Upload folder using huggingface_hub

.ipynb_checkpoints/mod_lora_scale-checkpoint.yaml

@@ -0,0 +1,48 @@
+---
+job: mod
+config:
+  name: name_of_your_model_v1
+  process:
+    - type: rescale_lora
+      # path to your current lora model
+      input_path: "/path/to/lora/lora.safetensors"
+      # output path for your new lora model, can be the same as input_path to replace
+      output_path: "/path/to/lora/output_lora_v1.safetensors"
+      # replaces meta with the meta below (plus minimum meta fields)
+      # if false, we will leave the meta alone except for updating hashes (sd-script hashes)
+      replace_meta: true
+      # how to adjust, we can scale the up_down weights or the alpha
+      # up_down is the default and probably the best, they will both net the same outputs
+      # would only affect rare NaN cases and maybe merging with old merge tools
+      scale_target: 'up_down'
+      # precision to save, fp16 is the default and standard
+      save_dtype: fp16
+      # current_weight is the ideal weight you use as a multiplier when using the lora
+      # IE in automatic1111 <lora:my_lora:6.0> the 6.0 is the current_weight
+      # you can do negatives here too if you want to flip the lora
+      current_weight: 6.0
+      # target_weight is the ideal weight you use as a multiplier when using the lora
+      # instead of the one above. IE in automatic1111 instead of using <lora:my_lora:6.0>
+      # we want to use <lora:my_lora:1.0> so 1.0 is the target_weight
+      target_weight: 1.0
+
+      # base model for the lora
+      # this is just used to add meta so automatic111 knows which model it is for
+      # assume v1.5 if these are not set
+      is_xl: false
+      is_v2: false
+meta:
+  # this is only used if you set replace_meta to true above
+  name: "[name]"  # [name] gets replaced with the name above
+  description: A short description of your lora
+  trigger_words:
+  - put
+  - trigger
+  - words
+  - here
+  version: '0.1'
+  creator:
+    name: Your Name
+    email: your@email.com
+    website: https://yourwebsite.com
+  any: All meta data above is arbitrary, it can be whatever you want.

extract.example.yml

@@ -0,0 +1,75 @@
+---
+# this is in yaml format. You can use json if you prefer
+# I like both but yaml is easier to read and write
+# plus it has comments which is nice for documentation
+job: extract # tells the runner what to do
+config:
+  # the name will be used to create a folder in the output folder
+  # it will also replace any [name] token in the rest of this config
+  name: name_of_your_model
+  # can be hugging face model, a .ckpt, or a .safetensors
+  base_model: "/path/to/base/model.safetensors"
+  # can be hugging face model, a .ckpt, or a .safetensors
+  extract_model: "/path/to/model/to/extract/trained.safetensors"
+  # we will create folder here with name above so. This will create /path/to/output/folder/name_of_your_model
+  output_folder: "/path/to/output/folder"
+  is_v2: false
+  dtype: fp16 # saved dtype
+  device: cpu # cpu, cuda:0, etc
+
+  # processes can be chained like this to run multiple in a row
+  # they must all use same models above, but great for testing different
+  # sizes and typed of extractions. It is much faster as we already have the models loaded
+  process:
+  # process 1
+  - type: locon  # locon or lora (locon is lycoris)
+    filename: "[name]_64_32.safetensors" # will be put in output folder
+    dtype: fp16
+    mode: fixed
+    linear: 64
+    conv: 32
+
+  # process 2
+  - type: locon
+    output_path: "/absolute/path/for/this/output.safetensors" # can be absolute
+    mode: ratio
+    linear: 0.2
+    conv: 0.2
+
+  # process 3
+  - type: locon
+    filename: "[name]_ratio_02.safetensors"
+    mode: quantile
+    linear: 0.5
+    conv: 0.5
+
+  # process 4
+  - type: lora  # traditional lora extraction (lierla) with linear layers only
+    filename: "[name]_4.safetensors"
+    mode: fixed  # fixed, ratio, quantile supported for lora as well
+    linear: 4 # lora dim or rank
+    # no conv for lora
+
+  # process 5
+  - type: lora
+    filename: "[name]_q05.safetensors"
+    mode: quantile
+    linear: 0.5
+
+# you can put any information you want here, and it will be saved in the model
+# the below is an example. I recommend doing trigger words at a minimum
+# in the metadata. The software will include this plus some other information
+meta:
+  name: "[name]"  # [name] gets replaced with the name above
+  description: A short description of your model
+  trigger_words:
+  - put
+  - trigger
+  - words
+  - here
+  version: '0.1'
+  creator:
+    name: Your Name
+    email: your@email.com
+    website: https://yourwebsite.com
+  any: All meta data above is arbitrary, it can be whatever you want.

generate.example.yaml

@@ -0,0 +1,60 @@
+---
+
+job: generate # tells the runner what to do
+config:
+  name: "generate" # this is not really used anywhere currently but required by runner
+  process:
+    # process 1
+    - type: to_folder  # process images to a folder
+      output_folder: "output/gen"
+      device: cuda:0 # cpu, cuda:0, etc
+      generate:
+        # these are your defaults you can override most of them with flags
+        sampler: "ddpm" # ignored for now, will add later though ddpm is used regardless for now
+        width: 1024
+        height: 1024
+        neg: "cartoon, fake, drawing, illustration, cgi, animated, anime"
+        seed: -1 # -1 is random
+        guidance_scale: 7
+        sample_steps: 20
+        ext: ".png" # .png, .jpg, .jpeg, .webp
+
+        # here ate the flags you can use for prompts. Always start with
+        # your prompt first then add these flags after. You can use as many
+        # like
+        # photo of a baseball --n painting, ugly --w 1024 --h 1024 --seed 42 --cfg 7 --steps 20
+        # we will try to support all sd-scripts flags where we can
+
+        # FROM SD-SCRIPTS
+        # --n Treat everything until the next option as a negative prompt.
+        # --w Specify the width of the generated image.
+        # --h Specify the height of the generated image.
+        # --d Specify the seed for the generated image.
+        # --l Specify the CFG scale for the generated image.
+        # --s Specify the number of steps during generation.
+
+        # OURS and some QOL additions
+        # --p2 Prompt for the second text encoder (SDXL only)
+        # --n2 Negative prompt for the second text encoder (SDXL only)
+        # --gr Specify the guidance rescale for the generated image (SDXL only)
+        # --seed Specify the seed for the generated image same as --d
+        # --cfg Specify the CFG scale for the generated image same as --l
+        # --steps Specify the number of steps during generation same as --s
+
+        prompt_file: false # if true a txt file will be created next to images with prompt strings used
+        # prompts can also be a path to a text file with one prompt per line
+        # prompts: "/path/to/prompts.txt"
+        prompts:
+          - "photo of batman"
+          - "photo of superman"
+          - "photo of spiderman"
+          - "photo of a superhero --n batman superman spiderman"
+
+      model:
+        # huggingface name, relative prom project path, or absolute path to .safetensors or .ckpt
+        #      name_or_path: "runwayml/stable-diffusion-v1-5"
+        name_or_path: "/mnt/Models/stable-diffusion/models/stable-diffusion/Ostris/Ostris_Real_v1.safetensors"
+        is_v2: false  # for v2 models
+        is_v_pred: false # for v-prediction models (most v2 models)
+        is_xl: false  # for SDXL models
+        dtype: bf16

mod_lora_scale.yaml

@@ -0,0 +1,48 @@
+---
+job: mod
+config:
+  name: name_of_your_model_v1
+  process:
+    - type: rescale_lora
+      # path to your current lora model
+      input_path: "/path/to/lora/lora.safetensors"
+      # output path for your new lora model, can be the same as input_path to replace
+      output_path: "/path/to/lora/output_lora_v1.safetensors"
+      # replaces meta with the meta below (plus minimum meta fields)
+      # if false, we will leave the meta alone except for updating hashes (sd-script hashes)
+      replace_meta: true
+      # how to adjust, we can scale the up_down weights or the alpha
+      # up_down is the default and probably the best, they will both net the same outputs
+      # would only affect rare NaN cases and maybe merging with old merge tools
+      scale_target: 'up_down'
+      # precision to save, fp16 is the default and standard
+      save_dtype: fp16
+      # current_weight is the ideal weight you use as a multiplier when using the lora
+      # IE in automatic1111 <lora:my_lora:6.0> the 6.0 is the current_weight
+      # you can do negatives here too if you want to flip the lora
+      current_weight: 6.0
+      # target_weight is the ideal weight you use as a multiplier when using the lora
+      # instead of the one above. IE in automatic1111 instead of using <lora:my_lora:6.0>
+      # we want to use <lora:my_lora:1.0> so 1.0 is the target_weight
+      target_weight: 1.0
+
+      # base model for the lora
+      # this is just used to add meta so automatic111 knows which model it is for
+      # assume v1.5 if these are not set
+      is_xl: false
+      is_v2: false
+meta:
+  # this is only used if you set replace_meta to true above
+  name: "[name]"  # [name] gets replaced with the name above
+  description: A short description of your lora
+  trigger_words:
+  - put
+  - trigger
+  - words
+  - here
+  version: '0.1'
+  creator:
+    name: Your Name
+    email: your@email.com
+    website: https://yourwebsite.com
+  any: All meta data above is arbitrary, it can be whatever you want.

modal/modal_train_lora_flux_24gb.yaml

@@ -0,0 +1,97 @@
+---
+job: extension
+config:
+  # this name will be the folder and filename name
+  name: "my_first_flux_lora_v1"
+  process:
+    - type: 'sd_trainer'
+      # root folder to save training sessions/samples/weights
+      training_folder: "/root/ai-toolkit/modal_output" # must match MOUNT_DIR from run_modal.py
+      # 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_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"
+        # your dataset must be placed in /ai-toolkit and /root is for modal to find the dir:
+        - folder_path: "/root/ai-toolkit/your-dataset"
+          caption_ext: "txt"
+          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
+          shuffle_tokens: false  # shuffle caption order, split by commas
+          cache_latents_to_disk: true  # leave this true unless you know what you're doing
+          resolution: [ 512, 768, 1024 ]  # flux enjoys multiple resolutions
+      train:
+        batch_size: 1
+        steps: 2000  # total number of steps to train 500 - 4000 is a good range
+        gradient_accumulation_steps: 1
+        train_unet: true
+        train_text_encoder: false  # probably won't work with flux
+        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
+        # uncomment to use new vell curved weighting. Experimental but may produce better results
+#        linear_timesteps: true
+
+        # ema will smooth out learning, but could slow it down. Recommended to leave on.
+        ema_config:
+          use_ema: true
+          ema_decay: 0.99
+
+        # will probably need this if gpu supports it for flux, other dtypes may not work correctly
+        dtype: bf16
+      model:
+        # huggingface model name or path
+        # if you get an error, or get stuck while downloading,
+        # check https://github.com/ostris/ai-toolkit/issues/84, download the model locally and
+        # place it like "/root/ai-toolkit/FLUX.1-dev"
+        name_or_path: "black-forest-labs/FLUX.1-dev"
+        is_flux: true
+        quantize: true  # run 8bit mixed precision
+#        low_vram: true  # uncomment this if the GPU is connected to your monitors. It will use less vram to quantize, but is slower.
+      sample:
+        sampler: "flowmatch" # must match train.noise_scheduler
+        sample_every: 250 # sample every this many steps
+        sample_start_step: 0 # start sampling at this step
+        width: 1024
+        height: 1024
+        prompts:
+          # you can add [trigger] to the prompts here and it will be replaced with the trigger word
+#          - "[trigger] holding a sign that says 'I LOVE PROMPTS!'"\
+          - "woman with red hair, playing chess at the park, bomb going off in the background"
+          - "a woman holding a coffee cup, in a beanie, sitting at a cafe"
+          - "a horse is a DJ at a night club, fish eye lens, smoke machine, lazer lights, holding a martini"
+          - "a man showing off his cool new t shirt at the beach, a shark is jumping out of the water in the background"
+          - "a bear building a log cabin in the snow covered mountains"
+          - "woman playing the guitar, on stage, singing a song, laser lights, punk rocker"
+          - "hipster man with a beard, building a chair, in a wood shop"
+          - "photo of a man, white background, medium shot, modeling clothing, studio lighting, white backdrop"
+          - "a man holding a sign that says, 'this is a sign'"
+          - "a bulldog, in a post apocalyptic world, with a shotgun, in a leather jacket, in a desert, with a motorcycle"
+        neg: ""  # not used on flux
+        seed: 42
+        walk_seed: true
+        guidance_scale: 4
+        sample_steps: 20
+# you can add any additional meta info here. [name] is replaced with config name at top
+meta:
+  name: "[name]"
+  version: '1.0'

