holly crap!, the first video is amazing! except for the blue banding in the shadows it was outstanding O.O!

Couldn’t you always go to ISO 12,800 and beyond in video by using shutter priority and Auto ISO even with an unhacked camera?

He has a link to a 1080p version with and without neat and it is still nice, but for me it is the result that count, do I find it good, yes for 10 000 iso it is damn good. The other one is very nice because even if it is only ........... 3200 iso it is still very very nice, more so that it was deliberately underexposed so this tend to show more of the noise. Sometime if it is very dark the crushed black would hide the noise, but even with the vimeo compression we can see detail in the shadows. I think someone more adverse to noise could use a little NR software with very good end result.

Having just applied a slightly lower bitrate version of Driftwood's GOP1 "Spanmybitchup" patch to my camera and using a wide lens (Nikon 1.4, stopped down usually to 2.4ish) I'm gobsmacked (or should that be GOPsmacked?) by the reduction in noise at high gain. I understand that the noise I've been used to, has been the result of B and P frames, and with GOP1 it's just amazingly clean. I really can't believe how good it is, even though GOP1 isn't as efficient a use of data as you get with P and B frames. I'm assuming it's the result of low GOP anyway, and that's got me completely sold on this low-GOP stuff from now on. I shot some bits and pieces around the house today and it's wonderful. I want to do a candlelit harp piece at some point, but I have a major gig to prepare for tomorrow and I now have to do that rather than messing around with bits of footage.

It was that first video (the Halloween one) that got me doing the patch, and I can quite believe the results he got. Looking around the dimmer areas of my house, with a wide lens this thing can see colours you can't see with the naked eye in low light.

Something I have seen for myself is how easy it is to flood the lens with light at these very high isos and get lots of blue flare (in the case of my Nikon). I think it's something to do with the light sources being more at eye level than they normally are (car headlights, candles etc)? Don't know, but the GH2 does very gracefully clip when it happens so any stray light reflecting in the lens certainly gets into the pictures. You just have to watch where you point the lens.

The most bizarre thing is that the footage almost doesn't look real - the graininess you normally get is such a good clue that you are filming in low light, and suddenly everything's different.

@alcomposer Thank you! I have to confess I didn't believe all that stuff about low GOP - but my wife (ex BBC camera person, ex 16mm camera operator and who has seen a ton of footage from the GH2) saw some lowlight GOP1 stuff I'd shot today, and unprompted said it looked "like 35mm" which is the first time she's made such a comment - and the first time I've used anything other than standard GOP.

Can't wait to get my Russian adaptor to try out my f1.4 zoom lens which I cannibalized from an early AKAI video camera. I think it'll be one of my "go-to" lenses if it works as well as I think it does and it'll be the perfect pairing with these new patches.

@danyyyel I think it won't stop lighting being important, but what's wonderful is that you can do great stuff in available light. I love the 14-140 for its all-round flexibility but I think the wider aperture lenses really come into their own for low-light work. I think you still need to keep the ISO as low as possible, to keep the noise down. I've only been using the high ISO patch for a day, and to be honest I've got a big shoot tomorrow and wasn't going to even consider using a new patch, but this is just perfect for what I want, because it's suddenly possible to do what I need in amongst a very busy, crowded area, where additional lighting would be very difficult to arrange.

@dannyel Exciting stuff. This simplifies the lighting requirement in low lighting. Huge time saving.

Screw efficient encoding. Long Live GOP1. LLG1!!!!!!!!!!! Sandisk 16GB Class10 cards are cheap.

CV 25mm 0.95 is wonderful, too. When I get the lens, I'll be ready to capture my friends picking nose at night. I will try to follow through booger flying in the air, too.

There is no inherent reason why h.264 P and B frames should introduce any more compression artifacts than P frames. It's simply the design of the GH2's encoder and our understanding of it that makes it easier to get less quantization with all I-frames than with P and B frames. A constant quality encoder will apply the exact same quantization to every single block of every single frame, regardless of frame type.

@balazer My understanding is that you're right for normal moving pictures, but noise (being random) is impossible to predict, hence looks uglier with higher GOPs. Or is that what you said anyway?

