Brom Bresenham

Tomb Sweeper Devlog - Launch Retrospective Here's a look back at the year I spent developing Tomb Sweeper for mobile, from inception to launch. Includes my development timeline and the tech stack I used as well as insights and lessons learned from working with AI, choosing a backend service, and working with advertisers for both monetization and marketing.

@ScreenFlow please add a "preview your project in full screen" option besides "hide the inspector, hide the timeline, then go into full screen editor, then manually zoom in to about the right level while still showing a few UI borders". 😩

I figured how to compensate for the color shift I was seeing when generating a new video clip from the thumbnail of an existing video clip. The video clips were coming back with no color profile specified, but if I use ffmpeg to assign the srgb color profile then the video colors stay much closer to the original thumbnail when I import it into Final Cut Pro. My shell command to assign a color profile: ffmpeg -i input.mp4 -c copy -color_primaries bt709 -color_trc iec61966-2-1 -colorspace bt709 -color_range pc output.mp4

@NanoBanana 2 is approaching Adobe Firefly levels of NSFW detection accuracy. Here are the four results for the same image modification request via Higgsfield.

Toma Mi Mano Devlog - Color Drift These days the second biggest challenge for AI filmmaking is maintaining consistent environments. The current tools have pretty much solved character consistency, but keeping the environment consistent across shots can be tough - including not only physical structure, but like background extras walking into a shot from an empty area the camera just finished panning over. The biggest challenge is maintaining color consistency across multiple shots. I'll use Nano Banana 2 to modify a frame grab to create a start frame for another shot and the colors will drift slightly. Multiple consecutive edits drift a little further each time, as shown in the video here. Prompting Nano Banana to keep the current color temperature etc. - or to match the color grade of the original frame, before edits - doesn't fix it. And it's not just Nano Banana - I've noticed that e.g. Kling 3.0's videos will have a small amount of additional color drift from the original start frame. For now the solution seems to be to color correct my edited start frames in Photoshop and/or color correct video sequences in DaVinci Resolve. Both of which are tedious... the automatic options aren't sufficient. This is on top of the color profile hell I went through for a few days on Mac. I was originally using ScreenFlow but I noticed that just exporting a frame and reimporting it would cause a color shift in the frame. I tried switching to Premiere and frame export/import worked fine, but it it would color shift video clips. Finally I've switched to Final Cut Pro, which does fine in both of those cases and is a great editing experience. The problem with ScreenFlow and Premiere was that images and/or videos would be missing a color profile and the software would assume the wrong color profile. I could have also switched entirely to DaVinci Resolve, but while it's very full-featured and I need to use its color correction options anyways, it just feels slightly clunky to me.

💯

Toma Mi Mano Devlog - Can We Have a Serious Conversation? Looks like I'll prolly be switching from Kling 3.0 to Higgsfield's own Cinema Studio 2.0, or at least using both. When dealing with animated-style characters, CS2 does a better job of with realistic dialog and body language than K3 does. All along I've wanted the cinematic hyper-real stylized aesthetic of Pixar films without the cartoonish faces and exaggerated body language. Generating my first couple of characters was tough. All my image gens wanted to create either anime-style characters with black outlines or Pixar-style characters with cartoonish faces. Once I had my first few characters, I was able to use them as references to help additional characters match the same style. Today I realized I would have the same problem doing animation with Kling. Apparently if it sees Pixar-esque characters, it gives them Pixar-style dialog delivery and body language. Even though there won't be any audible dialog in my music video, there will still be some muted conversations I don't want my characters having melodramatic interactions. Here's the side-by-side test I did. I gave the same exact prompt to Kling 3.0 Omni (without specifying a start frame) and then Cinema Studio (which requires a start frame, which I took from the Kling render; also I set the Genre to Intimate). The first video here is Kling, the second is Cinema Studio. Here's the prompt I have to both, with image references: >> BartenderGirl tells LaptopGuy I have a question for you". Her statement is matter-of-fact. BartenderGirl looks serious overall but also has a wry half smile as she says it. LaptopGuy looks up at BartenderGirl in mild surprise and asks: "What is it?" BartenderGirl looks slightly puzzled as she asks LaptopGuy "can we have a serious conversation?" LaptopGuy looks slightly puzzled himself and replies to BartenderGirl "I don't know, can we?"

At this point I'm just spinning my wheels. While another few generations might give me a winner, I decided to try create character sheets out of some of the customers and prerender them into the reference images to improve my chances. That's what I've been working on, but after taking a few minutes to post these WIP videos, I realize I should take another stab at a couple of them with Kling Video Edit and just see what happens.

Take 33/34: randomly there is a sorority party at the bar. Looks pretty good otherwise and I might be tempted to go with it if it wasn't for some notable gaffes at the beginning (two guys sitting inside the service aisle) and end (two women appearing out of nowhere, the first more obvious than the second).

Toma Mi Mano Devlog: Bar Scene I've been working on a new song called Toma Mi Mano and a GGI-animation-style AI music video to accompany it. I'm working primarily with Nano Banana and Kling 3.0 via Higgsfield. An establishing shot for a seaside restaurant bar is proving difficult to nail down. I've been burning a lot of credits on prompt iterations and regens, with nothing to show for it except for some occasionally hilarious results. What's supposed to happen: camera follows Bartender Girl down the bar's service aisle. She walks down to Laptop Guy who is sitting at the far end on the left side, absorbed in his work. There are 2 other people at the bar (near right side) and 2 people at the table on the far left. Here is my 3rd attempt (of 34 so far), which looks great and is almost good enough EXCEPT that no one has any drinks in a restaurant bar. There are other small AI-generation issues, such as the mangled hands of the guy on the far right, but I could live with that. But people should have drinks. It's on me though. I only specified that people should be eating, I did not say that they should be eating and drinking. It was so close in the beginning, and I've been chasing the dragon for a couple of days since then, trying to recreate this scene with that one small improvement.

Nano Banana and/or Higgsfield: you're killing me here! 😂🤣 Both image generation attempts are from today.

I think I'll go with the response on the left. 😉😂

Tomb Sweeper Devlog - iOS 15 Crashes Turns out BGFX's minimum iOS target is 16.0, which explains the crashes on iPhone 7 Plus, which is capped at iOS 15.8. Since BGFX builds as a '.a' lib and is pretty light (extremely light, dare I say feather-light) on the iOS documentation, it was non-obvious. I tried changing the BGFX build scripts to target 15.6 instead of 16.0, but got a different kind of crash when running ("*** Terminating app due to uncaught exception 'CAMetalLayerInvalid', reason: 'invalid pixel format 70'"). So I'll bump the minimum iOS version to 16.0 and release an update.