Add Qwen Image Edit Plus

config/examples/temp_examples_qwen_edit_plus/qwen_image_edit_plus_2controls.yaml

@@ -0,0 +1,93 @@
+---
+job: "extension"
+config:
+  name: "my_first_qwen_image_edit_plus2509_lora_v3"
+  process:
+    - type: "sd_trainer"
+      training_folder: "/home/ubuntu/ai-toolkit/output"
+      device: "cuda:0"
+      network:
+        type: "lora"
+        linear: 4
+        linear_alpha: 4
+      save:
+        dtype: "float16"
+        save_every: 500
+        max_step_saves_to_keep: 12
+      datasets:
+        - folder_path: "/home/ubuntu/ai-toolkit/datasets/end"
+          control_path: "/home/ubuntu/ai-toolkit/datasets/start"
+          control_path2: "/home/ubuntu/ai-toolkit/datasets/reference"
+          full_size_control_images: true
+          caption_ext: "txt"
+          caption_dropout_rate: 0.05
+          buckets: true
+          cache_latents_to_disk: true
+          keep_native_size: true
+          random_scale: false
+          random_crop: false
+          square_crop: false
+          scale: 0.1
+      train:
+        batch_size: 1
+        unload_text_encoder: false
+        cache_text_embeddings: true
+        steps: 3000
+        gradient_accumulation: 1
+        timestep_type: "weighted"
+        train_unet: true
+        train_text_encoder: false
+        gradient_checkpointing: false
+        noise_scheduler: "flowmatch"
+        optimizer: "adamw8bit"
+        lr: 0.0001
+        dtype: "bf16"
+      model:
+        name_or_path: "Qwen/Qwen-Image-Edit-2509"
+        arch: "qwen_image_edit_plus2509"
+        quantize: true
+        qtype: "qfloat8"
+        quantize_te: true
+        qtype_te: "qfloat8"
+        low_vram: true
+      sample:
+        sampler: "flowmatch"
+        sample_every: 500
+        width: 1024
+        height: 1024
+        samples:
+          - prompt: "Swap the furniture inside the red bounding box of the Image 1 by the reference furniture of the Image 2"
+            ctrl_img: "/home/ubuntu/ai-toolkit/datasets/start_full/94a31bd5-b0ce-42bc-a12e-a76a676c1314_2_to_1.png"
+            ctrl_img2: "/home/ubuntu/ai-toolkit/datasets/reference_full/94a31bd5-b0ce-42bc-a12e-a76a676c1314_2_to_1.png"
+            width: 1024
+            height: 535
+          # - prompt: "Swap the furniture inside the red bounding box of the Image 1 by the reference furniture of the Image 2"
+          #   width: 796
+          #   height: 1024
+          #   ctrl_img: "/home/ubuntu/ai-toolkit/datasets/start_full/95dabfac-acf1-43cb-a85a-aa795bc52739_1_to_2.png"
+          #   ctrl_img2: "/home/ubuntu/ai-toolkit/datasets/reference_full/95dabfac-acf1-43cb-a85a-aa795bc52739_1_to_2.png"
+          # - prompt: "Swap the furniture inside the red bounding box of the Image 1 by the reference furniture of the Image 2"
+          #   width: 961
+          #   height: 1024
+          #   ctrl_img: "/home/ubuntu/ai-toolkit/datasets/start_full/a3912256-a57f-4254-89df-fd06cdfe1c3c_2_to_1.png"
+          #   ctrl_img2: "/home/ubuntu/ai-toolkit/datasets/reference_full/a3912256-a57f-4254-89df-fd06cdfe1c3c_2_to_1.png"
+          # - prompt: "Swap the furniture inside the red bounding box of the Image 1 by the reference furniture of the Image 2"
+          #   width: 1024
+          #   height: 1024
+          #   ctrl_img: "/home/ubuntu/ai-toolkit/datasets/start_full/a71024fc-89bc-4f57-a1da-50a4b0c4ef3b_2_to_1.png"
+          #   ctrl_img2: "/home/ubuntu/ai-toolkit/datasets/reference_full/a71024fc-89bc-4f57-a1da-50a4b0c4ef3b_2_to_1.png"
+          # - prompt: "Swap the furniture inside the red bounding box of the Image 1 by the reference furniture of the Image 2"
+          #   width: 1024
+          #   height: 489
+          #   ctrl_img: "/home/ubuntu/ai-toolkit/datasets/start_full/a8ddc0e8-be33-4717-9bec-1a99b35593c3_1_to_2.png"
+          #   ctrl_img2: "/home/ubuntu/ai-toolkit/datasets/reference_full/a8ddc0e8-be33-4717-9bec-1a99b35593c3_1_to_2.png"
+        neg: ""
+        seed: 42
+        walk_seed: false
+        guidance_scale: 3
+        sample_steps: 25
+      sqlite_db_path: "./aitk_db.db"
+      performance_log_every: 10
+meta:
+  name: "[name]"
+  version: "1.0"