modal/modal_train_lora_flux_schnell_24gb.yaml

@@ -0,0 +1,99 @@
+---
+job: extension
+config:
+  # this name will be the folder and filename name
+  name: "my_first_flux_lora_v1"
+  process:
+    - type: 'sd_trainer'
+      # root folder to save training sessions/samples/weights
+      training_folder: "/root/ai-toolkit/modal_output" # must match MOUNT_DIR from run_modal.py
+      # 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_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"
+        # your dataset must be placed in /ai-toolkit and /root is for modal to find the dir:
+        - folder_path: "/root/ai-toolkit/your-dataset"
+          caption_ext: "txt"
+          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
+          shuffle_tokens: false  # shuffle caption order, split by commas
+          cache_latents_to_disk: true  # leave this true unless you know what you're doing
+          resolution: [ 512, 768, 1024 ]  # flux enjoys multiple resolutions
+      train:
+        batch_size: 1
+        steps: 2000  # total number of steps to train 500 - 4000 is a good range
+        gradient_accumulation_steps: 1
+        train_unet: true
+        train_text_encoder: false  # probably won't work with flux
+        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
+        # uncomment to use new vell curved weighting. Experimental but may produce better results
+#        linear_timesteps: true
+
+        # ema will smooth out learning, but could slow it down. Recommended to leave on.
+        ema_config:
+          use_ema: true
+          ema_decay: 0.99
+
+        # will probably need this if gpu supports it for flux, other dtypes may not work correctly
+        dtype: bf16
+      model:
+        # huggingface model name or path
+        # if you get an error, or get stuck while downloading,
+        # check https://github.com/ostris/ai-toolkit/issues/84, download the models locally and
+        # place them like "/root/ai-toolkit/FLUX.1-schnell" and "/root/ai-toolkit/FLUX.1-schnell-training-adapter"
+        name_or_path: "black-forest-labs/FLUX.1-schnell"
+        assistant_lora_path: "ostris/FLUX.1-schnell-training-adapter" # Required for flux schnell training
+        is_flux: true
+        quantize: true  # run 8bit mixed precision
+        # low_vram is painfully slow to fuse in the adapter avoid it unless absolutely necessary
+#        low_vram: true  # uncomment this if the GPU is connected to your monitors. It will use less vram to quantize, but is slower.
+      sample:
+        sampler: "flowmatch" # must match train.noise_scheduler
+        sample_every: 250 # sample every this many steps
+        sample_start_step: 0 # start sampling at this step
+        width: 1024
+        height: 1024
+        prompts:
+          # you can add [trigger] to the prompts here and it will be replaced with the trigger word
+#          - "[trigger] holding a sign that says 'I LOVE PROMPTS!'"\
+          - "woman with red hair, playing chess at the park, bomb going off in the background"
+          - "a woman holding a coffee cup, in a beanie, sitting at a cafe"
+          - "a horse is a DJ at a night club, fish eye lens, smoke machine, lazer lights, holding a martini"
+          - "a man showing off his cool new t shirt at the beach, a shark is jumping out of the water in the background"
+          - "a bear building a log cabin in the snow covered mountains"
+          - "woman playing the guitar, on stage, singing a song, laser lights, punk rocker"
+          - "hipster man with a beard, building a chair, in a wood shop"
+          - "photo of a man, white background, medium shot, modeling clothing, studio lighting, white backdrop"
+          - "a man holding a sign that says, 'this is a sign'"
+          - "a bulldog, in a post apocalyptic world, with a shotgun, in a leather jacket, in a desert, with a motorcycle"
+        neg: ""  # not used on flux
+        seed: 42
+        walk_seed: true
+        guidance_scale: 1  # schnell does not do guidance
+        sample_steps: 4  # 1 - 4 works well
+# you can add any additional meta info here. [name] is replaced with config name at top
+meta:
+  name: "[name]"
+  version: '1.0'

train_flex_redux.yaml

@@ -0,0 +1,113 @@
+---
+job: extension
+config:
+  # this name will be the folder and filename name
+  name: "my_first_flex_redux_finetune_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
+      adapter:
+        type: "redux"
+        # you can finetune an existing adapter or start from scratch. Set to null to start from scratch
+        name_or_path: '/local/path/to/redux_adapter_to_finetune.safetensors'
+        # name_or_path: null
+        # image_encoder_path: 'google/siglip-so400m-patch14-384' # Flux.1 redux adapter
+        image_encoder_path: 'google/siglip2-so400m-patch16-512' # Flex.1 512 redux adapter
+        # image_encoder_arch: 'siglip' # for Flux.1
+        image_encoder_arch: 'siglip2'
+        # You need a control input for each sample. Best to do squares for both images
+        test_img_path: 
+          - "/path/to/x_01.jpg"
+          - "/path/to/x_02.jpg"
+          - "/path/to/x_03.jpg"
+          - "/path/to/x_04.jpg"
+          - "/path/to/x_05.jpg"
+          - "/path/to/x_06.jpg"
+          - "/path/to/x_07.jpg"
+          - "/path/to/x_08.jpg"
+          - "/path/to/x_09.jpg"
+          - "/path/to/x_10.jpg"
+        clip_layer: 'last_hidden_state'
+        train: true
+      save:
+        dtype: bf16 # precision to save
+        save_every: 250 # save every this many steps
+        max_step_saves_to_keep: 4 
+      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"
+          # clip_image_path is directory containting your control images. They must have filename as their train image. (extension does not matter)
+          # for normal redux, we are just recreating the same image, so you can use the same folder path above
+          clip_image_path: "/path/to/control/images/folder"
+          caption_ext: "txt"
+          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
+          resolution: [ 512, 768, 1024 ]  # flex enjoys multiple resolutions
+      train:
+        # this is what I used for the 24GB card, but feel free to adjust
+        # total batch size is 6 here
+        batch_size: 3
+        gradient_accumulation: 2
+
+        # captions are not needed for this training, we cache a blank proompt and rely on the vision encoder
+        unload_text_encoder: true
+
+        loss_type: "mse"
+        train_unet: true
+        train_text_encoder: false
+        steps: 4000000  # I set this very high and stop when I like the results
+        content_or_style: balanced  # content, style, balanced
+        gradient_checkpointing: true
+        noise_scheduler: "flowmatch" # or "ddpm", "lms", "euler_a"
+        timestep_type: "flux_shift" 
+        optimizer: "adamw8bit"
+        lr: 1e-4
+
+        # this is for Flex.1, comment this out for FLUX.1-dev
+        bypass_guidance_embedding: true
+
+        dtype: bf16
+        ema_config:
+          use_ema: true
+          ema_decay: 0.99
+      model:
+        name_or_path: "ostris/Flex.1-alpha"
+        is_flux: true
+        quantize: true
+        text_encoder_bits: 8
+      sample:
+        sampler: "flowmatch" # must match train.noise_scheduler
+        sample_every: 250 # sample every this many steps
+        sample_start_step: 0 # start sampling at this step
+        width: 1024
+        height: 1024
+        # I leave half blank to test prompt and unprompted
+        prompts:
+          - "woman with red hair, playing chess at the park, bomb going off in the background"
+          - "a woman holding a coffee cup, in a beanie, sitting at a cafe"
+          - "a horse is a DJ at a night club, fish eye lens, smoke machine, lazer lights, holding a martini"
+          - "a man showing off his cool new t shirt at the beach, a shark is jumping out of the water in the background"
+          - "a bear building a log cabin in the snow covered mountains"
+          - ""
+          - ""
+          - ""
+          - ""
+          - ""
+        neg: ""
+        seed: 42
+        walk_seed: true
+        guidance_scale: 4
+        sample_steps: 25
+        network_multiplier: 1.0
+
+# you can add any additional meta info here. [name] is replaced with config name at top
+meta:
+  name: "[name]"
+  version: '1.0'

train_full_fine_tune_flex.yaml

@@ -0,0 +1,108 @@
+---
+# This configuration requires 48GB of VRAM or more to operate
+job: extension
+config:
+  # this name will be the folder and filename name
+  name: "my_first_flex_finetune_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_word: "p3r5on"
+      save:
+        dtype: bf16 # precision to save
+        save_every: 250 # save every this many steps
+        max_step_saves_to_keep: 2 # how many intermittent saves to keep
+        save_format: 'diffusers' # 'diffusers'
+      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"
+          caption_ext: "txt"
+          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
+          shuffle_tokens: false  # shuffle caption order, split by commas
+          # cache_latents_to_disk: true  # leave this true unless you know what you're doing
+          resolution: [ 512, 768, 1024 ]  # flex enjoys multiple resolutions
+      train:
+        batch_size: 1
+        # IMPORTANT! For Flex, you must bypass the guidance embedder during training
+        bypass_guidance_embedding: true
+
+        # can be 'sigmoid', 'linear', or 'lognorm_blend'
+        timestep_type: 'sigmoid' 
+        
+        steps: 2000  # total number of steps to train 500 - 4000 is a good range
+        gradient_accumulation: 1
+        train_unet: true
+        train_text_encoder: false  # probably won't work with flex
+        gradient_checkpointing: true  # need the on unless you have a ton of vram
+        noise_scheduler: "flowmatch" # for training only
+        optimizer: "adafactor"
+        lr: 3e-5
+
+        # Paramiter swapping can reduce vram requirements. Set factor from 1.0 to 0.0. 
+        # 0.1 is 10% of paramiters active at easc step. Only works with adafactor
+
+        # do_paramiter_swapping: true
+        # paramiter_swapping_factor: 0.9
+
+        # uncomment this to skip the pre training sample
+        # skip_first_sample: true
+        # uncomment to completely disable sampling
+        # disable_sampling: true
+
+        # ema will smooth out learning, but could slow it down. Recommended to leave on if you have the vram
+        ema_config:
+          use_ema: true
+          ema_decay: 0.99
+
+        # will probably need this if gpu supports it for flex, other dtypes may not work correctly
+        dtype: bf16
+      model:
+        # huggingface model name or path
+        name_or_path: "ostris/Flex.1-alpha"
+        is_flux: true # flex is flux architecture
+        # full finetuning quantized models is a crapshoot and results in subpar outputs
+        # quantize: true
+        # you can quantize just the T5 text encoder here to save vram
+        quantize_te: true
+        # only train the transformer blocks
+        only_if_contains: 
+          - "transformer.transformer_blocks."
+          - "transformer.single_transformer_blocks."
+      sample:
+        sampler: "flowmatch" # must match train.noise_scheduler
+        sample_every: 250 # sample every this many steps
+        sample_start_step: 0 # start sampling at this step
+        width: 1024
+        height: 1024
+        prompts:
+          # you can add [trigger] to the prompts here and it will be replaced with the trigger word
+          # - "[trigger] holding a sign that says 'I LOVE PROMPTS!'"\
+          - "woman with red hair, playing chess at the park, bomb going off in the background"
+          - "a woman holding a coffee cup, in a beanie, sitting at a cafe"
+          - "a horse is a DJ at a night club, fish eye lens, smoke machine, lazer lights, holding a martini"
+          - "a man showing off his cool new t shirt at the beach, a shark is jumping out of the water in the background"
+          - "a bear building a log cabin in the snow covered mountains"
+          - "woman playing the guitar, on stage, singing a song, laser lights, punk rocker"
+          - "hipster man with a beard, building a chair, in a wood shop"
+          - "photo of a man, white background, medium shot, modeling clothing, studio lighting, white backdrop"
+          - "a man holding a sign that says, 'this is a sign'"
+          - "a bulldog, in a post apocalyptic world, with a shotgun, in a leather jacket, in a desert, with a motorcycle"
+        neg: ""  # not used on flex
+        seed: 42
+        walk_seed: true
+        guidance_scale: 4
+        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'

