This week the team completed both expansions, Heart of Winter and Trial of the Luremaster. They’re making a fast pace through the game, and have grown to be pretty fearsomely capable. We’re setting up for the endgame, just Lower Dorn’s Deep and the return to Easthaven. Funny thing, for those familiar with Lava’s Rediscovery of Kuldahar mod. I won’t be getting the star gazer’s special atlas this run, I’ll finish the game before he finishes the book.
Leezel does occasionally use her Long Sword of Action instead of a mace. She can do serious damage any way, and is easily the teams best tank.Turpin and Hildy have grown very close, and will not part.Leezel and Grace are as close as sisters, and a very dangerous front rank.Gretchen sweltering under a desert sun. The team spent a couple days in the Anauroch Desert, about 100 degrees different than what they’d grown used to.The Luremaster’s riddles were mostly in Hildy’s purview, and Turpin and Henri were a little help!
Soon enough we were back in the frozen north.
“Just, give me a minute… there’s traps everywhere” Henri was kept busy as we got close to a dragon’s lair.Grace is not our biggest warrior! But she is skilled… and cuts in every direction at once.
*****
I should be back early next week with the conclusion of this run.
I really like the style of that second image. Wish I could figure out how to stabilize that more. Even using specific artists and such it seems like there’s a lot of variation, and sometimes I stumble on something I really like but can’t consistently reproduce in other scenarios.
That last dragon picture is great! I love that composition. The movie is fun too. The pause for a frost breath is a really nice touch.
I’ve been experimenting more with what I discovered about CoPilot re-writing queries before passing them to DALLE-3. So now sometimes I do my queries the old way, and sometimes I do them the new way – writing as if describing things to a human artist. I’ve found when the AI is completely failing to do what I want one way, switching to the other often helps.
I’ve been using “Clyde Caldwell style fantasy art” as the style prompt exclusively on this run. But yeah, as you say, it is often all over the place on what it chooses to do.
I think all I actually said for the icy breath was “pauses, and let’s out a breath”. It interpreted the icy cold itself! Some of the later video renderers are getting much smarter (like the dropped, falling staff in the previous set). This is an area of much improvement just in the couple months since I started doing this.
it does seem that *how* you say things to the renderer does not always matter in the ways you think it will. Presumably the AI re-writing the prompts has something to do with this. I’ve not figured out how to work this to any advantage.
It actually does make a big difference sometimes. Reviewing the “peek” I get with the CTRL+A,CTRL+C trick I found that many times when it went off the rails its because the text translator either misunderstood and mangled the DALL-E prompt or ran out of characters and dropped important stuff. Rephrasing it in a way that gets through properly helps a lot there. The big lesson was that there are actually *four* AIs between you and your image:
AI Text Filter decides if it objects to words in your text
ChatGPT AI translates your query to a “natural language” query for DALL-E
DALLE-AI generates an image
AI Image Filter decides if it objects to the image
So when things go wonky the trick is figuring out which AI is getting psyched out.
In yesterday’s episode my most difficult renders were a simple hug – which was solved by re-writing the query to make explicit who was who, and Jaheira’s introduction – which failed completely in my usual query style but worked much better the new way.
DavesGaming 
Leave a reply to Zeno Cancel reply