NDSP Quad Cortex

[Whizzinby]

Not really. Their fully realized amp models (as opposed to captures) have as many params as, for instance, Helix or other “white box” modelers. Whether those QC amp models are composites of black box captures or not, I couldn’t say. I know that’s been the rumor for some time.

Yeah in terms of standard controls it’s all there, but you can’t say “swap tubes” or disable power amp simulation, because their modeling method doesn’t distinguish between various components of the amp. It’s just an end sound. “If this gets turned, the resulting sound is this.” Their method is kind of cool, but has its advantages and disadvantages.

 

[mbenigni]

Yeah in terms of standard controls it’s all there, but you can’t say “swap tubes” or disable power amp simulation, because their modeling method doesn’t distinguish between various components of the amp. It’s just an end sound. “If this gets turned, the resulting sound is this.” Their method is kind of cool, but has its advantages and disadvantages.

Agreed, but I don’t thing any of this, in itself, evidences black box modeling conclusively (even though it’s likely true, based on NDSP’s own comments.) If it did, we’d have to conclude Helix and a half dozen other platforms were also black box, which we know isn’t true.

 

[MirrorProfiles]

Yeah in terms of standard controls it’s all there, but you can’t say “swap tubes” or disable power amp simulation, because their modeling method doesn’t distinguish between various components of the amp. It’s just an end sound. “If this gets turned, the resulting sound is this.” Their method is kind of cool, but has its advantages and disadvantages.

Depends if they are black boxing the entire amp end to end or splitting into sections. Pretty sure the plugins have separate models for preamps and power amps. I know some companies will mix and match black box and white box, only using black box for highly non-linear parts of the circuit.

 

[Whizzinby]

Agreed, but I don’t thing any of this, in itself, evidences black box modeling conclusively (even though it’s likely true, based on NDSP’s own comments.) If it did, we’d have to conclude Helix and a half dozen other platforms were also black box, which we know isn’t true.

Yeah I don’t know about white/black/taupe-box modeling. I just know Doug said they don’t have that sort of granular customization because their modeling approach doesn’t account for it. Which, makes sense to my dumb brain if it’s largely a capture/mlm approach.

I desperately need another cup of coffee.

 

[JiveTurkey]

I think the whole "we don't do it like that" approach allows them to play selectively dumb as they feel and then avoid getting into any sort of accuracy discussions with their product. Which, oddly enough; never seem to come up?

 

[MirrorProfiles]

any sort of accuracy discussions with their product. Which, oddly enough; never seem to come up?

Their plugins have always checked out as extremely accurate for me. They even posted A/B examples of their Nameless plugin vs the real amp which way too few companies are prepared to do:

 

[Whizzinby]

I think the whole "we don't do it like that" approach allows them to play selectively dumb as they feel and then avoid getting into any sort of accuracy discussions with their product. Which, oddly enough; never seem to come up?

Disclosure: I’m a dumbass… but it seems “simple” that they are “basically” feeding an input, having Tina work the knobs (NSFW) and they measure the resulting output. They do this with enough control variations to feel confident to interpolate the rest of them.

Which I don’t think is a knock. Their models sound great. (But explains why there are some limitations if you want to go full fantasy-amp-designer)

To my knowledge you can’t disable power amp sim on QC models. (Or couldn’t when I had it) When asked Doug said they couldn’t (at that time) because they don’t account for specific sections of the amp. But my memory grows dim.

 

[JiveTurkey]

Disclosure: I’m a dumbass… but it seems “simple” that they are “basically” feeding an input, having Tina work the knobs (NSFW) and they measure the resulting output. They do this with enough control variations to feel confident to interpolate the rest of them.

Which I don’t think is a knock. Their models sound great. (But explains why there are some limitations if you want to go full fantasy-amp-designer)

To my knowledge you can’t disable power amp sim on QC models. (Or couldn’t when I had it) When asked Doug said they couldn’t (at that time) because they don’t account for specific sections of the amp. But my memory grows dim.

 

[laxu]

Disclosure: I’m a dumbass… but it seems “simple” that they are “basically” feeding an input, having Tina work the knobs (NSFW) and they measure the resulting output. They do this with enough control variations to feel confident to interpolate the rest of them.

Which I don’t think is a knock. Their models sound great. (But explains why there are some limitations if you want to go full fantasy-amp-designer)

To my knowledge you can’t disable power amp sim on QC models. (Or couldn’t when I had it) When asked Doug said they couldn’t (at that time) because they don’t account for specific sections of the amp. But my memory grows dim.

It's anything but simple. There was a research paper (can't find the link atm) that discusses their approach to this and it's definitely innovative stuff to get around the fact that it would take a shitload of time to analyze every possible permutation of even a single channel amp's knob positions. Their method seems like a good shortcut to avoid that.

It's an approach that allows them to shortcut things like carefully measuring the real amp at various points, building simulations of gain stages and stuff like that. If you look at e.g Fractal, Cliff is constantly correcting, improving and evolving their modeling system as he gains new insight. A machine learning model doesn't know about any of that stuff. Instead the hard work is figuring out the math to make those models do the things they need.

In the end the development effort probably ends up being similar, just from totally different perspectives. There's a possibility both companies will eventually end up at their respective models that are so close to the real amps that any further development becomes unnecessary. Then it comes just a question of how all that scales to adding new amps and whatnot.

So far for Neural DSP it seems it hasn't. There's still no updates to existing models and very few new models released since the announcement of TINA.

 

[Whizzinby]

It's anything but simple. There was a research paper (can't find the link atm) that discusses their approach to this and it's definitely innovative stuff to get around the fact that it would take a shitload of time to analyze every possible permutation of even a single channel amp's knob positions. Their method seems like a good shortcut to avoid that.

It's an approach that allows them to shortcut things like carefully measuring the real amp at various points, building simulations of gain stages and stuff like that. If you look at e.g Fractal, Cliff is constantly correcting, improving and evolving their modeling system as he gains new insight. A machine learning model doesn't know about any of that stuff. Instead the hard work is figuring out the math to make those models do the things they need.

In the end the development effort probably ends up being similar, just from totally different perspectives. There's a possibility both companies will eventually end up at their respective models that are so close to the real amps that any further development becomes unnecessary. Then it comes just a question of how all that scales to adding new amps and whatnot.

So far for Neural DSP it seems it hasn't. There's still no updates to existing models and very few new models released since the announcement of TINA.

I thought air quoting “simple” would imply I was talking conceptually, not that it was actually “simple” to execute.

Though I think we come to the same conclusion on it.