train_full_fine_tune_lumina.yaml

@@ -0,0 +1,100 @@
+---
+# This configuration requires 24GB of VRAM or more to operate
+job: extension
+config:
+  # this name will be the folder and filename name
+  name: "my_first_lumina_finetune_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_word: "p3r5on"
+      save:
+        dtype: bf16 # precision to save
+        save_every: 250 # save every this many steps
+        max_step_saves_to_keep: 2 # how many intermittent saves to keep
+        save_format: 'diffusers' # 'diffusers'
+      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"
+          caption_ext: "txt"
+          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
+          shuffle_tokens: false  # shuffle caption order, split by commas
+          # cache_latents_to_disk: true  # leave this true unless you know what you're doing
+          resolution: [ 512, 768, 1024 ]  # lumina2 enjoys multiple resolutions
+      train:
+        batch_size: 1
+
+        # can be 'sigmoid', 'linear', or 'lumina2_shift'
+        timestep_type: 'lumina2_shift' 
+        
+        steps: 2000  # total number of steps to train 500 - 4000 is a good range
+        gradient_accumulation: 1
+        train_unet: true
+        train_text_encoder: false  # probably won't work with lumina2
+        gradient_checkpointing: true  # need the on unless you have a ton of vram
+        noise_scheduler: "flowmatch" # for training only
+        optimizer: "adafactor"
+        lr: 3e-5
+
+        # Paramiter swapping can reduce vram requirements. Set factor from 1.0 to 0.0. 
+        # 0.1 is 10% of paramiters active at easc step. Only works with adafactor
+
+        # do_paramiter_swapping: true
+        # paramiter_swapping_factor: 0.9
+
+        # uncomment this to skip the pre training sample
+        # skip_first_sample: true
+        # uncomment to completely disable sampling
+        # disable_sampling: true
+
+        # ema will smooth out learning, but could slow it down. Recommended to leave on if you have the vram
+        # ema_config:
+        #   use_ema: true
+        #   ema_decay: 0.99
+
+        # will probably need this if gpu supports it for lumina2, other dtypes may not work correctly
+        dtype: bf16
+      model:
+        # huggingface model name or path
+        name_or_path: "Alpha-VLLM/Lumina-Image-2.0"
+        is_lumina2: true # lumina2 architecture
+        # you can quantize just the Gemma2 text encoder here to save vram
+        quantize_te: true
+      sample:
+        sampler: "flowmatch" # must match train.noise_scheduler
+        sample_every: 250 # sample every this many steps
+        sample_start_step: 0 # start sampling at this step
+        width: 1024
+        height: 1024
+        prompts:
+          # you can add [trigger] to the prompts here and it will be replaced with the trigger word
+          # - "[trigger] holding a sign that says 'I LOVE PROMPTS!'"\
+          - "woman with red hair, playing chess at the park, bomb going off in the background"
+          - "a woman holding a coffee cup, in a beanie, sitting at a cafe"
+          - "a horse is a DJ at a night club, fish eye lens, smoke machine, lazer lights, holding a martini"
+          - "a man showing off his cool new t shirt at the beach, a shark is jumping out of the water in the background"
+          - "a bear building a log cabin in the snow covered mountains"
+          - "woman playing the guitar, on stage, singing a song, laser lights, punk rocker"
+          - "hipster man with a beard, building a chair, in a wood shop"
+          - "photo of a cat that is half black and half orange tabby, split down the middle. The cat has on a blue tophat. They are holding a martini glass with a pink ball of yarn in it with green knitting needles sticking out, in one paw. In the other paw, they are holding a DVD case for a movie titled, \"This is a test\" that has a golden robot on it. In the background is a busy night club with a giant mushroom man dancing with a bear."
+          - "a man holding a sign that says, 'this is a sign'"
+          - "a bulldog, in a post apocalyptic world, with a shotgun, in a leather jacket, in a desert, with a motorcycle"
+        neg: ""
+        seed: 42
+        walk_seed: true
+        guidance_scale: 4.0
+        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'

train_lora_chroma_24gb.yaml

@@ -0,0 +1,105 @@
+---
+job: extension
+config:
+  # this name will be the folder and filename name
+  name: "my_first_chroma_lora_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_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
+        push_to_hub: false #change this to True to push your trained model to Hugging Face.
+        # You can either set up a HF_TOKEN env variable or you'll be prompted to log-in         
+#       hf_repo_id: your-username/your-model-slug
+#       hf_private: true #whether the repo is private or public
+      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"
+          caption_ext: "txt"
+          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
+          shuffle_tokens: false  # shuffle caption order, split by commas
+          cache_latents_to_disk: true  # leave this true unless you know what you're doing
+          resolution: [ 512, 768, 1024 ]  # chroma enjoys multiple resolutions
+      train:
+        batch_size: 1
+        steps: 2000  # total number of steps to train 500 - 4000 is a good range
+        gradient_accumulation: 1
+        train_unet: true
+        train_text_encoder: false  # probably won't work with chroma
+        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
+        # uncomment to use new vell curved weighting. Experimental but may produce better results
+#        linear_timesteps: true
+
+        # ema will smooth out learning, but could slow it down. Recommended to leave on.
+        ema_config:
+          use_ema: true
+          ema_decay: 0.99
+
+        # will probably need this if gpu supports it for chroma, other dtypes may not work correctly
+        dtype: bf16
+      model:
+        # Download the whichever model you prefer from the Chroma repo
+        # https://huggingface.co/lodestones/Chroma/tree/main
+        # point to it here.
+        # name_or_path: "/path/to/chroma/chroma-unlocked-vVERSION.safetensors"
+
+        # using lodestones/Chroma will automatically use the latest version
+        name_or_path: "lodestones/Chroma"
+
+        # # You can also select a version of Chroma like so
+        # name_or_path: "lodestones/Chroma/v28"
+        
+        arch: "chroma"
+        quantize: true  # run 8bit mixed precision
+      sample:
+        sampler: "flowmatch" # must match train.noise_scheduler
+        sample_every: 250 # sample every this many steps
+        sample_start_step: 0 # start sampling at this step
+        width: 1024
+        height: 1024
+        prompts:
+          # you can add [trigger] to the prompts here and it will be replaced with the trigger word
+#          - "[trigger] holding a sign that says 'I LOVE PROMPTS!'"\
+          - "woman with red hair, playing chess at the park, bomb going off in the background"
+          - "a woman holding a coffee cup, in a beanie, sitting at a cafe"
+          - "a horse is a DJ at a night club, fish eye lens, smoke machine, lazer lights, holding a martini"
+          - "a man showing off his cool new t shirt at the beach, a shark is jumping out of the water in the background"
+          - "a bear building a log cabin in the snow covered mountains"
+          - "woman playing the guitar, on stage, singing a song, laser lights, punk rocker"
+          - "hipster man with a beard, building a chair, in a wood shop"
+          - "photo of a man, white background, medium shot, modeling clothing, studio lighting, white backdrop"
+          - "a man holding a sign that says, 'this is a sign'"
+          - "a bulldog, in a post apocalyptic world, with a shotgun, in a leather jacket, in a desert, with a motorcycle"
+        neg: ""  # negative prompt, optional
+        seed: 42
+        walk_seed: true
+        guidance_scale: 4
+        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'

