Llaasseerr

Just tried the old link, but I'm only seeing the old files. Do you have a new link? 🙏 EDIT: my bad, downloading now 😅

Your last round of test files was instructive in that I still found the clipping point in each clip, and it was uniform across all clips within a small level of variance. So I'm expecting with exposure differences in a new data set, that the clipping point will also remain uniform. Still, it's always helpful to have another tranche of test clips to play with, so bring it on! I don't want to push it, but it would be interesting to have similar tests with internal DNG footage if you're up for it.

Thanks for flagging that! It makes me more hopeful that this isn't a widespread issue. Plus if pushing by only +2 stops at ISO 800 would mean more bits able to be allocated at the low end.

Thanks, I revisited your ISO equivalence table to better understand how the values map from the Ninja to internal as per your tests. If I factor out that Atomos may have the gain levels set incorrectly and also that the false colour is not behaving predictably, what I see is the same exposure and highlight clipping point (except for 3200, see below) and the only difference is with noise level. I realise that you are stopping down the lens when changing the ISO level to compensate (I can see the iris change), which is really helpful in this case, so that means given the highlights still clip in the same place, the noise level is the only variable. The ISO 3200 clip was the only one with inconsistent exposure, which seems to be due to the switch to the higher base ISO 640. To make it consistent with the ISO 1600 clip, I adjusted the exposure by 1.67 stops (based on underlying ISO 200 vs 640), then lowered by -1 stop to go from 3200 to 1600. When accounting for the exposure difference, the highlight clipping point is about the same as ISO 1600, but with slightly more gain - so there doesn't seem to be an advantage to using the higher base ISO, meaning there was not a sudden drop in image noise when switching. As I mentioned in my post above, from my tests I think the sweet spot is ISO 6400 (internal 800) based on the reduced gain level of the noise. What do you think?

I just wanted to update this since I was really tired when I posted it. I made the observation that the Ninja V ISO 800 recording (internal ISO 100) +2 stops has what looks to me as the same gain level as the Ninja V ISO 6400 recording (internal ISO 800). Obviously it's not that useful to push +2 stops if the image then looks as noisy as +3 stops ISO (800 vs 6400). It suggest that the 6400/800 image has better noise performance than 800/100 if pushing 800 to 6400 +3 stops. I've taken another look, and this time I lowered the exposure of the 6400 image by -2 stops and compared to a similar luminance level area of the 800 image. Definitely the noise is a lot cleaner! Note also that if this was -3 stops to reflect the exposure rating difference between ISO 6400 and 800, then the noise would be even cleaner. So it seems best to shoot at 6400 (Ninja V)/800 (Internal DNG) and push +2 stops if you want, as per @Ryan Earl. I checked 12800/1600 -1 stop vs 6400/800, and they seem to have the same gain structure, so the sweet spot seems to lie with 6400/800. @Ryan Earl do ISO 800 internal DNGs also have lower gain than ISO 100, but with the same highlight clipping point? That would align with the fact you like to shoot at 800 internally and push +2 stops. That seems very good. The caveat with the PRR tests is that there's no highlight recovery as far as I know, so if this gain behavior hold true, that may give an edge to the internal DNGs.

Meant to say 1/3 to 1/2 stop below Alexa sensor clipping point, but could not edit the post. This is obviously a very good result, if the shadow noise/detail is acceptable! And it might be higher with internal DNGs with highlight recovery in Resolve.

I can assure you that all that extra range is appreciated and used if you're not just shooting to edit for Rec709 delivery. Definitely in post and vfx it makes a huge difference, but also just for highlight rendering vs clipping. Just doing an accurate film emulation will very aggressively roll off the highlights in a way that would break the image in most 12 stop cameras. I always notice how clippy bright scene elements look for shows shot on Venice on Netflix because it doesn't have that extra range.

