers/loaders/lora_pipeline.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5472,6 +5472,16 @@ def lora_state_dict(
logger.warning(warn_msg)
state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}

is_kohya = any(".lora_down.weight" in k for k in state_dict)
if is_kohya:
state_dict = _convert_kohya_flux_lora_to_diffusers(
state_dict,
version_flux2=True,
)
# Kohya already takes care of scaling the LoRA parameters with alpha.
for k in state_dict:
assert "alpha" not in k, f"Found key with alpha: {k}"

is_ai_toolkit = any(k.startswith("diffusion_model.") for k in state_dict)
if is_ai_toolkit:
state_dict = _convert_non_diffusers_flux2_lora_to_diffusers(state_dict)
Expand Down