train_lora_flex2_24gb.yaml

@@ -0,0 +1,166 @@
+# Note, Flex2 is a highly experimental WIP model. Finetuning a model with built in controls and inpainting has not 
+# been done before, so you will be experimenting with me on how to do it. This is my recommended setup, but this is highly
+# subject to change as we learn more about how Flex2 works.
+
+---
+job: extension
+config:
+  # this name will be the folder and filename name
+  name: "my_first_flex2_lora_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_word: "p3r5on"
+      network:
+        type: "lora"
+        linear: 32
+        linear_alpha: 32
+      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
+        push_to_hub: false #change this to True to push your trained model to Hugging Face.
+        # You can either set up a HF_TOKEN env variable or you'll be prompted to log-in         
+#       hf_repo_id: your-username/your-model-slug
+#       hf_private: true #whether the repo is private or public
+      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"
+          # Flex2 is trained with controls and inpainting. If you want the model to truely understand how the 
+          # controls function with your dataset, it is a good idea to keep doing controls during training. 
+          # this will automatically generate the controls for you before training. The current script is not 
+          # fully optimized so this could be rather slow for large datasets, but it caches them to disk so it
+          # only needs to be done once. If you want to skip this step, you can set the controls to [] and it will
+          controls:
+            - "depth"
+            - "line"
+            - "pose"
+            - "inpaint"
+          
+          # you can make custom inpainting images as well. These images must be webp or png format with an alpha.
+          # just erase the part of the image you want to inpaint and save it as a webp or png. Again, erase your
+          # train target. So the person if training a person. The automatic controls above with inpaint will
+          # just run a background remover mask and erase the foreground, which works well for subjects.
+
+          # inpaint_path: "/my/impaint/images"
+          
+          # you can also specify existing control image pairs. It can handle multiple groups and will randomly
+          # select one for each step.
+
+          # control_path:
+          #   - "/my/custom/control/images"
+          #   - "/my/custom/control/images2"
+          
+          caption_ext: "txt"
+          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
+          resolution: [ 512, 768, 1024 ]  # flex2 enjoys multiple resolutions
+      train:
+        batch_size: 1
+        # IMPORTANT! For Flex2, you must bypass the guidance embedder during training
+        bypass_guidance_embedding: true
+        
+        steps: 3000  # total number of steps to train 500 - 4000 is a good range
+        gradient_accumulation: 1
+        train_unet: true
+        train_text_encoder: false  # probably won't work with flex2
+        gradient_checkpointing: true  # need the on unless you have a ton of vram
+        noise_scheduler: "flowmatch" # for training only
+        # shift works well for training fast and learning composition and style. 
+        # for just subject, you may want to change this to sigmoid
+        timestep_type: 'shift'  # 'linear', 'sigmoid', 'shift'
+        optimizer: "adamw8bit"
+        lr: 1e-4
+
+        optimizer_params:
+          weight_decay: 1e-5
+        # uncomment this to skip the pre training sample
+#        skip_first_sample: true
+        # uncomment to completely disable sampling
+#        disable_sampling: true
+        # uncomment to use new vell curved weighting. Experimental but may produce better results
+#        linear_timesteps: true
+
+        # ema will smooth out learning, but could slow it down. Defaults off
+        ema_config:
+          use_ema: false
+          ema_decay: 0.99
+
+        # will probably need this if gpu supports it for flex, other dtypes may not work correctly
+        dtype: bf16
+      model:
+        # huggingface model name or path
+        name_or_path: "ostris/Flex.2-preview"
+        arch: "flex2"
+        quantize: true  # run 8bit mixed precision
+        quantize_te: true
+
+        # you can pass special training infor for controls to the model here
+        # percentages are decimal based so 0.0 is 0% and 1.0 is 100% of the time.
+        model_kwargs:
+          # inverts the inpainting mask, good to learn outpainting as well, recommended 0.0 for characters
+          invert_inpaint_mask_chance: 0.5
+          # this will do a normal t2i training step without inpaint when dropped out. REcommended if you want
+          # your lora to be able to inference with and without inpainting.
+          inpaint_dropout: 0.5
+          # randomly drops out the control image. Dropout recvommended if your want it to work without controls as well.
+          control_dropout: 0.5
+          # does a random inpaint blob. Usually a good idea to keep. Without it, the model will learn to always 100%
+          # fill the inpaint area with your subject. This is not always a good thing.
+          inpaint_random_chance: 0.5
+          # generates random inpaint blobs if you did not provide an inpaint image for your dataset. Inpaint breaks down fast
+          # if you are not training with it. Controls are a little more robust and can be left out,
+          # but when in doubt, always leave this on
+          do_random_inpainting: false
+          # does random blurring of the inpaint mask. Helps prevent weird edge artifacts for real workd inpainting. Leave on.
+          random_blur_mask: true
+          # applies a small amount of random dialition and restriction to the inpaint mask. Helps with edge artifacts.
+          # Leave on.
+          random_dialate_mask: true
+      sample:
+        sampler: "flowmatch" # must match train.noise_scheduler
+        sample_every: 250 # sample every this many steps
+        sample_start_step: 0 # start sampling at this step
+        width: 1024
+        height: 1024
+        prompts:
+          # you can add [trigger] to the prompts here and it will be replaced with the trigger word
+          # - "[trigger] holding a sign that says 'I LOVE PROMPTS!'"\
+
+          # you can use a single inpaint or single control image on your samples. 
+          # for controls, the ctrl_idx is 1, the images can be any name and image format. 
+          # use either a pose/line/depth image or whatever you are training with. An example is
+          # - "photo of [trigger] --ctrl_idx 1 --ctrl_img /path/to/control/image.jpg"
+
+          # for an inpainting image, it must be png/webp. Erase the part of the image you want to inpaint
+          # IMPORTANT! the inpaint images must be ctrl_idx 0 and have .inpaint.{ext} in the name for this to work right.
+          # - "photo of [trigger] --ctrl_idx 0 --ctrl_img /path/to/inpaint/image.inpaint.png"
+
+          - "woman with red hair, playing chess at the park, bomb going off in the background"
+          - "a woman holding a coffee cup, in a beanie, sitting at a cafe"
+          - "a horse is a DJ at a night club, fish eye lens, smoke machine, lazer lights, holding a martini"
+          - "a man showing off his cool new t shirt at the beach, a shark is jumping out of the water in the background"
+          - "a bear building a log cabin in the snow covered mountains"
+          - "woman playing the guitar, on stage, singing a song, laser lights, punk rocker"
+          - "hipster man with a beard, building a chair, in a wood shop"
+          - "photo of a man, white background, medium shot, modeling clothing, studio lighting, white backdrop"
+          - "a man holding a sign that says, 'this is a sign'"
+          - "a bulldog, in a post apocalyptic world, with a shotgun, in a leather jacket, in a desert, with a motorcycle"
+        neg: ""  # not used on flex2
+        seed: 42
+        walk_seed: true
+        guidance_scale: 4
+        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'

train_lora_flex_24gb.yaml

@@ -0,0 +1,102 @@
+---
+job: extension
+config:
+  # this name will be the folder and filename name
+  name: "my_first_flex_lora_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_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
+        push_to_hub: false #change this to True to push your trained model to Hugging Face.
+        # You can either set up a HF_TOKEN env variable or you'll be prompted to log-in         
+#       hf_repo_id: your-username/your-model-slug
+#       hf_private: true #whether the repo is private or public
+      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"
+          caption_ext: "txt"
+          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
+          shuffle_tokens: false  # shuffle caption order, split by commas
+          cache_latents_to_disk: true  # leave this true unless you know what you're doing
+          resolution: [ 512, 768, 1024 ]  # flex enjoys multiple resolutions
+      train:
+        batch_size: 1
+        # IMPORTANT! For Flex, you must bypass the guidance embedder during training
+        bypass_guidance_embedding: true
+        
+        steps: 2000  # total number of steps to train 500 - 4000 is a good range
+        gradient_accumulation: 1
+        train_unet: true
+        train_text_encoder: false  # probably won't work with flex
+        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
+        # uncomment to use new vell curved weighting. Experimental but may produce better results
+#        linear_timesteps: true
+
+        # ema will smooth out learning, but could slow it down. Recommended to leave on.
+        ema_config:
+          use_ema: true
+          ema_decay: 0.99
+
+        # will probably need this if gpu supports it for flex, other dtypes may not work correctly
+        dtype: bf16
+      model:
+        # huggingface model name or path
+        name_or_path: "ostris/Flex.1-alpha"
+        is_flux: true
+        quantize: true  # run 8bit mixed precision
+        quantize_kwargs:
+          exclude:
+            - "*time_text_embed*"  # exclude the time text embedder from quantization
+      sample:
+        sampler: "flowmatch" # must match train.noise_scheduler
+        sample_every: 250 # sample every this many steps
+        sample_start_step: 0 # start sampling at this step
+        width: 1024
+        height: 1024
+        prompts:
+          # you can add [trigger] to the prompts here and it will be replaced with the trigger word
+#          - "[trigger] holding a sign that says 'I LOVE PROMPTS!'"\
+          - "woman with red hair, playing chess at the park, bomb going off in the background"
+          - "a woman holding a coffee cup, in a beanie, sitting at a cafe"
+          - "a horse is a DJ at a night club, fish eye lens, smoke machine, lazer lights, holding a martini"
+          - "a man showing off his cool new t shirt at the beach, a shark is jumping out of the water in the background"
+          - "a bear building a log cabin in the snow covered mountains"
+          - "woman playing the guitar, on stage, singing a song, laser lights, punk rocker"
+          - "hipster man with a beard, building a chair, in a wood shop"
+          - "photo of a man, white background, medium shot, modeling clothing, studio lighting, white backdrop"
+          - "a man holding a sign that says, 'this is a sign'"
+          - "a bulldog, in a post apocalyptic world, with a shotgun, in a leather jacket, in a desert, with a motorcycle"
+        neg: ""  # not used on flex
+        seed: 42
+        walk_seed: true
+        guidance_scale: 4
+        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'

train_lora_flux_24gb.yaml

@@ -0,0 +1,97 @@
+---
+job: extension
+config:
+  # this name will be the folder and filename name
+  name: "my_first_flux_lora_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_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
+        push_to_hub: false #change this to True to push your trained model to Hugging Face.
+        # You can either set up a HF_TOKEN env variable or you'll be prompted to log-in         
+#       hf_repo_id: your-username/your-model-slug
+#       hf_private: true #whether the repo is private or public
+      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"
+          caption_ext: "txt"
+          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
+          shuffle_tokens: false  # shuffle caption order, split by commas
+          cache_latents_to_disk: true  # leave this true unless you know what you're doing
+          resolution: [ 512, 768, 1024 ]  # flux enjoys multiple resolutions
+      train:
+        batch_size: 1
+        steps: 2000  # total number of steps to train 500 - 4000 is a good range
+        gradient_accumulation_steps: 1
+        train_unet: true
+        train_text_encoder: false  # probably won't work with flux
+        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
+        # uncomment to use new vell curved weighting. Experimental but may produce better results
+#        linear_timesteps: true
+
+        # ema will smooth out learning, but could slow it down. Recommended to leave on.
+        ema_config:
+          use_ema: true
+          ema_decay: 0.99
+
+        # will probably need this if gpu supports it for flux, other dtypes may not work correctly
+        dtype: bf16
+      model:
+        # huggingface model name or path
+        name_or_path: "black-forest-labs/FLUX.1-dev"
+        is_flux: true
+        quantize: true  # run 8bit mixed precision
+#        low_vram: true  # uncomment this if the GPU is connected to your monitors. It will use less vram to quantize, but is slower.
+      sample:
+        sampler: "flowmatch" # must match train.noise_scheduler
+        sample_every: 250 # sample every this many steps
+        sample_start_step: 0 # start sampling at this step
+        width: 1024
+        height: 1024
+        prompts:
+          # you can add [trigger] to the prompts here and it will be replaced with the trigger word
+#          - "[trigger] holding a sign that says 'I LOVE PROMPTS!'"\
+          - "woman with red hair, playing chess at the park, bomb going off in the background"
+          - "a woman holding a coffee cup, in a beanie, sitting at a cafe"
+          - "a horse is a DJ at a night club, fish eye lens, smoke machine, lazer lights, holding a martini"
+          - "a man showing off his cool new t shirt at the beach, a shark is jumping out of the water in the background"
+          - "a bear building a log cabin in the snow covered mountains"
+          - "woman playing the guitar, on stage, singing a song, laser lights, punk rocker"
+          - "hipster man with a beard, building a chair, in a wood shop"
+          - "photo of a man, white background, medium shot, modeling clothing, studio lighting, white backdrop"
+          - "a man holding a sign that says, 'this is a sign'"
+          - "a bulldog, in a post apocalyptic world, with a shotgun, in a leather jacket, in a desert, with a motorcycle"
+        neg: ""  # not used on flux
+        seed: 42
+        walk_seed: true
+        guidance_scale: 4
+        sample_steps: 20
+# you can add any additional meta info here. [name] is replaced with config name at top
+meta:
+  name: "[name]"
+  version: '1.0'