I've had a chance to do some tests, here's what I found (let me know if this lines up with your own findings). I'm assuming -3 stops to correlate the Ninja V recordings to internal ISO values. I ended up converting from PRR to Prores 4444 through EditReady, since it gave me consistent exposure for all clips except for EI 3200 (internal ISO 400?) which was a little darker. For some reason I had an issue with Assimilate Play Pro (possibly I made a mistake). I exported as Alexa LogC/AWG and reimported to ACES. The highlights all clipped at around the same linear floating point value, in the 10.5-11.5 range which is a bit lower than the theoretical max based on the DNG clipping point. So the major difference is the noise floor. The most low noise (cleanest) image was EI 800 (internal ISO 100?). It has about 2 stops less noise than EI 6400 (internal ISO 800?), so it can be pushed +2 stops and have about the same noise profile. Here's a screengrab showing that the noise level is about equal for +2 stops more highlight capture. This puts the fp about 1.3 to 1/2 stop below Alexa highlight clipping point. For comparison, I have a DNG ISO 800 shot I found and the DR in the highlights is clipping at about 17-18. This extra range must be because of highlight recovery in Resolve. So I would need to do more tests to determine what the relationship is between the DNG raw and the ProRes raw recordings, and if internal is better for pure image data. Here's the internal ISO 800 clip I'm referring to: https://drive.google.com/drive/folders/1-z-R-8377SCVI6OxVVa5AybmcerEpQ9h Also, here's the clip I was referring to earlier, when I called out the green patches and quantizing in the shadows. https://drive.google.com/drive/folders/1PqyifY3J9HLKEUTUhIr75SJU3uBKo8cv

As I mentioned, I'm definitely interested to understand the real world differences between 100 and 800 if I can push the 100 to 800 in post. So basically, assuming ISO invariance. Like, what - if any - are the differences between the noise and the highlight clipping. It sounds like in your experience that there is an advantage from shooting 800 and only pushing 2 stops though, which sounds positive. 800 is obviously a comfort zone, given the Alexa and how other cameras have begun to settle on that as their rated base ISO even if the sensor native ISO is different. There is a nice indoor scene of a musical duo shot at 800 in that raw footage I linked to off YouTube, and it does look good. It does seem like the highlights are clipping slightly higher than if ISO 100 was pushed +3 stops, but it could also be because of some random behavior by the highlight recovery in Resolve. I would need to do more tests. https://drive.google.com/drive/folders/1-z-R-8377SCVI6OxVVa5AybmcerEpQ9h Also, here's the original footage I screengrabbed where I saw the green tint and posterization in the shadows of the hair when I pushed the exposure up: https://drive.google.com/drive/folders/1PqyifY3J9HLKEUTUhIr75SJU3uBKo8cv

Thanks, this is very helpful! I can't believe how well this camera holds up when pushed 4 stops in daylight. Even the 7 stop push is usable with Neat Video, assuming a film grain emulation would be added afterward. What I have not been clear on is if there's any effective reason to shoot ISO 800 vs just pushing ISO 100 by +3 stops, assuming ISO invariance. Definitely keeping an open mind on that though. I've only been able to check this with ISO 100,125 and 160, but the highlights clip at the same place if you take into account the ISO variation. But I haven't had a chance to confirm the behaviour of 200, 400, 800. This is assuming DNG images, not PRR from the Ninja V that I will check out with the files @OleB uploaded.

Thank you, this is freaking awesome! Will take a look. Also, thanks for the note about the metadata. So for whatever reason, Atomos decided to rate it 3 stops over. Can I assume that the higher base ISO is kicking in at 3200 (25600 in Atomos metadata)? Do you think that this can be converted to DNG equivalence just by subtracting 3 stops? I'm unsure, and also I don't know if one or the other clips the highlights at a different value either.

Thanks for your real-world feedback. Yeah, I think my math was not sounds because I know the sensors are capable of capturing significant HDR values depending on the sensitivity and whether using ND filters. If middle grey is pegged at a code value of 727/4095 in a normal exposure, then that only gives about 2 2/3 stops over and obviously it can do way more than that. So there's something I don't understand about the way sensor sensitivity works, and how it translates to linear code values in a 12-bit raw file. I'm talking about uniform exposure pushes rather than just affecting the shadows, but effectively it's the same area being influenced (shadows) in this example. You own an Alexa Classic, right? Do you have any opinion on how far you can push the ISO on the fp to get a noise floor level like the Alexa which is relatively high? To me, that's still acceptable in a good many situations. Although the Alexa noise also looks pretty organic.

Alternatively it could just be that I'm taking the bits buried right at the left side of the histogram, and expecting there to be enough detail when expanded out by +4 stops into the midrange. Perhaps 12 linear bit just isn't enough for this kind of image manipulation, hence banding and posterization! Maybe these numbers will reveal something: Assuming 4096 code values in a 12 bit integer linear raw image converted to linear floating point EXR where middle grey is pegged at 0.18: So if I underexpose by an extreme -4 stops, then that pushes middle grey down to .0112. With 4096 code values, middle grey at "correct" exposure would be at around 737. At -4 stops below, it would be at 46. So that's probably my answer - LOL. 46 code values for everything from middle grey and below, and 4050 for everything above middle grey! Assuming my maths is right. If so, -3 stops would be about 92 code values for middle grey in linear light encoding.

