I published the model, download here:
https://civitai.com/models/1401415
and here is an example workflow:
https://civitai.com/models/1401477
#comfyui

Finally got around to testing the ChatGPT's new image features (in case you slept through the last week, you can now edit and create pretty much any image you want just using words), and it all looks pretty cool. But let's dig deeper, and see what the drawbacks are. I'll compare it to #comfyui using two examples:
1. Winter render
2. Filling an empty interior with furniture

But if you look closely, you can see another problem, ChatGPT changes the proportions of the objects. There is no ControlNet and therefore it's hard for it to keep everything exactly as it is on the source. This is important, because the details in any case in many places are incorrect, and they need to be erased in Photoshop to adjust with the original image. Unfortunately this can not be done with the result from ChatGPT. Because of which it seems there is no sense at all to use the model for such tasks.

In this case ChatGPT has a big advantage in that the workflow is extremely simple, you throw everything together and tell it what to do. But the same problem, the original room changes in details and proportions. I would not say that for this task such changes are critical, for example, if we want at the initial stage of the project just to estimate how the interior will look like, ChatGPT gives an idea of it.

And yes, no model can handle some complex objects, the carpet is different here and there, but GPT is closer to the original. Also GPT better fits the lighting in the scene.

In #comfyui, the proportions of the interior and its original details are, of course, in place. But there is no speed advantage here, because you have to insert items one by one, not all at once as in ChatGPT, and wait for rendering each time. Besides, such workflow can be built only on Flux, and it is not fast by itself.

There are a lot of “Photoshop is dead” posts on the internet right now - you can edit images in LLM using just words. The story here is that even if the accuracy will be corrected in the future, the input of commands using text or voice is not suitable for any complex editing. Actually, people in the office do not communicate with each other in this way: to explain something properly you need sketches, masks, marks... in short, the functionality of Photoshop, otherwise no one will understand the task, neither a human nor AI (even if the AI will be smarter than a human, the problem here is not in intelligence, but in the fact that not everything can be formulated in words in a reasonable time). So AI model is not enough, you also need an interface with tools for graphical editing, that is AI inside the analogs of Photoshop, Revit, Rhino and everything else.

But, overall, sure cool, I still have some ideas on what can be done with these new ChatGPT functions