Flux Dev — open-source fine-tuning

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-modelx-flux-dev-finetuning-creators

1. What technical reason makes Flux Schnell unsuitable for stable fine-tuning?

   1. It automatically generates incorrect captions for training data
   2. It uses a distilled model that loses too much original capability during training
   3. It requires more than 100GB of VRAM to run
   4. It lacks the necessary API endpoints for custom modifications

2. What is the recommended number of consistent-style reference images for training a Flux Dev style LoRA?

   1. Between 200 and 500 images
   2. Exactly 100 images
   3. Between 5 and 10 images
   4. Between 20 and 50 images

3. In LoRA training terminology, what is the purpose of a trigger word?

   1. To prevent the model from generating harmful content
   2. To reduce the file size of the resulting LoRA
   3. To automatically caption the training images
   4. To invoke the trained style when generating new images

4. Approximately how many training steps are recommended when training a style LoRA on Flux Dev?

   1. Around 100 steps
   2. Around 500 steps
   3. Around 10,000 steps
   4. Around 1500 steps

5. What commercial restriction applies to outputs generated using Flux Dev weights?

   1. Commercial use is completely prohibited with no path to authorization
   2. Outputs can be sold freely but the LoRA itself cannot be sold
   3. Commercial use is permitted but requires a monthly subscription
   4. The weights ship under a non-commercial license, requiring negotiation for selling outputs

6. What happens when you attempt to apply a Flux Dev-trained LoRA to Flux Pro for production renders?

   1. The LoRA works but produces lower quality output than on Dev
   2. The LoRA works identically since Pro and Dev share the same architecture
   3. Flux Pro does not accept LoRA adapters — you must use it as a style reference instead
   4. Flux Pro automatically converts the LoRA to the correct format

7. Which tool is mentioned in the lesson as being used for training LoRAs on Flux Dev?

   1. TensorFlow Lite
   2. kohya_ss or ai-toolkit
   3. PyTorch Lightning
   4. HuggingFace Diffusers only

8. What is the approximate training cost for a Flux Dev LoRA on a single consumer GPU like an RTX 4090?

   1. Between $50 and $100
   2. Less than $1
   3. Over $200
   4. Between $5 and $15 in GPU compute costs

9. What distinguishes Flux Dev's position in the Flux model family from Flux Pro?

   1. Dev is faster but lower quality, while Pro is slower but more accurate
   2. Dev can generate video but Pro cannot
   3. Dev requires a subscription, while Pro is free to use
   4. Dev offers open weights suitable for LoRA training, while Pro is a closed commercial API

10. Why is it important to keep your trigger word consistent when collaborating with others on a LoRA project?

    1. To prevent the LoRA from forgetting the style
    2. Because changing the trigger word corrupts the LoRA weights
    3. So collaborators can reproduce your exact results with the same prompts
    4. Because trigger words are encrypted into the LoRA file

11. What does the 'rank' parameter control in LoRA training?

    1. The learning rate schedule during training
    2. The dimensionality of the low-rank adaptation matrices added to the model
    3. The number of images in each training batch
    4. The resolution of images the LoRA can process

12. What is the primary advantage of training a LoRA over using reference images alone for consistent styling?

    1. The LoRA works with any model, not just Flux
    2. The LoRA is free to use while reference images cost money
    3. The LoRA produces higher resolution outputs
    4. The LoRA allows hundreds of renders with identical aesthetic DNA without per-image reference setup

13. What is the recommended workflow for producing final commercial images after training a style LoRA on Flux Dev?

    1. Use Flux Dev for everything since it's the highest quality
    2. Export the LoRA to Stable Diffusion for commercial use
    3. Train the LoRA on Flux Pro directly
    4. Apply the LoRA to Flux Dev for iteration, then run final prompts on Flux Pro using fine-tune endpoints

14. What happens if you sell images generated with a Flux Dev LoRA without negotiating a commercial license?

    1. Nothing happens because LoRA outputs are exempt from the base model license
    2. Black Forest Labs automatically charges you retroactively
    3. You would be violating the non-commercial license terms of the Dev weights
    4. The images are automatically converted to a commercial license

15. What is the primary reason to train on Flux Dev rather than directly on Flux Pro for LoRA development?

    1. Dev allows local training on personal GPUs while Pro is closed and API-only
    2. Pro does not support the LoRA training process
    3. Pro produces lower quality LoRAs than Dev
    4. Dev training is faster than Pro training