Just following up on that last post, I realise I was simultaneously saying the shadows hold up really well, then saying that they had banding and weird tinting issues. What I should have said is that generally the shadow detail seems very low noise and very good, but I did run into the artifact issue in the case of this hair crop in the post above. So hopefully that kind of thing is not native to the sensor.

Thanks for the detailed answer! I've been testing a lot more raw files in the last day that I found linked off youtube recently (see linked video). I mentioned that I would lean towards re-pegging the DR by at least +3 stops since the camera holds shadow detail very well. Assuming base ISO 100, it seems you can cleanly push +4 stops with EI 100-800 in daylight and then maybe +3 with ISO 3200 in a darker scene if managed. Not sure about ISO 1600, it apparently has a bit less range that 800 and 3200? My latest question is around bad posterizing in the shadows with a green tint. I'm attaching a crop of a +4 stop push in post on one of the test clips from the upload I linked to. I could be imagining this, but I think I've read here on eoshd before that people are seeing a green issue in shadows sometimes. I don't know if this is the issue here, though, or if it's to do with the fact that these files were passed through slimRAW with lossy DNG compression, where generally lossy DCT compression can unduly affect shadows because there's an assumption that it won't be seen. What I'm hoping is that with lossless or losslessly compressed DNGs, this would not be an issue. But additionally, I've noticed that very high contrast lenses (Sigma) and night time scenes make the blacks fall to zero. This isn't necessarily an issue, but I would rather have them just off the zero floor. Not just because any compression algorithm can cause problems all the way down there, but also that if shooting to push 2-4 stops in post then too much of the image is clipped on the zero end and it looks excessively contrasty. So it seems a good solution regardless of any compression issues is to shoot with something like a Tiffen Ultra Con 1 and/or a lower contrast vintage lens, assuming you can't just add a bit of bounce fill in a controlled lighting situation. Obviously this appeals for more of a run and gun setup. If the green/banding issue is because of the sensor, that obviously sucks!

No he hasn't, but I've worked on a lot of movies shot on Alexa where generally the footage is ingested to EXR as either linear gamma/Alexa Wide Gamut, or as ACES. I have about 10 TB of footage shot on Alexa with extreme dynamic range in the scene, so I got a good sense of how it performs. We are really pixel peeping the plates and tearing them apart, so we know what's what.

I know people take issue with a Sony "look", but I've found a very neutral image when imported to ACES just with the default S-log3/Sgamut3 input transform, displayed through the standard rrt/odt. I think a lot of people just don't do that one basic thing and then have issue with skin tones and weird tints. I had been planning to buy an FX6, but they're really hard to come by due to the supply chain issues. But a friend has one and it's been interesting to do a ton of tests with it. In my mind, when shooting raw it's close to an Alexa, except as I stated the lower DR.

What do you think about the idea of adding an after-market OLPF? Or is it best just to blur the image a tad 🙃. Not that there is presently an OLFP for this camera that I know of. I ask because I know it can experience moire in some cases, and Sigma decided not to add an OLPF to retain sharpness. But tbh 2k resolving power is enough for me if the pixels are good pixels.

Just a note IMO about linear Raw vs Log encoding. Let's keep in mind that log encoding was developed by Kodak for Cineon scanning of film negative, and is basically what was passed on to digital cameras. It's not an inferior image encoding method developed for hybrid 4k mirrorless cameras. The LogC encoding on the Alexa is probably the most prominent log curve for a digital camera to be very close to Cineon, and Red and Sony eventually capitulated to that. It's not inferior to linear raw, unless the individual log curve implementation by that camera manufacturer is left wanting. It's a more efficient encoding of highlights and shadows because it gives less code values to highlights, rather than half of a linear image for the brightest stop. It also allocates more code values to the shadows than linear raw. I would rather take 12 bit log ProRes RAW from the Sony cams/Ninja V than 12 bit linear DNGs, but obviously that's a much more expensive solution than the fp - like 5x the price. The reason that log footage is often inferior in the prosumer cams is because it's probably chroma subsampled, denoised and DCT compressed. ProRes 4444 log from the Alexa Classic is a thing of wonder though. If you apply the inverse transform of the log encoding curve, then your image is back in linear space and if you factor out the ways the image has been decimated, then it is very close to an original raw recording.