train_lora_flux_kontext_24gb.yaml

@@ -0,0 +1,107 @@
+---
+job: extension
+config:
+  # this name will be the folder and filename name
+  name: "my_first_flux_kontext_lora_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_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
+        push_to_hub: false #change this to True to push your trained model to Hugging Face.
+        # You can either set up a HF_TOKEN env variable or you'll be prompted to log-in         
+#       hf_repo_id: your-username/your-model-slug
+#       hf_private: true #whether the repo is private or public
+      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 is the input images for kontext for a paired dataset. These are the source images you want to change.  
+          # You can comment this out and only use normal images if you don't have a paired dataset. 
+          # Control images need to match the filenames on the folder path but in
+          # a different folder. These do not need captions.
+          control_path: "/path/to/control/folder"
+          caption_ext: "txt"
+          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
+          shuffle_tokens: false  # shuffle caption order, split by commas
+          cache_latents_to_disk: true  # leave this true unless you know what you're doing
+          # Kontext runs images in at 2x the latent size. It may OOM at 1024 resolution with 24GB vram.
+          resolution: [ 512, 768 ]  # flux enjoys multiple resolutions
+          # resolution: [ 512, 768, 1024 ]
+      train:
+        batch_size: 1
+        steps: 3000  # total number of steps to train 500 - 4000 is a good range
+        gradient_accumulation_steps: 1
+        train_unet: true
+        train_text_encoder: false  # probably won't work with flux
+        gradient_checkpointing: true  # need the on unless you have a ton of vram
+        noise_scheduler: "flowmatch" # for training only
+        optimizer: "adamw8bit"
+        lr: 1e-4
+        timestep_type: "weighted" # sigmoid, linear, or weighted.
+        # uncomment this to skip the pre training sample
+#        skip_first_sample: true
+        # uncomment to completely disable sampling
+#        disable_sampling: true
+
+        # ema will smooth out learning, but could slow it down.
+
+        # ema_config:
+        #   use_ema: true
+        #   ema_decay: 0.99
+
+        # will probably need this if gpu supports it for flux, other dtypes may not work correctly
+        dtype: bf16
+      model:
+        # huggingface model name or path. This model is gated.
+        # visit https://huggingface.co/black-forest-labs/FLUX.1-Kontext-dev to accept the terms and conditions
+        # and then you can use this model.
+        name_or_path: "black-forest-labs/FLUX.1-Kontext-dev"
+        arch: "flux_kontext"
+        quantize: true  # run 8bit mixed precision
+#        low_vram: true  # uncomment this if the GPU is connected to your monitors. It will use less vram to quantize, but is slower.
+      sample:
+        sampler: "flowmatch" # must match train.noise_scheduler
+        sample_every: 250 # sample every this many steps
+        sample_start_step: 0 # start sampling at this step
+        width: 1024
+        height: 1024
+        prompts:
+          # you can add [trigger] to the prompts here and it will be replaced with the trigger word
+          # the --ctrl_img path is the one loaded to apply the kontext editing to
+#          - "[trigger] holding a sign that says 'I LOVE PROMPTS!'"\
+          - "make the person smile  --ctrl_img /path/to/control/folder/person1.jpg"
+          - "give the person an afro  --ctrl_img /path/to/control/folder/person1.jpg"
+          - "turn this image into a cartoon  --ctrl_img /path/to/control/folder/person1.jpg"
+          - "put this person in an action film  --ctrl_img /path/to/control/folder/person1.jpg"
+          - "make this person a rapper in a rap music video  --ctrl_img /path/to/control/folder/person1.jpg"
+          - "make the person smile  --ctrl_img /path/to/control/folder/person1.jpg"
+          - "give the person an afro  --ctrl_img /path/to/control/folder/person1.jpg"
+          - "turn this image into a cartoon  --ctrl_img /path/to/control/folder/person1.jpg"
+          - "put this person in an action film  --ctrl_img /path/to/control/folder/person1.jpg"
+          - "make this person a rapper in a rap music video  --ctrl_img /path/to/control/folder/person1.jpg"
+        neg: ""  # not used on flux
+        seed: 42
+        walk_seed: true
+        guidance_scale: 4
+        sample_steps: 20
+# you can add any additional meta info here. [name] is replaced with config name at top
+meta:
+  name: "[name]"
+  version: '1.0'

train_lora_flux_schnell_24gb.yaml

@@ -0,0 +1,99 @@
+---
+job: extension
+config:
+  # this name will be the folder and filename name
+  name: "my_first_flux_lora_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_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
+        push_to_hub: false #change this to True to push your trained model to Hugging Face.
+        # You can either set up a HF_TOKEN env variable or you'll be prompted to log-in         
+#       hf_repo_id: your-username/your-model-slug
+#       hf_private: true #whether the repo is private or public        
+      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"
+          caption_ext: "txt"
+          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
+          shuffle_tokens: false  # shuffle caption order, split by commas
+          cache_latents_to_disk: true  # leave this true unless you know what you're doing
+          resolution: [ 512, 768, 1024 ]  # flux enjoys multiple resolutions
+      train:
+        batch_size: 1
+        steps: 2000  # total number of steps to train 500 - 4000 is a good range
+        gradient_accumulation_steps: 1
+        train_unet: true
+        train_text_encoder: false  # probably won't work with flux
+        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
+        # uncomment to use new bell curved weighting. Experimental but may produce better results
+#        linear_timesteps: true
+
+        # ema will smooth out learning, but could slow it down. Recommended to leave on.
+        ema_config:
+          use_ema: true
+          ema_decay: 0.99
+
+        # will probably need this if gpu supports it for flux, other dtypes may not work correctly
+        dtype: bf16
+      model:
+        # huggingface model name or path
+        name_or_path: "black-forest-labs/FLUX.1-schnell"
+        assistant_lora_path: "ostris/FLUX.1-schnell-training-adapter" # Required for flux schnell training
+        is_flux: true
+        quantize: true  # run 8bit mixed precision
+        # low_vram is painfully slow to fuse in the adapter avoid it unless absolutely necessary
+#        low_vram: true  # uncomment this if the GPU is connected to your monitors. It will use less vram to quantize, but is slower.
+      sample:
+        sampler: "flowmatch" # must match train.noise_scheduler
+        sample_every: 250 # sample every this many steps
+        sample_start_step: 0 # start sampling at this step
+        width: 1024
+        height: 1024
+        prompts:
+          # you can add [trigger] to the prompts here and it will be replaced with the trigger word
+#          - "[trigger] holding a sign that says 'I LOVE PROMPTS!'"\
+          - "woman with red hair, playing chess at the park, bomb going off in the background"
+          - "a woman holding a coffee cup, in a beanie, sitting at a cafe"
+          - "a horse is a DJ at a night club, fish eye lens, smoke machine, lazer lights, holding a martini"
+          - "a man showing off his cool new t shirt at the beach, a shark is jumping out of the water in the background"
+          - "a bear building a log cabin in the snow covered mountains"
+          - "woman playing the guitar, on stage, singing a song, laser lights, punk rocker"
+          - "hipster man with a beard, building a chair, in a wood shop"
+          - "photo of a man, white background, medium shot, modeling clothing, studio lighting, white backdrop"
+          - "a man holding a sign that says, 'this is a sign'"
+          - "a bulldog, in a post apocalyptic world, with a shotgun, in a leather jacket, in a desert, with a motorcycle"
+        neg: ""  # not used on flux
+        seed: 42
+        walk_seed: true
+        guidance_scale: 1  # schnell does not do guidance
+        sample_steps: 4  # 1 - 4 works well
+# you can add any additional meta info here. [name] is replaced with config name at top
+meta:
+  name: "[name]"
+  version: '1.0'

train_lora_hidream_48.yaml

@@ -0,0 +1,113 @@
+# HiDream training is still highly experimental. The settings here will take ~35.2GB of vram to train.
+# It is not possible to train on a single 24GB card yet, but I am working on it. If you have more VRAM
+# I highly recommend first disabling quantization on the model itself if you can. You can leave the TEs quantized.
+# HiDream has a mixture of experts that may take special training considerations that I do not
+# have implemented properly. The current implementation seems to work well for LoRA training, but
+# may not be effective for longer training runs. The implementation could change in future updates
+# so your results may vary when this happens.
+
+---
+job: extension
+config:
+  # this name will be the folder and filename name
+  name: "my_first_hidream_lora_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_word: "p3r5on"
+      network:
+        type: "lora"
+        linear: 32
+        linear_alpha: 32
+        network_kwargs:
+          # it is probably best to ignore the mixture of experts since only 2 are active each block. It works activating it, but I wouldnt.
+          # proper training of it is not fully implemented
+          ignore_if_contains:
+            - "ff_i.experts"
+            - "ff_i.gate"
+      save:
+        dtype: bfloat16 # 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"
+          caption_ext: "txt"
+          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
+          resolution: [ 512, 768, 1024 ]  # hidream enjoys multiple resolutions
+      train:
+        batch_size: 1
+        steps: 3000  # total number of steps to train 500 - 4000 is a good range
+        gradient_accumulation_steps: 1
+        train_unet: true
+        train_text_encoder: false  # wont work with hidream
+        gradient_checkpointing: true  # need the on unless you have a ton of vram
+        noise_scheduler: "flowmatch" # for training only
+        timestep_type: shift # sigmoid, shift, linear
+        optimizer: "adamw8bit"
+        lr: 2e-4
+        # uncomment this to skip the pre training sample
+#        skip_first_sample: true
+        # uncomment to completely disable sampling
+#        disable_sampling: true
+        # uncomment to use new vell curved weighting. Experimental but may produce better results
+#        linear_timesteps: true
+
+        # ema will smooth out learning, but could slow it down. Defaults off
+        ema_config:
+          use_ema: false
+          ema_decay: 0.99
+
+        # will probably need this if gpu supports it for hidream, other dtypes may not work correctly
+        dtype: bf16
+      model:
+        # the transformer will get grabbed from this hf repo
+        # warning ONLY train on Full. The dev and fast models are distilled and will break 
+        name_or_path: "HiDream-ai/HiDream-I1-Full"
+        # the extras will be grabbed from this hf repo. (text encoder, vae) 
+        extras_name_or_path: "HiDream-ai/HiDream-I1-Full"
+        arch: "hidream"
+        # both need to be quantized to train on 48GB currently
+        quantize: true
+        quantize_te: true
+        model_kwargs:
+          # llama is a gated model, It defaults to unsloth version, but you can set the llama path here
+          llama_model_path: "unsloth/Meta-Llama-3.1-8B-Instruct"
+      sample:
+        sampler: "flowmatch" # must match train.noise_scheduler
+        sample_every: 250 # sample every this many steps
+        sample_start_step: 0 # start sampling at this step
+        width: 1024
+        height: 1024
+        prompts:
+          # you can add [trigger] to the prompts here and it will be replaced with the trigger word
+#          - "[trigger] holding a sign that says 'I LOVE PROMPTS!'"\
+          - "woman with red hair, playing chess at the park, bomb going off in the background"
+          - "a woman holding a coffee cup, in a beanie, sitting at a cafe"
+          - "a horse is a DJ at a night club, fish eye lens, smoke machine, lazer lights, holding a martini"
+          - "a man showing off his cool new t shirt at the beach, a shark is jumping out of the water in the background"
+          - "a bear building a log cabin in the snow covered mountains"
+          - "woman playing the guitar, on stage, singing a song, laser lights, punk rocker"
+          - "hipster man with a beard, building a chair, in a wood shop"
+          - "photo of a man, white background, medium shot, modeling clothing, studio lighting, white backdrop"
+          - "a man holding a sign that says, 'this is a sign'"
+          - "a bulldog, in a post apocalyptic world, with a shotgun, in a leather jacket, in a desert, with a motorcycle"
+        neg: ""
+        seed: 42
+        walk_seed: true
+        guidance_scale: 4
+        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'

