2025-10-18-old-man-fit
詳細
ファイルをダウンロード (1)
このバージョンについて
モデル説明
⚡ Flux Model Training Summary (LoRA)
This job customized the Flux base model, which is a different architecture from Stable Diffusion XL. This run featured several key changes in its configuration, resulting in a longer duration.
1. Job Status and Timeline
Base Model Used: Flux (A modern, likely newer or different image generation model).
Job ID: 563759-20251018091302678
Training Start: October 18, 2025 at 05:13:06 PM
Completion Time (Ready): October 19, 2025 at 12:03:14 PM
Total Duration: Approximately 18 hours and 50 minutes (significantly longer than the previous job).
Dataset Used: 27 Files / 27 Labels.
Label Type: Caption (using full descriptive sentences, rather than just short tags, which is ideal for a complex model like Flux).
2. Key Training Parameters (The Differences)
This configuration is notably different from the previous SDXL job, suggesting an attempt to train a very lightweight, highly specialized LoRA.
Model Intensity & Focus
Network Dimension (networkDim): 2 (This is extremely low compared to 32 in the last job. This makes the resulting LoRA file much smaller but potentially less capable of capturing complex details).
Network Alpha (networkAlpha): 16 (Used for LoRA stabilization).
LoRA Type: lora.
Engine: kohya.
Learning Process
Image Resolution: 512 (Lower resolution than the previous 1024).
Maximum Epochs: 35 (Higher than the previous 20, meaning the data was shown to the model more often to compensate for the small network size).
Text Encoder Learning Rate (textEncoderLR): 0 (This is crucial: The text understanding part of the model was frozen and did not learn. Only the image generation part (U-Net) was updated).
U-Net Learning Rate (unetLR): 0.0005.
Optimizer Type: AdamW8Bit (Different from Adafactor).
Shuffle Caption: false (Captions were fed in the same order each time).
Keep Tokens: 0 (No specific caption words were prioritized, unlike 3 previously).