config/examples/train_lora_omnigen2_24gb.yaml

@@ -61,7 +61,7 @@ config:
         # will probably need this if gpu supports it for omnigen2, other dtypes may not work correctly
         dtype: bf16
       model:
-        name_or_path: "OmniGen2/OmniGen2
+        name_or_path: "OmniGen2/OmniGen2"
         arch: "omnigen2"
         quantize_te: true  # quantize_only te
         # quantize: true  # quantize transformer

config/examples/train_lora_qwen_image_edit_32gb.yaml

@@ -58,12 +58,11 @@ config:
         name_or_path: "Qwen/Qwen-Image-Edit"
         arch: "qwen_image_edit"
         quantize: true
-        # qtype_te: "qfloat8" Default float8 qquantization
+        # qtype_te: "qfloat8" # Default float8 qquantization
         # to use the ARA use the | pipe to point to hf path, or a local path if you have one.
         # 3bit is required for 32GB
         qtype: "uint3|qwen_image_edit_torchao_uint3.safetensors"
         quantize_te: true
-        qtype_te: "qfloat8"
         low_vram: true
       sample:
         sampler: "flowmatch" # must match train.noise_scheduler

config/examples/train_lora_qwen_image_edit_plus2509_32gb_2controls.yaml

@@ -0,0 +1,117 @@
+---
+job: extension
+config:
+  # this name will be the folder and filename name
+  name: "my_first_qwen_image_edit_plus_20lora_v1"
+  process:
+    - type: 'sd_trainer'
+      # root folder to save training sessions/samples/weights
+      training_folder: "output"
+      # uncomment to see performance stats in the terminal every N steps
+#      performance_log_every: 1000
+      device: cuda:0
+      # if a trigger word is specified, it will be added to captions of training data if it does not already exist
+      # alternatively, in your captions you can add [trigger] and it will be replaced with the trigger word
+      # Trigger words will not work when caching text embeddings
+#      trigger_word: "p3r5on"
+      network:
+        type: "lora"
+        linear: 16
+        linear_alpha: 16
+      save:
+        dtype: float16 # precision to save
+        save_every: 250 # save every this many steps
+        max_step_saves_to_keep: 4 # how many intermittent saves to keep
+      datasets:
+        # datasets are a folder of images. captions need to be txt files with the same name as the image
+        # for instance image2.jpg and image2.txt. Only jpg, jpeg, and png are supported currently
+        # images will automatically be resized and bucketed into the resolution specified
+        # on windows, escape back slashes with another backslash so
+        # "C:\\path\\to\\images\\folder"
+        - folder_path: "/path/to/images/folder"
+          control_path: "/path/to/control/images/folder"
+          control_path2: "/path/to/control/images/folder2"
+          # control_path3: "/path/to/control/images/folder3"
+          caption_ext: "txt"
+          # default_caption: "a person" # if caching text embeddings, if you don't have captions, this will get cached
+          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
+          resolution: [ 512, 768, 1024 ]  # qwen image enjoys multiple resolutions
+      train:
+        batch_size: 1
+        # caching text embeddings is required for 32GB
+        cache_text_embeddings: true
+
+        steps: 3000  # total number of steps to train 500 - 4000 is a good range
+        gradient_accumulation: 1
+        timestep_type: "weighted"
+        train_unet: true
+        train_text_encoder: false  # probably won't work with qwen image
+        gradient_checkpointing: true  # need the on unless you have a ton of vram
+        noise_scheduler: "flowmatch" # for training only
