From bbe9aa7f3e16ba320b3eaff2e0b8c7ddeaa58b41 Mon Sep 17 00:00:00 2001
From: Jayce-Ping <315229706@qq.com>
Date: Fri, 23 Jan 2026 11:38:18 +0800
Subject: [PATCH 1/2] First commit

---
 inference/example_full.py | 30 ++++++++++++++++++++++++++++++
 inference/example_lora.py | 33 +++++++++++++++++++++++++++++++++
 2 files changed, 63 insertions(+)
 create mode 100644 inference/example_full.py
 create mode 100644 inference/example_lora.py

diff --git a/inference/example_full.py b/inference/example_full.py
new file mode 100644
index 00000000..adf19d78
--- /dev/null
+++ b/inference/example_full.py
@@ -0,0 +1,30 @@
+# Modified from https://huggingface.co/black-forest-labs/FLUX.1-dev usage example
+import torch
+from diffusers import FluxPipeline, FluxTransformer2DModel
+
+# 1. Load pipeline
+pipe = FluxPipeline.from_pretrained(
+    "black-forest-labs/FLUX.1-dev",
+    torch_dtype=torch.bfloat16
+)
+pipe.enable_model_cpu_offload()
+
+# 2. Load full fine-tuned transformer weights
+checkpoint = 'path/to/full/checkpoint'  # replace with your checkpoint directory
+pipe.transformer = FluxTransformer2DModel.from_pretrained(
+    checkpoint,
+    torch_dtype=torch.bfloat16
+)
+
+# 3. Generate image
+prompt = "A cat holding a sign that says hello world"
+image = pipe(
+    prompt,
+    height=1024,
+    width=1024,
+    guidance_scale=3.5,
+    num_inference_steps=50,
+    max_sequence_length=512,
+    generator=torch.Generator("cpu").manual_seed(0)
+).images[0]
+image.save("flux-dev-full.png")
\ No newline at end of file
diff --git a/inference/example_lora.py b/inference/example_lora.py
new file mode 100644
index 00000000..36cc054c
--- /dev/null
+++ b/inference/example_lora.py
@@ -0,0 +1,33 @@
+# Modified from https://huggingface.co/black-forest-labs/FLUX.1-dev usage example
+import torch
+from diffusers import FluxPipeline
+from peft import PeftModel
+
+
+# 1. Load pipeline
+pipe = FluxPipeline.from_pretrained(
+    "black-forest-labs/FLUX.1-dev",
+    torch_dtype=torch.bfloat16
+)
+pipe.enable_model_cpu_offload() #save some VRAM by offloading the model to CPU. Remove this if you have enough GPU power
+
+# 2. Load LoRA weights
+checkpoint = 'path/to/lora/checkpoint'  # replace with your checkpoint directory
+pipe.transformer = PeftModel.from_pretrained(
+    pipe.transformer,
+    checkpoint,
+    torch_dtype=torch.bfloat16
+)
+
+# 3. Generate image
+prompt = "A cat holding a sign that says hello world"
+image = pipe(
+    prompt,
+    height=1024,
+    width=1024,
+    guidance_scale=3.5,
+    num_inference_steps=50,
+    max_sequence_length=512,
+    generator=torch.Generator("cpu").manual_seed(0)
+).images[0]
+image.save("flux-dev.png")

From a48fb0ec19e32719f3c857180be7a3d46a44e16f Mon Sep 17 00:00:00 2001
From: Jayce-Ping <315229706@qq.com>
Date: Fri, 23 Jan 2026 12:37:27 +0800
Subject: [PATCH 2/2] update

---
 inference/example_full.py | 7 ++++---
 inference/example_lora.py | 8 ++++----
 2 files changed, 8 insertions(+), 7 deletions(-)

diff --git a/inference/example_full.py b/inference/example_full.py
index adf19d78..e7c0ce1b 100644
--- a/inference/example_full.py
+++ b/inference/example_full.py
@@ -7,15 +7,16 @@
     "black-forest-labs/FLUX.1-dev",
     torch_dtype=torch.bfloat16
 )
-pipe.enable_model_cpu_offload()
 
 # 2. Load full fine-tuned transformer weights
-checkpoint = 'path/to/full/checkpoint'  # replace with your checkpoint directory
+checkpoint = 'path/to/checkpoint'  # replace with your checkpoint directory
 pipe.transformer = FluxTransformer2DModel.from_pretrained(
     checkpoint,
     torch_dtype=torch.bfloat16
 )
 
+pipe.enable_model_cpu_offload()
+
 # 3. Generate image
 prompt = "A cat holding a sign that says hello world"
 image = pipe(
@@ -23,7 +24,7 @@
     height=1024,
     width=1024,
     guidance_scale=3.5,
-    num_inference_steps=50,
+    num_inference_steps=28,
     max_sequence_length=512,
     generator=torch.Generator("cpu").manual_seed(0)
 ).images[0]
diff --git a/inference/example_lora.py b/inference/example_lora.py
index 36cc054c..7746368f 100644
--- a/inference/example_lora.py
+++ b/inference/example_lora.py
@@ -3,22 +3,22 @@
 from diffusers import FluxPipeline
 from peft import PeftModel
 
-
 # 1. Load pipeline
 pipe = FluxPipeline.from_pretrained(
     "black-forest-labs/FLUX.1-dev",
     torch_dtype=torch.bfloat16
 )
-pipe.enable_model_cpu_offload() #save some VRAM by offloading the model to CPU. Remove this if you have enough GPU power
 
 # 2. Load LoRA weights
-checkpoint = 'path/to/lora/checkpoint'  # replace with your checkpoint directory
+checkpoint = 'path/to/checkpoint'  # replace with your checkpoint directory
 pipe.transformer = PeftModel.from_pretrained(
     pipe.transformer,
     checkpoint,
     torch_dtype=torch.bfloat16
 )
 
+pipe.enable_model_cpu_offload() #save some VRAM by offloading the model to CPU. Remove this if you have enough GPU power
+
 # 3. Generate image
 prompt = "A cat holding a sign that says hello world"
 image = pipe(
@@ -26,7 +26,7 @@
     height=1024,
     width=1024,
     guidance_scale=3.5,
-    num_inference_steps=50,
+    num_inference_steps=28,
     max_sequence_length=512,
     generator=torch.Generator("cpu").manual_seed(0)
 ).images[0]