Don't know if I'm right but it makes sense in terms of the difference I see with all I-frames. It's sort of subtle but at the same time immediately noticeable when you have been used to the long GOP footage and for a given bitrate, looks better to me even if strictly speaking it's less efficient.

the first video is amazing, but largely due to the fact that the guy used the voigtländer F0.95 lens :-D that thing is amazing!

@Mark_the_Harp, that's not what I'm saying. What I said is true regardless of the amount of noise.

A constant quality encoder applies the same level of quantization to every single block of every single frame, regardless of frame type. That means that you would not be able to tell the difference between an I-frame, P or B frame in the decoded picture, and you'd have no visible indication of the GOP length. Compression efficiency goes up when there is more similarity between frames and when the GOP is longer. If a video stream is pure noise, there is no similarity between frames, so using longer GOPs would give you no more efficiency: for some encoder output quality level, a GOP1 stream would have the same bit rate as GOP10. The encoded P and B frames would be the same size as the I-frames.

But the GH2 encoder isn't constant quality. It has all kinds of biases built into it, because it was designed to be variable within certain bit rate ranges, with certain GOP lengths and patterns. I haven't followed the development of the hacks in detail, but I'm guessing that it turned out to be impossible to get the GH2 to boost the sizes of its encoded P and B frames. To preserve the most quality under very noisy conditions, the P and B frames should be nearly as large as the I-frames, and the bit rate would go up accordingly.

Using all I-frames eliminates one component of an encoder's variable performance: now the encoder is doing nothing to try to find similarities between frames, so the amount of similarity no longer has any bearing on the encoder's compression efficiency. But even having eliminated one component of the encoder's variable performance, some frames are more compressible than other frames, so using all I-frames doesn't make the encoder constant quality. It just gets us closer, in some sense. And since the GH2 lets us write very large I-frames, we get high quality regardless of the amount of similarity between frames. That helps a lot with high ISOs.

Notice I didn't use the word "motion" anywhere in there. The encoder has a very limited notion of motion. Basically, when I said that the encoder looks for similarity between frames, similarity includes the case when an object visible in one frame is visible in another frame, but in a different position within the frame. The encoder is able to search for where the object moved to, and compensate for its motion when it builds the base frame that a P or B frame is encoded relative to. So when you hear people talk about encoder settings that are optimized for high motion, it's really not accurate. Motion is just one thing that can make two frames less similar even after compensation, and then only in specific cases, e.g. when new objects are moving into frame, or when an object has moved farther than the encoder is able to search for.

@danyyyel - "Just need for a hack to get the best Dynamic range from the sensor for the gh2 to be perfect. "

The downside to shooting at ISO 12800 is a dramatic reduction in dynamic range. This, along with its background noise, is an inherent aspect of the physical properties of the image sensor:

http://www.sensorgen.info/PanasonicDMC_GH2.html

At ISO 12800, the GH2 has only 5 stops of dynamic range, but as we see in the video samples in this thread, that's enough for low-light shooting. If you look at the Sensorgen link's Read Noise data, you may spot the GH2's saving grace: the sensor has only half as much random noise at ISO 12800 than at ISO 160! At the dark end of the dynamic range, this is equivalent to opening the lens aperture a full stop wider, making shadow detail significantly cleaner at ISO 12800.

With the limited dynamic range, however, you should be prepared to blow out the highlights whenever a direct source of illumination is in the frame. Black & white Film Modes can handle this more gracefully as there is no chance of color shift in the blown highlights.

If you think those conditions are bad for low light, here's my night: shooting an interview with a homeless man, 10 minute hike away from my car, down the river near a bridge, with so little light from the nearest street illumination in some shots that I needed to shoot ISO 12,800 at 1/8 to get enough light.

That is until the shoot was interrupted by the sound of 5 bullets being fired in close proximity and we had to go silent until we could be sure the shooter had left. Oh, and then trying to get stills to from atop a bridge to see whether a particular shape was a log or a dead body...

I think I'll go back to worrying about poorly controlled lighting in a nice room again, thank you very much. :)