+        optimizer: "adamw8bit"
+        lr: 1e-4
+        # uncomment this to skip the pre training sample
+#        skip_first_sample: true
+        # uncomment to completely disable sampling
+#        disable_sampling: true
+        dtype: bf16
+      model:
+        # huggingface model name or path
+        name_or_path: "Qwen/Qwen-Image-Edit-2509"
+        arch: "qwen_image_edit_plus2509"
+        quantize: true
+        # qtype_te: "qfloat8" # Default float8 qquantization
+        # to use the ARA use the | pipe to point to hf path, or a local path if you have one.
+        # 3bit is required for 32GB
+        #! TODO: Quantize and add this to hf
+        # qtype: "uint3|qwen_image_edit_torchao_uint3.safetensors"
+        qtype: "qfloat8"
+        quantize_te: true
+        low_vram: true
+      sample:
+        sampler: "flowmatch" # must match train.noise_scheduler
+        sample_every: 250 # sample every this many steps
+        width: 1024
+        height: 1024
+        samples:
+          - prompt: "do the thing to it"
+            ctrl_img: "/path/to/control/image.jpg"
+            ctrl_img2: "/path/to/control2/image.jpg"
+            # ctrl_img3: "/path/to/control3/image.jpg"
+          - prompt: "do the thing to it"
+            ctrl_img: "/path/to/control/image.jpg"
+            ctrl_img2: "/path/to/control2/image.jpg"
+          - prompt: "do the thing to it"
+            ctrl_img: "/path/to/control/image.jpg"
+            ctrl_img2: "/path/to/control2/image.jpg"
+          - prompt: "do the thing to it"
+            ctrl_img: "/path/to/control/image.jpg"
+            ctrl_img2: "/path/to/control2/image.jpg"
+          - prompt: "do the thing to it"
+            ctrl_img: "/path/to/control/image.jpg"
+            ctrl_img2: "/path/to/control2/image.jpg"
+          - prompt: "do the thing to it"
+            ctrl_img: "/path/to/control/image.jpg"
+          - prompt: "do the thing to it"
+            ctrl_img: "/path/to/control/image.jpg"
+            ctrl_img2: "/path/to/control2/image.jpg"
+          - prompt: "do the thing to it"
+            ctrl_img: "/path/to/control/image.jpg"
+          - prompt: "do the thing to it"
+            ctrl_img: "/path/to/control/image.jpg"
+            ctrl_img2: "/path/to/control2/image.jpg"
+          - prompt: "do the thing to it"
+            ctrl_img: "/path/to/control/image.jpg"
+            ctrl_img2: "/path/to/control2/image.jpg"
+          - prompt: "do the thing to it"
+            ctrl_img: "/path/to/control/image.jpg"
+            ctrl_img2: "/path/to/control2/image.jpg"
+        neg: ""
+        seed: 42
+        walk_seed: true
+        guidance_scale: 3
+        sample_steps: 25
+# you can add any additional meta info here. [name] is replaced with config name at top
+meta:
+  name: "[name]"
+  version: '1.0'

extensions_built_in/diffusion_models/__init__.py

@@ -4,7 +4,7 @@
 from .omnigen2 import OmniGen2Model
 from .flux_kontext import FluxKontextModel
 from .wan22 import Wan225bModel, Wan2214bModel, Wan2214bI2VModel
-from .qwen_image import QwenImageModel, QwenImageEditModel
+from .qwen_image import QwenImageModel, QwenImageEditModel, QwenImageEditPlus2509Model
 
 AI_TOOLKIT_MODELS = [
     # put a list of models here
@@ -20,4 +20,5 @@
     Wan2214bModel,
     QwenImageModel,
     QwenImageEditModel,
+    QwenImageEditPlus2509Model,
 ]

extensions_built_in/diffusion_models/qwen_image/__init__.py

@@ -1,2 +1,3 @@
 from .qwen_image import QwenImageModel
-from .qwen_image_edit import QwenImageEditModel
+from .qwen_image_edit import QwenImageEditModel
+from .qwen_image_edit_plus2509 import QwenImageEditPlus2509Model