We built a compiler that turns intent into a deterministic blueprint. The architect doesn't start until the user approves (and you can interject at anytime to make updates or reshape the blueprint).
Very cool. I really need to get back into recording to my cassettes. It was a lot of fun just as a way to explore music and then burn what I liked into a fun format. This project has not gotten a lot of love from me since some life changes but this is what I was using to generate printable J-Cards from my navidrome instance.
What a wonderful space you have created. I have so much mindlessness in where I have things placed. I feel inspired to be a bit more purposeful and organize some things now.
Very neat. I have a love-hate relationship with a unique type of digital sensor called the Foveon X3, which needs a particular (very sluggish and limited capability) software made by Sigma to convert the RAWs. Recently I have been using AI to create a converter that processes the RAWs in a similar fashion with some improvements.
This is one of those cases where its a means-to-an-end (I just want to take more photos without being bogged down) and less of a project for me to learn how to reverse-engineer and design signal processing pipelines. I am very grateful for what recent models are enabling me to do.
Everyone has a love/hate relationship with Foveon, that's the beauty of the sensor isn't it ? Would you mind sharing your project ? That's awesome. Have you been able to compare crispness of files ? Dng compatibility from latest cameras, although it results in huge files, was pretty useful but files were not as sharp as native X3F developed in Sigma Photo Pro.
Likely because SPP bakes in "detail recovery" with a high pass filter in its early processing stage alongside noise suppression. SPP does a lot beyond prepping and calibrating the raw layer data. Some aspects of that can be desirable but for portraits I find the lack of control over that processing step to be a real downer.
Have you tried a Foveon sensor camera for film scanning? I was wondering about it recently and how it would perform - the easily removable IR filter and different way the sensor works could be really interesting.
Remind us what the Foveon sensor's quirk was? Something in the subpixel layout as I recall. A non-bayer pattern since abandoned and largely unsupported, yet effective somehow. But maybe I'm imagining this.
Foveon sensor is composed of three Red/blue/green layers, so each pixels know exactly its color, while the Bayer sensor used by other manufacturers is composed of pixels that are either red, blue or green, and that implies some reading out of its color and its neighbour's one... and some glitches (typically moire or purple fringes). In order to reduce those, a Low-pass filter that blurs the picture a little bit is added on top of the sensor. Therefore, foveon pictures are way sharper, with true colors. This is the theory for colors... in fact it's a very hard to tame sensor and results can be very random. Plus the cameras are very slow because they have to process 3x more information. Sigma being a very very small company versus the other big guys, their R&D is limited. Still, it's a very nice camera to use for geeks who takes time to understand it. You hate it or love it. No in-between.
Hey Devin! There's a blog called Leicaphilia written by a—sadly now passed—photographer named Timothy Vanderweert who I would argue was actually more attached to the Sigma Foveon than he was to the Leica. Kind of ironic.
I think you'd like the blog in general, but the Foveon sensor posts are definitely worth reading if you're curious about it: https://leicaphilia.com/?s=Foveon
What improvements have you made? Processing RAW sluggishness was always an issue from de-Bayering. Before RAW, my hell was DPX image sequences. This was pre-SSD, so I/O was painful opening/closing individual frames in realtime. The best improvement of processing RAW I've experienced was NVMe SSDs.
Foveons don't actually de-bayer bc they use a stacked 3-layer sensor, where each layer is sensitive to different colors. The top layer is RGB, the middle is just RG, and the bottom is just R. We calculate the three values by subtracting the layers from each other.
Traditionally all processing of these X3F RAW files had to go thru software provided by Sigma, and its clunky and bakes in some processing you can't really turn off. Some folks (myself among them) are focused on creating new ways of processing and converting these X3F files to be used more flexibly. I have some links in the threads other replies to some changes I made to one processor.
I was able to significantly improve the highlight recovery. It's kind of a game changer since the sensor architecture requires a LOT of light to get the best results. To minimize noise you expose to the right at least one stop and then pull down in post. With the new processing I am able to get pleasing results closer to +1.7 and sometimes even 2ev. This sensor was not known for having great dynamic range, but with extra processing you can tease some information out into something attractive. It does leave the upper end of specular highlights flattened compared to a proper exposure but it often looks "believable" so I'm happy with it.
It never ceases to amaze me the pushback to ETTR I still get to this day. Every sensor has its own limits to how much can be pushed. People just don't want to work and just shoot down the middle. My favorite quote from a lazy editor "I don't work that way" when asking why they didn't help the photog out and save some pushed exposures rather than just omitting the shots from selections.
It makes sense to me if you are a mostly JPEG shooter and just keep RAWs for archiving. It also is less pleasant to look at the preview when using an EVF if its 30% blown. My photo style is often ETTL bc I like to crush shadows in a chiaroscuro kind of way. Also modern cameras look great at higher ISOs, and if you need it to be cleaner bc of large prints, AI powered de-noisers that operate on RAW files are also incredible.
One thing you can do with a raw workflow is use the EV compensation metadata to compensate the render you see. So you can shoot +1ev and then automatically reduce exposure by that much when viewing. This is a decent ETTR workflow and its never been easier to whip up a custom culling/importing tool.
That said, its basically a requirement to get proper signal-to-noise on a Merrill. The LCD-screen is dog-water anyway so its not like you can see anything of value on it other than blurry composition. So for those reasons ETTR is very natural. If you want to be extra careful you can spot meter with EV compensation , but I find the evaluative mode works pretty well enough.
Some aspects of the process seem more like labor than a form of self-expression. I think if you are into the limitations of shooting film photos it makes sense.
As someone who printed B&W in a darkroom until last year, yeah that. My sanity seems to have been improved by selling it all and just shooting digital.
Hell I even shoot jpegs with Nikon picture control baked in now. No post editing at all.
They state it's for HBM wafer inspection equipment:
OKI Group printed circuit board (PCB) company, has successfully developed 124-layer PCB technology for wafer inspection equipment designed for next-generation high bandwidth memory, such as HBM
Since the latest semiconductors process an enormous number of signals and the number of wafer-mounted chips increases due to process miniaturization, it is necessary to increase density and more layers on the PCBs used in inspection equipment.
In their linked semiconductor testing page[1], they mention conductive paste. So seems they test whole wafers by having the whole wafer put on such a PCB, using conductive paste instead of solder or similar, and the high layer count is then to be able to bring out all the connections (pins) from each chip to the testing circuits while maintaining the high signal integrity required for the high-speed signals.
HBM makes it worse since it is fast due to its very high bus width, 1024 bits wide, moving to 2048 bits wide[2]. So to test a HBM chip on a wafer before its cut and bonded to a GPU chip or such, you'd need to bring out those thousands of connections per chip on the wafer. Hence the need for so many layers compared to a regular PCB.
The number of levels of traces and gates laid out is an important aspect of how much sophistication and capability available to connect ICs on a pcb. You are laying out power and ground and communication. There are lots of other considerations like the complexity of what you can plan to layout using software. But you stack layers of traces to add capability. More layers is naively more capability.
i guess, but to me its not so much "this is ai this is terrible" and more "i'm sick of hearing this phrase all the time". it's not that I never want to see an em-dash ever again, its just that i don't want them used every other sentence.
reply