train_lora_lumina.yaml

@@ -0,0 +1,97 @@
+---
+# This configuration requires 20GB of VRAM or more to operate
+job: extension
+config:
+  # this name will be the folder and filename name
+  name: "my_first_lumina_lora_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_word: "p3r5on"
+      network:
+        type: "lora"
+        linear: 16
+        linear_alpha: 16
+      save:
+        dtype: bf16 # precision to save
+        save_every: 250 # save every this many steps
+        max_step_saves_to_keep: 2 # how many intermittent saves to keep
+        save_format: 'diffusers' # 'diffusers'
+      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"
+          caption_ext: "txt"
+          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
+          shuffle_tokens: false  # shuffle caption order, split by commas
+          # cache_latents_to_disk: true  # leave this true unless you know what you're doing
+          resolution: [ 512, 768, 1024 ]  # lumina2 enjoys multiple resolutions
+      train:
+        batch_size: 1
+
+        # can be 'sigmoid', 'linear', or 'lumina2_shift'
+        timestep_type: 'lumina2_shift' 
+        
+        steps: 2000  # total number of steps to train 500 - 4000 is a good range
+        gradient_accumulation: 1
+        train_unet: true
+        train_text_encoder: false  # probably won't work with lumina2
+        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
+
+        # ema will smooth out learning, but could slow it down. Recommended to leave on if you have the vram
+        ema_config:
+          use_ema: true
+          ema_decay: 0.99
+
+        # will probably need this if gpu supports it for lumina2, other dtypes may not work correctly
+        dtype: bf16
+      model:
+        # huggingface model name or path
+        name_or_path: "Alpha-VLLM/Lumina-Image-2.0"
+        is_lumina2: true # lumina2 architecture
+        # you can quantize just the Gemma2 text encoder here to save vram
+        quantize_te: true
+      sample:
+        sampler: "flowmatch" # must match train.noise_scheduler
+        sample_every: 250 # sample every this many steps
+        sample_start_step: 0 # start sampling at this step
+        width: 1024
+        height: 1024
+        prompts:
+          # you can add [trigger] to the prompts here and it will be replaced with the trigger word
+          # - "[trigger] holding a sign that says 'I LOVE PROMPTS!'"\
+          - "woman with red hair, playing chess at the park, bomb going off in the background"
+          - "a woman holding a coffee cup, in a beanie, sitting at a cafe"
+          - "a horse is a DJ at a night club, fish eye lens, smoke machine, lazer lights, holding a martini"
+          - "a man showing off his cool new t shirt at the beach, a shark is jumping out of the water in the background"
+          - "a bear building a log cabin in the snow covered mountains"
+          - "woman playing the guitar, on stage, singing a song, laser lights, punk rocker"
+          - "hipster man with a beard, building a chair, in a wood shop"
+          - "photo of a cat that is half black and half orange tabby, split down the middle. The cat has on a blue tophat. They are holding a martini glass with a pink ball of yarn in it with green knitting needles sticking out, in one paw. In the other paw, they are holding a DVD case for a movie titled, \"This is a test\" that has a golden robot on it. In the background is a busy night club with a giant mushroom man dancing with a bear."
+          - "a man holding a sign that says, 'this is a sign'"
+          - "a bulldog, in a post apocalyptic world, with a shotgun, in a leather jacket, in a desert, with a motorcycle"
+        neg: ""
+        seed: 42
+        walk_seed: true
+        guidance_scale: 4.0
+        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'

train_lora_omnigen2_24gb.yaml

@@ -0,0 +1,95 @@
+---
+job: extension
+config:
+  # this name will be the folder and filename name
+  name: "my_first_omnigen2_lora_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_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
+        push_to_hub: false #change this to True to push your trained model to Hugging Face.
+        # You can either set up a HF_TOKEN env variable or you'll be prompted to log-in         
+#       hf_repo_id: your-username/your-model-slug
+#       hf_private: true #whether the repo is private or public
+      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"
+          caption_ext: "txt"
+          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
+          shuffle_tokens: false  # shuffle caption order, split by commas
+          cache_latents_to_disk: true  # leave this true unless you know what you're doing
+          resolution: [ 512, 768, 1024 ]  # omnigen2 should work with multiple resolutions
+      train:
+        batch_size: 1
+        steps: 3000  # total number of steps to train 500 - 4000 is a good range
+        gradient_accumulation: 1
+        train_unet: true
+        train_text_encoder: false  # probably won't work with omnigen2
+        gradient_checkpointing: true  # need the on unless you have a ton of vram
+        noise_scheduler: "flowmatch" # for training only
+        optimizer: "adamw8bit"
+        lr: 1e-4
+        timestep_type: 'sigmoid' # sigmoid, linear, shift
+        # uncomment this to skip the pre training sample
+#        skip_first_sample: true
+        # uncomment to completely disable sampling
+#        disable_sampling: true
+
+        # ema will smooth out learning, but could slow it down.
+        # ema_config:
+        #   use_ema: true
+        #   ema_decay: 0.99
+
+        # will probably need this if gpu supports it for omnigen2, other dtypes may not work correctly
+        dtype: bf16
+      model:
+        name_or_path: "OmniGen2/OmniGen2
+        arch: "omnigen2"
+        quantize_te: true  # quantize_only te
+        # quantize: true  # quantize transformer
+      sample:
+        sampler: "flowmatch" # must match train.noise_scheduler
+        sample_every: 250 # sample every this many steps
+        sample_start_step: 0 # start sampling at this step
+        width: 1024
+        height: 1024
+        prompts:
+          # you can add [trigger] to the prompts here and it will be replaced with the trigger word
+#          - "[trigger] holding a sign that says 'I LOVE PROMPTS!'"\
+          - "woman with red hair, playing chess at the park, bomb going off in the background"
+          - "a woman holding a coffee cup, in a beanie, sitting at a cafe"
+          - "a horse is a DJ at a night club, fish eye lens, smoke machine, lazer lights, holding a martini"
+          - "a man showing off his cool new t shirt at the beach, a shark is jumping out of the water in the background"
+          - "a bear building a log cabin in the snow covered mountains"
+          - "woman playing the guitar, on stage, singing a song, laser lights, punk rocker"
+          - "hipster man with a beard, building a chair, in a wood shop"
+          - "photo of a man, white background, medium shot, modeling clothing, studio lighting, white backdrop"
+          - "a man holding a sign that says, 'this is a sign'"
+          - "a bulldog, in a post apocalyptic world, with a shotgun, in a leather jacket, in a desert, with a motorcycle"
+        neg: ""  # negative prompt, optional
+        seed: 42
+        walk_seed: true
+        guidance_scale: 4
+        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'

train_lora_qwen_image_24gb.yaml

@@ -0,0 +1,96 @@
+---
+job: extension
+config:
+  # this name will be the folder and filename name
+  name: "my_first_qwen_image_lora_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"
+          caption_ext: "txt"
+          # default_caption: "a person" # if caching text embeddings, if you dont have captions, this will get cached
+          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
+          shuffle_tokens: false  # shuffle caption order, split by commas
+          cache_latents_to_disk: true  # leave this true unless you have a large dataset
+          # if you OOM, 1024 may be too much, but should work
+          resolution: [ 512, 768, 1024 ]  # qwen image enjoys multiple resolutions
+      train:
+        batch_size: 1
+        # caching text embeddings is required for 24GB
+        cache_text_embeddings: true
+        
+        steps: 2000  # total number of steps to train 500 - 4000 is a good range
+        gradient_accumulation: 1
+        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"
+        arch: "qwen_image"
+        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 24GB
+        qtype: "uint3|ostris/accuracy_recovery_adapters/qwen_image_torchao_uint3.safetensors"
+        quantize_te: true
+        qtype_te: "qfloat8"
+        low_vram: true
+      sample:
+        sampler: "flowmatch" # must match train.noise_scheduler
+        sample_every: 250 # sample every this many steps
+        sample_start_step: 0 # start sampling at this step
+        width: 1024
+        height: 1024
+        prompts:
+          # you can add [trigger] to the prompts here and it will be replaced with the trigger word
+          # - "[trigger] holding a sign that says 'I LOVE PROMPTS!'"\ 
+          - "woman with red hair, playing chess at the park, bomb going off in the background"
+          - "a woman holding a coffee cup, in a beanie, sitting at a cafe"
+          - "a horse is a DJ at a night club, fish eye lens, smoke machine, lazer lights, holding a martini"
+          - "a man showing off his cool new t shirt at the beach, a shark is jumping out of the water in the background"
+          - "a bear building a log cabin in the snow covered mountains"
+          - "woman playing the guitar, on stage, singing a song, laser lights, punk rocker"
+          - "hipster man with a beard, building a chair, in a wood shop"
+          - "photo of a man, white background, medium shot, modeling clothing, studio lighting, white backdrop"
+          - "a man holding a sign that says, 'this is a sign'"
+          - "a bulldog, in a post apocalyptic world, with a shotgun, in a leather jacket, in a desert, with a motorcycle"
+        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'

train_lora_qwen_image_edit_2509_32gb.yaml

@@ -0,0 +1,106 @@
+---
+job: extension
+config:
+  # this name will be the folder and filename name
+  name: "my_first_qwen_image_edit_2509_lora_v1"
+  process:
+    - type: 'diffusion_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
+      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"
+          # can do up to 3 control image folders, file names must match target file names, but aspect/size can be different
+          control_path:
+            - "/path/to/control/images/folder1"
+            - "/path/to/control/images/folder2"
+            - "/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
+          # a trigger word that can be cached with the text embeddings 
+          # trigger_word: "optional trigger word"
+      train:
+        batch_size: 1
+        # caching text embeddings is required for 32GB
+        cache_text_embeddings: true
+        # unload_text_encoder: 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_plus"
+        quantize: true
+        # 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|ostris/accuracy_recovery_adapters/qwen_image_edit_2509_torchao_uint3.safetensors"
+        quantize_te: true
+        qtype_te: "qfloat8"
+        low_vram: true
+      sample:
+        sampler: "flowmatch" # must match train.noise_scheduler
+        sample_every: 250 # sample every this many steps
+        sample_start_step: 0 # start sampling at this step
+        width: 1024
+        height: 1024
+        # you can provide up to 3 control images here
+        samples:
+          - prompt: "Do whatever with Image1 and Image2"
+            ctrl_img_1: "/path/to/image1.png"
+            ctrl_img_2: "/path/to/image2.png"
+            # ctrl_img_3: "/path/to/image3.png"
+          - prompt: "Do whatever with Image1 and Image2"
+            ctrl_img_1: "/path/to/image1.png"
+            ctrl_img_2: "/path/to/image2.png"
+            # ctrl_img_3: "/path/to/image3.png"
+          - prompt: "Do whatever with Image1 and Image2"
+            ctrl_img_1: "/path/to/image1.png"
+            ctrl_img_2: "/path/to/image2.png"
+            # ctrl_img_3: "/path/to/image3.png"
+          - prompt: "Do whatever with Image1 and Image2"
+            ctrl_img_1: "/path/to/image1.png"
+            ctrl_img_2: "/path/to/image2.png"
+            # ctrl_img_3: "/path/to/image3.png"
+          - prompt: "Do whatever with Image1 and Image2"
+            ctrl_img_1: "/path/to/image1.png"
+            ctrl_img_2: "/path/to/image2.png"
+            # ctrl_img_3: "/path/to/image3.png"
+        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'

train_lora_qwen_image_edit_32gb.yaml

@@ -0,0 +1,103 @@
+---
+job: extension
+config:
+  # this name will be the folder and filename name
+  name: "my_first_qwen_image_edit_lora_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"
+          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"
+        arch: "qwen_image_edit"
+        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
+        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
+        sample_every: 250 # sample every this many steps
+        sample_start_step: 0 # start sampling at this step
+        width: 1024
+        height: 1024
+        samples:
+          - 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"
+          - 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"
+          - 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"
+          - 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"
+          - 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"
+        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'

