Upload train_lora_flux_kontext_24gb.yaml

Kontext_pixel_art/train_lora_flux_kontext_24gb.yaml

@@ -0,0 +1,99 @@
+---
+job: extension
+# 特别注意: 云端关闭镜像后会清除内容, 关闭前先把数据保存到本地, 或者自己存储镜像
+# 特别注意: 云端关闭镜像后会清除内容, 关闭前先把数据保存到本地, 或者自己存储镜像
+# 特别注意: 云端关闭镜像后会清除内容, 关闭前先把数据保存到本地, 或者自己存储镜像
+# 特别注意: 云端关闭镜像后会清除内容, 关闭前先把数据保存到本地, 或者自己存储镜像
+config:
+  name: "Kontext_pixel_art"   # 你的Lora名称
+  process:
+    - type: sd_trainer  
+      training_folder: output
+#      performance_log_every: 1000  
+      device: cuda:0                                                    #用哪张显卡训练, 默认单卡不要改
+#      trigger_word: "p3r5on"                                           #触发词, 打标的时候没有写可以删掉前面的 # 号, 在双引号中自己输入
+      network:
+        type: lora
+        linear: 16
+        linear_alpha: 16                                                #跟学习率有关, 越低越难练(没做大数据测试, 欢迎数据反馈), 跟下面 lr 学习率成反相关
+        lokr_full_rank: true
+        lokr_factor: -1
+        network_kwargs:
+          ignore_if_contains: []
+      save:
+        dtype: bf16                                                  # precision to save
+        save_every: 250                                                 #  多少步保存一次模型
+        max_step_saves_to_keep: 4                                       # 只保留最新的几个模型
+        save_format: diffusers
+        push_to_hub: false
+      datasets:
+        - folder_path: "/root/style/img"        # 原始数据路径, 仅支持jpg, jpeg, and png, 不需要打标文件
+          control_path: "/root/root/style/edit"        # 训练数据路径, 仅支持jpg, jpeg, and png  需要打标txt文件, 例如"Let him hold a sword in his hand"
+          mask_path: null
+          mask_min_value: 0.1
+          default_caption: ''
+          caption_ext: "txt"                                            # 数据集打标文件的格式
+          # num_repeates: 20
+          caption_dropout_rate: 0.05                                    # 随机删除千分之5的关键词
+          shuffle_tokens: false                                         # 是否打乱提示词, 会降低触发词的强度, 同时增加其他关键词的稳定性
+          cache_latents_to_disk: true                                   # leave this true unless you know what you're doing
+          resolution: [ 512, 768]                                          # 数据集图片尺寸, 多个用[512,1024,...]  , 不改也没事, 好像会自动分, 数据集图片尺寸太大了可能会爆显存, 不建议超过1500
+      train:
+        lr: 1e-4                                                        #学习率, 1e-4(0.0001)到6e-4(0.0006)之间, 根据数据集大小, 训练步数, 适当调整     
+        batch_size: 1                                                   # 并行训练数, 量力而行, 不怕爆显存可以调高
+        steps: 10000                                                     # 训练步数, 推荐1000-4000之间, 根据数据集大小调整, 也可逐步加大训练, 比如先练2000, 看看效果, 不行再加再训练  
+        gradient_accumulation_steps: 1
+        train_unet: true  
+        train_text_encoder: false                                       # probably won't work with flux
+        #content_or_style: balanced                                      # content, style, balanced  类似于风格Lora还是人物Lora, content侧重于内容, style侧重于风格, balanced在两者之间平衡 (没做测试)
+        gradient_checkpointing: true                                    # need the on unless you have a ton of vram
+        noise_scheduler: flowmatch                                    # for training only
+        optimizer: adamw8bit
+        timestep_type: weighted
+        content_or_style: balanced
+        optimizer_params:
+          weight_decay: 0.0001  
+        # lr_scheduler: polynomial                                                
+        skip_first_sample: true                                        #默认开始训练前会画一次图, 把前面 # 去掉就是不画图
+        unload_text_encoder: false
+        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
+        diff_output_preservation: false
+        diff_output_preservation_multiplier: 1
+        diff_output_preservation_class: person
+      model:
+        # huggingface model name or path
+        name_or_path: /root/FLUX.1-Kontext-dev 
+        arch: flux_kontext
+        is_flux: true
+        quantize: true  # run 8bit mixed precision
+        quantize_te: true
+        model_kwargs: {}
+        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                                             # 多少步测试一次图片, 建议和上面save_every相同
+        width: 1024
+        height: 1024
+        prompts:
+      # ⭐ 使用正确的----ctrl_img格式
+        - Convert the image to ChatGPT 4o style pixel art ----ctrl_img /root/style/sample/img1.jpg
+        - Convert the image to ChatGPT 4o style pixel art ----ctrl_img /root/style/sample/img2.jpg
+        - Convert the image to ChatGPT 4o style pixel art ----ctrl_img /root/style/sample/img3.jpg
+        neg: ''
+        seed: 42
+        walk_seed: true
+        guidance_scale: 4  #画图的CFG
+        sample_steps: 25   #画图的采样步数
+        num_frames: 1
+        fps: 1
+# you can add any additional meta info here. [name] is replaced with config name at top
+meta:
+  name: "Kontext_pixel_art"
+  version: '1.0'