[HN Gopher] Building an agentic image generator that improves it...
       ___________________________________________________________________
        
       Building an agentic image generator that improves itself
        
       Hey HN! We recently graduated from YC, and have been building
       customer personas for large e-commerce companies. We recently
       expanded into the image generation space, and have been working on
       research about how to automatically improve the quality of
       generated images.
        
       Author : palashshah
       Score  : 51 points
       Date   : 2025-05-21 13:12 UTC (9 hours ago)
        
 (HTM) web link (simulate.trybezel.com)
 (TXT) w3m dump (simulate.trybezel.com)
        
       | average_r_user wrote:
       | Quite interesting, do you have some documentation of your
       | platform and capabilities? Your landing page is quite synthetic
        
         | palashshah wrote:
         | hey! we're working with an initial set of customers, and plan
         | to launch full capabilities soon. stay tuned :)
        
       | ramesh31 wrote:
       | This is a wonderful writeup of building a simple agentic system
       | in general. What OP describes is more or less the bare minimum
       | you should be doing at this point to get good (consistent)
       | results from an LLM; single-shot prompting is a thing of the
       | past.
        
         | palashshah wrote:
         | appreciate the compliment! yep, it's definitely necessary and
         | is the bare minimum for building image generation systems in
         | production.
        
       | shmoogy wrote:
       | I'm surprised you landed on using o3 as the judge - we found it
       | way too expensive. I use llm as a judge for generating color
       | variations of products, definitely hoping for some improvements -
       | it can be brutal to get non hallucinated features along with
       | proper final rendering.
        
         | omneity wrote:
         | Have you tried open weights vision models such as Qwen VL,
         | MiniCPM, PaliGemma...?
         | 
         | I'm also curious how usable are simpler vision models such as
         | Florence in case you explored this direction.
        
           | palashshah wrote:
           | we're currently in the process of doing this. i think
           | something that could potentially work is to iterate upon the
           | initial image composition / structure using cheaper models,
           | and then upscale at the end. this way you're saving on that
           | iteration cost, but eventually land on a higher-scale image.
        
       | elif wrote:
       | This is great and provides a good starting point for any similar
       | efforts.
       | 
       | However I think the temptation to lean all tasks on AI is perhaps
       | a little naive if not lazy.
       | 
       | For mask generation, there is really not much reason to use AI.
       | In this example, simple stochastic blob detection, a trivial
       | function you could get from openCV or ask a college sophomore to
       | write would generate much better quality masks.
        
         | palashshah wrote:
         | totally agreed here. i think my goal primarily with the mask
         | generation was to test out how effective openai's capabilities
         | were.
         | 
         | we're currently working on pipelines that limit the the
         | involvement of AI to various tasks. for example, when
         | generating an ad there's usually logo, some banner text, and
         | background image.
         | 
         | we can use gpt-image-1 to generate the background image,
         | another LLM to identify the coordinates of where we place the
         | logo, and just add the logo onto the image. this is just one
         | example!
        
           | jackphilson wrote:
           | Why do you agree? I think we should outsource as much as we
           | can to abstraction. We've been doing it forever.
        
       | mentalgear wrote:
       | Again another example of "the unreasonable effectiveness of LLMs
       | in a loop". At with time, the tasks for loop become bigger and
       | more complex, until we find ourselves "outlooped" at least job
       | wise.
        
       | ramoz wrote:
       | Nice retrospective but I guess this process is no longer needed
       | as model's get better; esp as they start enabling features like
       | consistent subjects. Seems like a lot of overhead to correct text
       | for inspirational images, but I can imagine you need to always
       | present some form of _quality_ to your clients.
       | 
       | Feel like control nets and some minimal photoshop work would've
       | been better.
        
         | palashshah wrote:
         | totally. it got to a point where most of the text generated in
         | our images was incorrect, and so it wasn't a great look showing
         | that to our clients.
         | 
         | we're actually working on some form of what you described where
         | we take images generated from LLMs + add consistent logos
         | discretely rather than generatively.
        
       | abshkbh wrote:
       | Palash this is a great post, I learnt a lot as an image gen noob!
       | Keep writing more :)
        
         | palashshah wrote:
         | this is incredible to hear! i plan to keep writing on a weekly
         | basis, and will be posting them on twitter.
        
       | t_mann wrote:
       | I was kind of hoping this would be in the 'Dreambooth mold' of
       | finetuning open weights models. I have used that with some
       | success some ~2 years ago, does anyone know what improvements
       | there have been in that direction since Dreambooth?
        
       | zahlman wrote:
       | It's frankly amazing to me that "ask another LLM to evaluate the
       | image" actually produces useful feedback that results in actual
       | improvement from the first LLM.
       | 
       | But then, I guess it's not much different of an idea from the
       | earlier use of GANs, or of telling LLMs to "stop hallucinating",
       | etc.
        
         | palashshah wrote:
         | totally. the way i think about it (purely based on intuition)
         | is that asking an LLM to do understanding + image generation is
         | too complex for it to be effective. if we separate out the
         | tasks into discrete steps, the evaluation becomes better, and
         | the generation simply becomes instruction following.
        
       ___________________________________________________________________
       (page generated 2025-05-21 23:01 UTC)