train_lora_sd35_large_24gb.yaml

@@ -0,0 +1,98 @@
+---
+# NOTE!! THIS IS CURRENTLY EXPERIMENTAL AND UNDER DEVELOPMENT. SOME THINGS WILL CHANGE
+job: extension
+config:
+  # this name will be the folder and filename name
+  name: "my_first_sd3l_lora_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_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
+        push_to_hub: false #change this to True to push your trained model to Hugging Face.
+        # You can either set up a HF_TOKEN env variable or you'll be prompted to log-in         
+#       hf_repo_id: your-username/your-model-slug
+#       hf_private: true #whether the repo is private or public
+      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"
+          caption_ext: "txt"
+          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
+          shuffle_tokens: false  # shuffle caption order, split by commas
+          cache_latents_to_disk: true  # leave this true unless you know what you're doing
+          resolution: [ 1024 ] 
+      train:
+        batch_size: 1
+        steps: 2000  # total number of steps to train 500 - 4000 is a good range
+        gradient_accumulation_steps: 1
+        train_unet: true
+        train_text_encoder: false  # May not fully work with SD3 yet
+        gradient_checkpointing: true  # need the on unless you have a ton of vram
+        noise_scheduler: "flowmatch"
+        timestep_type: "linear" # linear or sigmoid
+        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
+        # uncomment to use new vell curved weighting. Experimental but may produce better results
+#        linear_timesteps: true
+
+        # ema will smooth out learning, but could slow it down. Recommended to leave on.
+        ema_config:
+          use_ema: true
+          ema_decay: 0.99
+
+        # will probably need this if gpu supports it for sd3, other dtypes may not work correctly
+        dtype: bf16
+      model:
+        # huggingface model name or path
+        name_or_path: "stabilityai/stable-diffusion-3.5-large"
+        is_v3: true
+        quantize: true  # run 8bit mixed precision
+      sample:
+        sampler: "flowmatch" # must match train.noise_scheduler
+        sample_every: 250 # sample every this many steps
+        sample_start_step: 0 # start sampling at this step
+        width: 1024
+        height: 1024
+        prompts:
+          # you can add [trigger] to the prompts here and it will be replaced with the trigger word
+#          - "[trigger] holding a sign that says 'I LOVE PROMPTS!'"\
+          - "woman with red hair, playing chess at the park, bomb going off in the background"
+          - "a woman holding a coffee cup, in a beanie, sitting at a cafe"
+          - "a horse is a DJ at a night club, fish eye lens, smoke machine, lazer lights, holding a martini"
+          - "a man showing off his cool new t shirt at the beach, a shark is jumping out of the water in the background"
+          - "a bear building a log cabin in the snow covered mountains"
+          - "woman playing the guitar, on stage, singing a song, laser lights, punk rocker"
+          - "hipster man with a beard, building a chair, in a wood shop"
+          - "photo of a man, white background, medium shot, modeling clothing, studio lighting, white backdrop"
+          - "a man holding a sign that says, 'this is a sign'"
+          - "a bulldog, in a post apocalyptic world, with a shotgun, in a leather jacket, in a desert, with a motorcycle"
+        neg: ""  
+        seed: 42
+        walk_seed: true
+        guidance_scale: 4
+        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'

train_lora_wan21_14b_24gb.yaml

@@ -0,0 +1,102 @@
+# IMPORTANT: The Wan2.1 14B model is huge. This config should work on 24GB GPUs. It cannot
+# support keeping the text encoder on GPU while training with 24GB, so it is only good
+# for training on a single prompt, for example a person with a trigger word.
+# to train on captions, you need more vran for now.
+---
+job: extension
+config:
+  # this name will be the folder and filename name
+  name: "my_first_wan21_14b_lora_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
+      # this is probably needed for 24GB cards when offloading TE to CPU
+      trigger_word: "p3r5on"
+      network:
+        type: "lora"
+        linear: 32
+        linear_alpha: 32
+      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
+        push_to_hub: false #change this to True to push your trained model to Hugging Face.
+        # You can either set up a HF_TOKEN env variable or you'll be prompted to log-in         
+#       hf_repo_id: your-username/your-model-slug
+#       hf_private: true #whether the repo is private or public
+      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"
+        # AI-Toolkit does not currently support video datasets, we will train on 1 frame at a time
+        # it works well for characters, but not as well for "actions"
+        - folder_path: "/path/to/images/folder"
+          caption_ext: "txt"
+          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
+          shuffle_tokens: false  # shuffle caption order, split by commas
+          cache_latents_to_disk: true  # leave this true unless you know what you're doing
+          resolution: [ 632 ]  # will be around 480p
+      train:
+        batch_size: 1
+        steps: 2000  # total number of steps to train 500 - 4000 is a good range
+        gradient_accumulation: 1
+        train_unet: true
+        train_text_encoder: false  # probably won't work with wan
+        gradient_checkpointing: true  # need the on unless you have a ton of vram
+        noise_scheduler: "flowmatch" # for training only
+        timestep_type: 'sigmoid'
+        optimizer: "adamw8bit"
+        lr: 1e-4
+        optimizer_params:
+          weight_decay: 1e-4
+        # uncomment this to skip the pre training sample
+#        skip_first_sample: true
+        # uncomment to completely disable sampling
+#        disable_sampling: true
+        # ema will smooth out learning, but could slow it down. Recommended to leave on.
+        ema_config:
+          use_ema: true
+          ema_decay: 0.99
+        dtype: bf16
+        # required for 24GB cards
+        # this will encode your trigger word and use those embeddings for every image in the dataset
+        unload_text_encoder: true
+      model:
+        # huggingface model name or path
+        name_or_path: "Wan-AI/Wan2.1-T2V-14B-Diffusers"
+        arch: 'wan21'
+        # these settings will save as much vram as possible
+        quantize: true
+        quantize_te: true
+        low_vram: true 
+      sample:
+        sampler: "flowmatch"
+        sample_every: 250 # sample every this many steps
+        sample_start_step: 0 # start sampling at this step
+        width: 832
+        height: 480
+        num_frames: 40
+        fps: 15
+        # samples take a long time. so use them sparingly
+        # samples will be animated webp files, if you don't see them animated, open in a browser.
+        prompts:
+          # you can add [trigger] to the prompts here and it will be replaced with the trigger word
+#          - "[trigger] holding a sign that says 'I LOVE PROMPTS!'"\
+          - "woman playing the guitar, on stage, singing a song, laser lights, punk rocker"
+        neg: ""  
+        seed: 42
+        walk_seed: true
+        guidance_scale: 5
+        sample_steps: 30
+# you can add any additional meta info here. [name] is replaced with config name at top
+meta:
+  name: "[name]"
+  version: '1.0'

train_lora_wan21_1b_24gb.yaml

@@ -0,0 +1,91 @@
+---
+job: extension
+config:
+  # this name will be the folder and filename name
+  name: "my_first_wan21_1b_lora_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_word: "p3r5on"
+      network:
+        type: "lora"
+        linear: 32
+        linear_alpha: 32
+      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
+        push_to_hub: false #change this to True to push your trained model to Hugging Face.
+        # You can either set up a HF_TOKEN env variable or you'll be prompted to log-in         
+#       hf_repo_id: your-username/your-model-slug
+#       hf_private: true #whether the repo is private or public
+      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"
+        # AI-Toolkit does not currently support video datasets, we will train on 1 frame at a time
+        # it works well for characters, but not as well for "actions"
+        - folder_path: "/path/to/images/folder"
+          caption_ext: "txt"
+          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
+          shuffle_tokens: false  # shuffle caption order, split by commas
+          cache_latents_to_disk: true  # leave this true unless you know what you're doing
+          resolution: [ 632 ]  # will be around 480p
+      train:
+        batch_size: 1
+        steps: 2000  # total number of steps to train 500 - 4000 is a good range
+        gradient_accumulation: 1
+        train_unet: true
+        train_text_encoder: false  # probably won't work with wan
+        gradient_checkpointing: true  # need the on unless you have a ton of vram
+        noise_scheduler: "flowmatch" # for training only
+        timestep_type: 'sigmoid'
+        optimizer: "adamw8bit"
+        lr: 1e-4
+        optimizer_params:
+          weight_decay: 1e-4
+        # uncomment this to skip the pre training sample
+#        skip_first_sample: true
+        # uncomment to completely disable sampling
+#        disable_sampling: true
+        # ema will smooth out learning, but could slow it down. Recommended to leave on.
+        ema_config:
+          use_ema: true
+          ema_decay: 0.99
+        dtype: bf16
+      model:
+        # huggingface model name or path
+        name_or_path: "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+        arch: 'wan21'
+        quantize_te: true # saves vram
+      sample:
+        sampler: "flowmatch"
+        sample_every: 250 # sample every this many steps
+        sample_start_step: 0 # start sampling at this step
+        width: 832
+        height: 480
+        num_frames: 40
+        fps: 15
+        # samples take a long time. so use them sparingly
+        # samples will be animated webp files, if you don't see them animated, open in a browser.
+        prompts:
+          # you can add [trigger] to the prompts here and it will be replaced with the trigger word
+#          - "[trigger] holding a sign that says 'I LOVE PROMPTS!'"\
+          - "woman playing the guitar, on stage, singing a song, laser lights, punk rocker"
+        neg: ""  
+        seed: 42
+        walk_seed: true
+        guidance_scale: 5
+        sample_steps: 30
+# you can add any additional meta info here. [name] is replaced with config name at top
+meta:
+  name: "[name]"
+  version: '1.0'

train_lora_wan22_14b_24gb.yaml

@@ -0,0 +1,112 @@
+# this example focuses mainly for training Wan2.2 14b on images. It will work for video as well by increasing
+# the number of frames in the dataset and samples. Training on and generating video is very VRAM intensive.
+---
+job: extension
+config:
+  # this name will be the folder and filename name
+  name: "my_first_wan22_14b_lora_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
+      # Use a trigger word if train.unload_text_encoder is true, however, if caching text embeddings, do not use a trigger word
+      # trigger_word: "p3r5on"
+      network:
+        type: "lora"
+        linear: 32
+        linear_alpha: 32
+      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.
+        # "C:\\path\\to\\images\\folder"
+        - folder_path: "/path/to/images/or/video/folder"
+          caption_ext: "txt"
+          caption_dropout_rate: 0.05  # will drop out the caption 5% of time
+          # number of frames to extract from your video. It will automatically extract them evenly spaced
+          # set to 1 frame for images
+          num_frames: 1
+          resolution: [ 512, 768, 1024]
+      train:
+        batch_size: 1
+        steps: 2000  # total number of steps to train 500 - 4000 is a good range
+        gradient_accumulation: 1
+        train_unet: true
+        train_text_encoder: false  # probably won't work with wan
+        gradient_checkpointing: true  # need the on unless you have a ton of vram
+        noise_scheduler: "flowmatch" # for training only
+        timestep_type: 'linear'
+        optimizer: "adamw8bit"
+        lr: 1e-4
+        optimizer_params:
+          weight_decay: 1e-4
+        # uncomment this to skip the pre training sample
+#        skip_first_sample: true
+        # uncomment to completely disable sampling
+#        disable_sampling: true
+        dtype: bf16
+
+        # IMPORTANT: this is for Wan 2.2 MOE. It will switch training one stage or the other every this many steps
+        switch_boundary_every: 10
+
+        # required for 24GB cards. You must do either unload_text_encoder or cache_text_embeddings but not both
+
+        # this will encode your trigger word and use those embeddings for every image in the dataset, captions will be ignored
+        # unload_text_encoder: true
+
+        # this will cache all captions in your dataset.
+        cache_text_embeddings: true
+
+      model:
+        # huggingface model name or path, this one if bf16, vs the float32 of the official repo
+        name_or_path: "ai-toolkit/Wan2.2-T2V-A14B-Diffusers-bf16"
+        arch: 'wan22_14b'
+        quantize: true
+        # This will pull and use a custom Accuracy Recovery Adapter to train at 4bit
+        qtype: "uint4|ostris/accuracy_recovery_adapters/wan22_14b_t2i_torchao_uint4.safetensors"
+        quantize_te: true
+        qtype_te: "qfloat8"
+        low_vram: true
+        model_kwargs:
+          # you can train high noise, low noise, or both. With low vram it will automatically unload the one not being trained.
+          train_high_noise: true
+          train_low_noise: true
+      sample:
+        sampler: "flowmatch"
+        sample_every: 250 # sample every this many steps
+        sample_start_step: 0 # start sampling at this step
+        width: 1024
+        height: 1024
+        # set to 1 for images
+        num_frames: 1
+        fps: 16
+        # samples take a long time. so use them sparingly
+        # samples will be animated webp files, if you don't see them animated, open in a browser.
+        prompts:
+          # you can add [trigger] to the prompts here and it will be replaced with the trigger word
+#          - "[trigger] holding a sign that says 'I LOVE PROMPTS!'"\
+          - "woman with red hair, playing chess at the park, bomb going off in the background"
+          - "a woman holding a coffee cup, in a beanie, sitting at a cafe"
+          - "a horse is a DJ at a night club, fish eye lens, smoke machine, lazer lights, holding a martini"
+          - "a man showing off his cool new t shirt at the beach, a shark is jumping out of the water in the background"
+          - "a bear building a log cabin in the snow covered mountains"
+          - "woman playing the guitar, on stage, singing a song, laser lights, punk rocker"
+          - "hipster man with a beard, building a chair, in a wood shop"
+          - "photo of a man, white background, medium shot, modeling clothing, studio lighting, white backdrop"
+          - "a man holding a sign that says, 'this is a sign'"
+          - "a bulldog, in a post apocalyptic world, with a shotgun, in a leather jacket, in a desert, with a motorcycle"
+        neg: ""  
+        seed: 42
+        walk_seed: true
+        guidance_scale: 3.5
+        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'

