Find the best Epochs after training your Flux Loras [Training Tools]

แชร์
ฝัง
  • เผยแพร่เมื่อ 13 ม.ค. 2025

ความคิดเห็น • 1

  • @wolpumba4099
    @wolpumba4099 หลายเดือนก่อน +1

    *Tips and Tricks for Streamlining Flux Lora Model Testing in ComfyUI*
    * *0:03** Introduction:* The video focuses on speeding up the process of testing LoRA models trained using various methods.
    * *0:27** Prompt Automation:* The creator recommends using a "caption collector" tool to create a prompt list from the training dataset. This allows for automated testing with prompts relevant to the training data using custom nodes like "zenai prompt V2" in ComfyUI.
    * *2:09** Organizing Intermediates:* Create a folder (starting with "A" for easy access) to store intermediate LoRA models (e.g., Epoch 3 through 10).
    * *2:40** Workflow Overview:* The demonstrated workflow uses the same seed, prompt, base model, and settings for each run, only changing the intermediate LoRA model (Epoch) to ensure a fair comparison.
    * *3:32** Default Settings:* The example uses Flux Schnell, T5 clip, 12 steps, Euler a sampler, and a CFG of 3.5.
    * *4:04** Image Comparison:* The workflow utilizes "CR image compare" nodes to create comparison charts of different Epoch outputs, visually highlighting the differences between them.
    * *5:37** Accuracy vs. Aesthetics:* The creator emphasizes the choice between prioritizing accuracy to the training dataset or the overall aesthetic appeal of the generated images, noting that the best-looking image may not always be the most accurate.
    * *6:16** Epoch Selection:* The creator suggests that Epochs 1 and 2 often have a strong bias towards the base model. They generally recommend Epochs 4-7 for good image quality and 8-9 for accuracy to the training data.
    * *7:00** Utilizing Different Clip Encoders:* The workflow can be modified to use the full T5 clip encoder for potentially better results, although it can increase processing time.
    * *17:52** Verifying LoRA File Integrity:* It's crucial to ensure that all downloaded intermediate LoRA files are the same size to avoid errors caused by incomplete downloads.
    * *18:37** Sequential Queuing:* To ensure a strictly sequential processing order, it's recommended to wait for the current queue to complete before adding more jobs from different tabs.
    * *20:47** Early Epoch Considerations:* Early Epochs might produce visually appealing results due to a stronger influence of the base model, but they might not accurately reflect the learned features of the LoRA.
    * *24:33** Identifying Outliers:* Be aware of outlier images that deviate significantly from the expected output. These can be caused by various factors and might not represent the overall performance of a particular Epoch.
    * *34:50** Raw Shacks Clip Exploit:* The creator briefly mentions a technique they call the "Raw Shacks Clip Exploit" or "Vision Game of Telephone" which involves using abstract images and intentionally misleading captions to create unique artistic styles.
    * *40:54** Maintaining Consistent Comparisons:* Avoid changing generation settings (CFG, steps, samplers, etc.) between Epoch comparisons to ensure a fair assessment of the LoRA's impact.
    I used gemini-1.5-pro-exp-0827 on rocketrecap dot com to summarize the transcript.
    Cost (if I didn't use the free tier): $0.03
    Input tokens: 24361
    Output tokens: 670