Why fine-tune LLMs? GPT-4o fine-tune for PERFECT FLUX Image Prompts

[@stereotyp9991]

I just started training LoRAs for Flux. This is so much fun. You have to try this!

[@relaygus]

Thoroughly enjoyed the video. One piece of feedback though: the first half felt like it was all about the image generation. So much so I forgot it's a video on fine-tuning. I reckon shortening the demo of the image generation would've taught us the same about fine-tuning.

[@KyleFES]

Legend IndyDevDan, loving the uploads G 🧐

[@Adrian_Galilea]

In order to validate how good the results are you should compare the image generations of your original prompt vs the expanded gpt4o fine-tuned prompt. Some image models are very good with anemic prompts so the only way to evaluate would be comparing both of them. (I have done this in the past) Love your channel, share it constantly, keep it up

[@indydevdan]

100% this is how I validate. The next step is to setup (best for productionization) is to setup large promptfoo tests 50-100+ and run both models side by side.

[@MedyGames]

Dev Dan is up there

[@techfixer1543]

DevDan...you've got a new subscriber with me! Great content, need to binge your vidz. Your unreleased tool "Vision Grid" rocks! Do you plan to make it available in th future? Also, I learned a great deal in your other vid on Cursor AI. Thnk you hermano! Props and high-five to you! 🖐️

[@Nadia-AIInsiders]

Did you put your long instructions into a system prompt while fine-tuning? If yes, don’t you need to use the same system prompt when using the model, or does it work well without the system prompt after the tuning? I’ve experimented with fine-tuning 4o mini and 4o for content creation - it works much better than few-shot prompting, and the model captures the creator’s style really well, making fine-tuning an ideal choice for scaling content :)

[@mrschmiklz]

This is flux fine tuned specific gpt4o? What docs for the fine tune, if you know... Every image gen model will have specific keywords and formatting, correct?

[@monkeyfish227]

What hardware were you using to generate these sky diving pics? Was this in ‘realtime’? I am in for getting a new macbook at the end of this month and wonder what would work best.

[@amit111978]

Hi, nice video. Will you be able to share the dataset you used for finetuning?

[@uhtexercises]

I'm also interested in seeing the data set or parts of it. It seems to be the secret sauce for fine tuning and it's very hard to find proper instructions on the net on how to compose your data set.

[@indydevdan]

Potentially - working on a membership/community plan for sharing more of the 'secret sauce' of the projects we share and discuss on the channel. LMK if interested.

[@amit111978]

@indydevdan what all will be included?

[@stonedizzleful]

Quality content

[@TaskSwitcherify]

What is the max. resolution of images out of Flux? Also, won't your prompt-expanding feature convert people's unique and original prompts into the same "expanded" prompts for everyone? Thanks

[@newfrontiers5673]

But don't the fine tuned models cost like 5x more via the API?

[@BurkhardReffeling]

50% more, but you safe on input and output tokens in the end

[@MacS7n]

Couldn't you have got the same result using a prompt that takes your input and generate a detailed prompt for flux?

[@indydevdan]

Yes but you would need a larger prompt with examples, output structure, and instructions. That's one of the advantages of fine-tuned models - you save tons of tokens.

[@Ha77778]

👌

[@mrschmiklz]

apologies. i am commenting as i am watching. so generic instructions and not img gen specific model instructions. would be interested in seeing model specific fine tuning as, to my knowledge, most models are trained differently with keywords and formatting of prompts can be different