train_slider.example.yml

@@ -0,0 +1,230 @@
+---
+# This is in yaml format. You can use json if you prefer
+# I like both but yaml is easier to write
+# Plus it has comments which is nice for documentation
+# This is the config I use on my sliders, It is solid and tested
+job: train
+config:
+  # the name will be used to create a folder in the output folder
+  # it will also replace any [name] token in the rest of this config
+  name: detail_slider_v1
+  # folder will be created with name above in folder below
+  # it can be relative to the project root or absolute
+  training_folder: "output/LoRA"
+  device: cuda:0 # cpu, cuda:0, etc
+  # for tensorboard logging, we will make a subfolder for this job
+  log_dir: "output/.tensorboard"
+  # you can stack processes for other jobs, It is not tested with sliders though
+  # just use one for now
+  process:
+    - type: slider # tells runner to run the slider process
+      # network is the LoRA network for a slider, I recommend to leave this be
+      network:
+        # network type lierla is traditional LoRA that works everywhere, only linear layers
+        type: "lierla"
+        # rank / dim of the network. Bigger is not always better. Especially for sliders. 8 is good
+        linear: 8
+        linear_alpha: 4 # Do about half of rank
+      # training config
+      train:
+        # this is also used in sampling. Stick with ddpm unless you know what you are doing
+        noise_scheduler: "ddpm" # or "ddpm", "lms", "euler_a"
+        # how many steps to train. More is not always better. I rarely go over 1000
+        steps: 500
+        # I have had good results with 4e-4 to 1e-4 at 500 steps
+        lr: 2e-4
+        # enables gradient checkpoint, saves vram, leave it on
+        gradient_checkpointing: true
+        # train the unet. I recommend leaving this true
+        train_unet: true
+        # train the text encoder. I don't recommend this unless you have a special use case
+        # for sliders we are adjusting representation of the concept (unet),
+        # not the description of it (text encoder)
+        train_text_encoder: false
+        # same as from sd-scripts, not fully tested but should speed up training
+        min_snr_gamma: 5.0
+        # just leave unless you know what you are doing
+        # also supports "dadaptation" but set lr to 1 if you use that,
+        # but it learns too fast and I don't recommend it
+        optimizer: "adamw"
+        # only constant for now
+        lr_scheduler: "constant"
+        # we randomly denoise random num of steps form 1 to this number
+        # while training. Just leave it
+        max_denoising_steps: 40
+        # works great at 1. I do 1 even with my 4090.
+        # higher may not work right with newer single batch stacking code anyway
+        batch_size: 1
+        # bf16 works best if your GPU supports it (modern)
+        dtype: bf16  # fp32, bf16, fp16
+        # if you have it, use it. It is faster and better
+        # torch 2.0 doesnt need xformers anymore, only use if you have lower version
+#        xformers: true
+        # I don't recommend using unless you are trying to make a darker lora. Then do 0.1 MAX
+        # although, the way we train sliders is comparative, so it probably won't work anyway
+        noise_offset: 0.0
+#        noise_offset: 0.0357  # SDXL was trained with offset of 0.0357. So use that when training on SDXL
+
+      # the model to train the LoRA network on
+      model:
+        # huggingface name, relative prom project path, or absolute path to .safetensors or .ckpt
+        name_or_path: "runwayml/stable-diffusion-v1-5"
+        is_v2: false  # for v2 models
+        is_v_pred: false # for v-prediction models (most v2 models)
+        # has some issues with the dual text encoder and the way we train sliders
+        # it works bit weights need to probably be higher to see it.
+        is_xl: false  # for SDXL models
+
+      # saving config
+      save:
+        dtype: float16 # precision to save. I recommend float16
+        save_every: 50 # save every this many steps
+        # this will remove step counts more than this number
+        # allows you to save more often in case of a crash without filling up your drive
+        max_step_saves_to_keep: 2
+
+      # sampling config
+      sample:
+        # must match train.noise_scheduler, this is not used here
+        # but may be in future and in other processes
+        sampler: "ddpm"
+        # sample every this many steps
+        sample_every: 20
+        # image size
+        width: 512
+        height: 512
+        # prompts to use for sampling. Do as many as you want, but it slows down training
+        # pick ones that will best represent the concept you are trying to adjust
+        # allows some flags after the prompt
+        #  --m [number]  # network multiplier. LoRA weight. -3 for the negative slide, 3 for the positive
+        #      slide are good tests. will inherit sample.network_multiplier if not set
+        #  --n [string]  # negative prompt, will inherit sample.neg if not set
+        # Only 75 tokens allowed currently
+        # I like to do a wide positive and negative spread so I can see a good range and stop
+        # early if the network is braking down
+        prompts:
+          - "a woman in a coffee shop, black hat, blonde hair, blue jacket --m -5"
+          - "a woman in a coffee shop, black hat, blonde hair, blue jacket --m -3"
+          - "a woman in a coffee shop, black hat, blonde hair, blue jacket --m 3"
+          - "a woman in a coffee shop, black hat, blonde hair, blue jacket --m 5"
+          - "a golden retriever sitting on a leather couch, --m -5"
+          - "a golden retriever sitting on a leather couch --m -3"
+          - "a golden retriever sitting on a leather couch --m 3"
+          - "a golden retriever sitting on a leather couch --m 5"
+          - "a man with a beard and red flannel shirt, wearing vr goggles, walking into traffic --m -5"
+          - "a man with a beard and red flannel shirt, wearing vr goggles, walking into traffic --m -3"
+          - "a man with a beard and red flannel shirt, wearing vr goggles, walking into traffic --m 3"
+          - "a man with a beard and red flannel shirt, wearing vr goggles, walking into traffic --m 5"
+        # negative prompt used on all prompts above as default if they don't have one
+        neg: "cartoon, fake, drawing, illustration, cgi, animated, anime, monochrome"
+        # seed for sampling. 42 is the answer for everything
+        seed: 42
+        # walks the seed so s1 is 42, s2 is 43, s3 is 44, etc
+        # will start over on next sample_every so s1 is always seed
+        # works well if you use same prompt but want different results
+        walk_seed: false
+        # cfg scale (4 to 10 is good)
+        guidance_scale: 7
+        # sampler steps (20 to 30 is good)
+        sample_steps: 20
+        # default network multiplier for all prompts
+        # since we are training a slider, I recommend overriding this with --m [number]
+        # in the prompts above to get both sides of the slider
+        network_multiplier: 1.0
+
+      # logging information
+      logging:
+        log_every: 10 # log every this many steps
+        use_wandb: false # not supported yet
+        verbose: false # probably done need unless you are debugging
+
+      # slider training config, best for last
+      slider:
+        # resolutions to train on. [ width, height ]. This is less important for sliders
+        # as we are not teaching the model anything it doesn't already know
+        # but must be a size it understands [ 512, 512 ] for sd_v1.5  and [ 768, 768 ] for sd_v2.1
+        # and [ 1024, 1024 ] for sd_xl
+        # you can do as many as you want here
+        resolutions:
+          - [ 512, 512 ]
+#          - [ 512, 768 ]
+#          - [ 768, 768 ]
+        # slider training uses 4 combined steps for a single round. This will do it in one gradient
+        # step. It is highly optimized and shouldn't take anymore vram than doing without it,
+        # since we break down batches for gradient accumulation now. so just leave it on.
+        batch_full_slide: true
+        # These are the concepts to train on. You can do as many as you want here,
+        # but they can conflict outweigh each other. Other than experimenting, I recommend
+        # just doing one for good results
+        targets:
+            # target_class is the base concept we are adjusting the representation of
+            # for example, if we are adjusting the representation of a person, we would use "person"
+            # if we are adjusting the representation of a cat, we would use "cat" It is not
+            # a keyword necessarily but what the model understands the concept to represent.
+            # "person" will affect men, women, children, etc but will not affect cats, dogs, etc
+            # it is the models base general understanding of the concept and everything it represents
+            # you can leave it blank to affect everything. In this example, we are adjusting
+            # detail, so we will leave it blank to affect everything
+          - target_class: ""
+            # positive is the prompt for the positive side of the slider.
+            # It is the concept that will be excited and amplified in the model when we slide the slider
+            # to the positive side and forgotten / inverted when we slide
+            # the slider to the negative side. It is generally best to include the target_class in
+            # the prompt. You want it to be the extreme of what you want to train on. For example,
+            # if you want to train on fat people, you would use "an extremely fat, morbidly obese person"
+            # as the prompt. Not just "fat person"
+            # max 75 tokens for now
+            positive: "high detail, 8k, intricate, detailed, high resolution, high res, high quality"
+            # negative is the prompt for the negative side of the slider and works the same as positive
+            # it does not necessarily work the same as a negative prompt when generating images
+            # these need to be polar opposites.
+            # max 76 tokens for now
+            negative: "blurry, boring, fuzzy, low detail, low resolution, low res, low quality"
+            # the loss for this target is multiplied by this number.
+            # if you are doing more than one target it may be good to set less important ones
+            # to a lower number like 0.1 so they don't outweigh the primary target
+            weight: 1.0
+            # shuffle the prompts split by the comma. We will run every combination randomly
+            # this will make the LoRA more robust. You probably want this on unless prompt order
+            # is important for some reason
+            shuffle: true
+
+
+        # anchors are prompts that we will try to hold on to while training the slider
+        # these are NOT necessary and can prevent the slider from converging if not done right
+        # leave them off if you are having issues, but they can help lock the network
+        # on certain concepts to help prevent catastrophic forgetting
+        # you want these to generate an image that is not your target_class, but close to it
+        # is fine as long as it does not directly overlap it.
+        # For example, if you are training on a person smiling,
+        # you could use "a person with a face mask" as an anchor. It is a person, the image is the same
+        # regardless if they are smiling or not, however, the closer the concept is to the target_class
+        # the less the multiplier needs to be. Keep multipliers less than 1.0 for anchors usually
+        # for close concepts, you want to be closer to 0.1 or 0.2
+        # these will slow down training. I am leaving them off for the demo
+
+#        anchors:
+#          - prompt: "a woman"
+#            neg_prompt: "animal"
+#            # the multiplier applied to the LoRA when this is run.
+#            # higher will give it more weight but also help keep the lora from collapsing
+#            multiplier: 1.0
+#          - prompt: "a man"
+#            neg_prompt: "animal"
+#            multiplier: 1.0
+#          - prompt: "a person"
+#            neg_prompt: "animal"
+#            multiplier: 1.0
+
+# You can put any information you want here, and it will be saved in the model.
+# The below is an example, but you can put your grocery list in it if you want.
+# It is saved in the model so be aware of that. The software will include this
+# plus some other information for you automatically
+meta:
+  # [name] gets replaced with the name above
+  name: "[name]"
+#  version: '1.0'
+#  creator:
+#    name: Your Name
+#    email: your@gmail.com
+#    website: https://your.website