Yes the most desirable scenario is to use this in as minimalist a fashion as possible like a Leica that can capture Vistavision, exposing for the middle. I do get the feeling this is why Greig Fraser is shooting with an FX3 right now. But it's great that the fp can shoot internal raw to keep things even smaller, because I don't actually really rate the internal recording on the FX3. The ProRes RAW looks great though. Owning the Digital Bolex is like owning a vintage motorbike that you're constantly maintaining. Like, it's so weird to get a decent rig that's still minimalist and functional, so I'm always experimenting with it. I have an irrational tolerance of its inefficiencies. It does produce a fantastic image.

I think the trick is to get something neutral and high dynamic range as a starting point based on the input transform knowing how to transform the sensor data to some baseline level of real world reflectance and exposure, then transform to the "looks good" version on the right.

OK I re-read what you're saying and I think I misunderstood you in my initial reply. I'm not looking to measure from the noise floor up, although yes that's a good point that there will sort of only be 12.5 stops available from there. Although I think it's subjective and that the camera companies except Arri and maybe Red are quite obsessed with images with clean shadows and low noise. I'm looking to push the highlights as much as possible, then look at the noise and see if I can live with the way it looks as I push that 12.5 stop range up. Obviously this is difficult because on a linear scale on the sensor, the shadows are already compressed into a tiny range compared to the highlights due to the brightest stop taking up half the sensor's available range (due to the doubling of light). And I'm proposing pushing the shadows into an even smaller allocation at the low end. But these sensors are just so sensitive to low light, that I'm okay doing that.

If I understand you, you're saying what if you don't need to catch very bright highlights? I'm a bit of a fetishist for stops over middle grey, where I like to capturing the highlights of naked practical sources, specular highlights etc. But if they're not needed for the subject matter then it's overkill. I'm just interested to see what can be captured in what I would consider to be a great many cases, and I don't mind a somewhat raised noise floor if it looks natural. Like you said, when the highlights clip, they clip whereas it's more subjective with the low end and it also depends on the look you like.

Definitely a valid opinion. I don't think it's as bad as the EOS-M, but there are issues for sure. Although the original BMPCC had a decent colour workflow that was a bit more predictable. The main issue for me though, was that BMD did not actually publish their log curve/gamut, so it was still a bit of a black box. I find the Sony alpha cams (a7sIII/FX3/FX6) to be great and predictable, and the most like an Alexa in their workflow, although having about 2 stops less DR.

Are you referring to my idea of using the built-in exposure compensation? Of course, without having the camera I'm making assumptions about how it would work and if it would work at all. But I think broadly speaking, I was thinking that the sensor clips too early in the highlights, but that the shadows were very clean. So the whole range could probably be pushed easily by +3 stops. So say, shooting 800 for 100 which is what you say it does anyway. But that could be pushed further. Beyond the fact its internally pushing 100 to 800, I'm saying maybe shoot with -3 ND and compensate for that in the light meter reading and the exposure compensation on the camera, so that it displays the image at +3 so middle grey looks correct. Again, assumptions. What this could synergistically offer is that if you can toggle exposure comp off then you could see the highlights without them clipping, then toggle it back on on and see the image exposed correctly for middle grey. Given the quirks of the camera, I don't know if this workflow is possible fully internally, or if a Ninja V or an external monitor and a separate LUT would need to be applied. Theoretically, outputting an underexposed image via HDMI could then allow a LUT to do the exposure difference and roll off the highlights that would otherwise be visually clipped while shooting. Yes PR Raw should absolutely be linear raw. I noticed the way it looks on the Mac can be interpreted as log or some other thing, but that's to do with metadata tags that are non-destructive. The underlying data in unclipped linear raw. I'm not sure about ISO invariance though. I'm taking a stab and saying that it only looks different based on whether the camera is using one or the other of the base ISOs, and then the "intended" ISO is just metadata. The overall idea is that I'm proposing shifting the highlight range beyond what Sigma recommends with their ISO/DR chart via a fixed underexposure (not ETTR with a histogram on a per-shot basis), to the point where you're still just about comfortable with the raised noise floor - and even then you can consider pushing it past that and applying a bit of denoising with Neat Video.