When it comes to the progress, the biggest roadblock is probably the vision capability of the model. It seems to miss assorted details from game screenshots that are provided, such as aesthetic flourishes and the layout of panels. The AI can replicate most words, identify monsters and furniture and make a game bible that includes information from decompiling the game and the manual.

Establishing the gameplay and the functionality of the interface is coming a bit more easily. With that, you can simply play the OG game and the recompile side by side in identical scenarios, then see where they diverge. For example, how the buttons interact with items on the floor. The AI made it so that the player can use the Holy Hand Grenades found on the floor. The OG only allies them to be used during combat. It is easy to iterate on such QA issues.

In some ways we were able to establish some QOL features as we went. One such example is an indicator for how many grenades and 'parleys' the player has, alongside with making the Blast spell a button option. In the registered version of the game, you could sacrifice health to cast a spell, but you had to use ALT+F1. Now we got a dedicated button, that also tracks how much health it costs to take this action.

The second picture here, is a mockup that improves over the AI's own attempt at replicating the OG, which I then passed back to the AI to be implemented. The AI botched it, as you can see from the strikethrough and mismatched furniture picture. That sucks, it takes the AI from half an hour to a couple hours to investigate and implement changes on my hardware.

To get more visual authenticity and cleanliness, probably requires a better vision projector, or perhaps remastering the assets inside ComfyUI. Hopefully, Qwen4 Flash can remove the need for visual remastering.