2dor
Shredder
- Messages
- 1,192
NAM only had an issue with full-rig captures for about a few months after it blew up. Steve came up with a fix shortly after (the fit_cab flag which is now on by default in the trainer and even on Tone3000).
NAM: Neural Amp Modeler
- Thread starter James Freeman
- Start date
MirrorProfiles
Rock Star
- Messages
- 4,092
Ha yeah I got so used to the decimals being in the same kind of ballpark. I usually accept around 0.1 for high gain tones, if its higher its usually a bit different sounding, but for a lot of gain it can be hard to get much lower.
Nice! Testing it now, I'll see how they turn out.
Yeah, B just sounds more satisfying and better to me although it's quite a small difference. Been a while since I fired up ToneX but I might as well include an advanced model to see how close it gets to these.
MirrorProfiles
Rock Star
- Messages
- 4,092
I have one model training with a "gated" parameter set to true, but Im trying to do another now and when I change any of them I get this message:
Is that because of what I have training at the moment? Seems strange that I could do it a few minutes ago but not now.
Is that because of what I have training at the moment? Seems strange that I could do it a few minutes ago but not now.
2dor
Shredder
- Messages
- 1,192
There's a cap on much tweaking Tone3000 allows so that some folks don't abuse the system lol.
Maybe try flipping it to "true" in the layer with only the 2 small dilations:
This layer should deal with the high-end anyway.
MirrorProfiles
Rock Star
- Messages
- 4,092
Hmm, yeah no dice. Would appreciate it if anyone else has something to test and can check. I literally did it a few seconds ago, but I guess I'll wait for my existing models to train in case that's a factor.
2dor
Shredder
- Messages
- 1,192
I wouldn't worry too much since that ESR is already pretty darn good anyway. The xSTD models you're training will probably dip slightly below it.
Dom Mcsweeney
Roadie
- Messages
- 277
0.1 is a type-o?
If getting '0.01', I'd be happy enough with that.
If getting '0.01', I'd be happy enough with that.
MirrorProfiles
Rock Star
- Messages
- 4,092
Yeah, it was 0.01. And actually, removing the amp switcher did help lower the ESR a little more. to 0.0087.
I'm usually OK with these values as they sound good, but unfortunately here it still sounds off (newest clip is C) - the low end isn't quite right.
Training 2 xSTD models, one with the Ampete, one without. The one with the Ampete had one of the gated=true but I haven't been able to apply that since. Also doing a ToneX model just to see. Hoping I can get the low end a bit more satisfying here
Last edited:
2dor
Shredder
- Messages
- 1,192
They're pretty close tho. Takes a bit of fiddling around tho & it doesn't look like Tone3000 will allow too much wiggle room (just tried a few combinations now in their UI for a few custom architectures).
MirrorProfiles
Rock Star
- Messages
- 4,092
A couple more with xSTD:
Definitely not unhappy with how they sound but this tone seems to show more differences to my ear than I'm maybe used to. There's a kind of characteristic heft on the palm mutes that's different (and the NAM models are a teeny bit more abrasive I think?).
Got the ESR down to 0.0071 on one model. Suppose I'd need to train on colab or locally to tweak things more. LOL at how much I've done while waiting for ToneX to do a single model. Makes it so much more appealing to use NAM.
Definitely not unhappy with how they sound but this tone seems to show more differences to my ear than I'm maybe used to. There's a kind of characteristic heft on the palm mutes that's different (and the NAM models are a teeny bit more abrasive I think?).
Got the ESR down to 0.0071 on one model. Suppose I'd need to train on colab or locally to tweak things more. LOL at how much I've done while waiting for ToneX to do a single model. Makes it so much more appealing to use NAM.
2dor
Shredder
- Messages
- 1,192
Training locally's the way to go for sure if you want to experiment. You can try some big a$$ architectures and incrementally trim it down until you strike that balance between resource usage and fidelity.
If you do train locally, I'm curious to head how many seconds an epoch takes (on the STANDARD architecture) on that Mac of yours
A few pages back, @DLC86 posted recommended learning rate values for xSTD. I’ve still achieved low ESRs if I forget to change them.
https://thegearforum.com/threads/tone3000-previously-tonehunt-and-tonezone3000.6919/post-320447
From what I’ve observed, xSTD brings the ESR down to the 0.001 to 0.002 range if the Standard capture is just under 0.01.
https://thegearforum.com/threads/tone3000-previously-tonehunt-and-tonezone3000.6919/post-320447
From what I’ve observed, xSTD brings the ESR down to the 0.001 to 0.002 range if the Standard capture is just under 0.01.
MirrorProfiles
Rock Star
- Messages
- 4,092
How did I miss those learning rate values? Nice one, trying another model now with the values posted there. Will add tonex to the playlist above in the next few minutes
2dor
Shredder
- Messages
- 1,192
ToneX is a time hogg.
laxu
Rock Star
- Messages
- 6,871
To my ears the "No Ampete" clips sound closer to the real amp. Agree that some of them sound a bit more abbrasive.
But I would have no complaints using any of these.
MirrorProfiles
Rock Star
- Messages
- 4,092
Last couple of clips added.
I used the custom learning rate values provided above - these got the ESR way down to 0.0037 for the example without the Ampete, and 0.008 for the version with. I think clip H gets the closest to the source tone. I can still hear some differences but it's the closest so far to getting the distinctive characteristic of the amplifier correct. The low end is still not quite there, and the real amp feels a hair smoother and less abrasive.
Thanks to all those who helped with the tweaks - starting at 0.01 and ending at 0.0037 is good going.
IMO capturing tech should be as close as it possibly can to the source, so I applaud those pushing boundaries on making that accessible to others. I'll definitely be using these tweaks going forward, the differences might be small but they're often in very important parts of the tone. Even though all those clips sound passable and good enough, many of them feel like a slightly different amp to me (if Im expecting that huge low end to punch through). I think this all just makes me want to use real amps even more on projects, all kinds of modelling are awesome for what they are but there's some magic when you turn an amp on and dial a tone in. You just get THERE faster and you have something distinct with personality at the end of it.
I used the custom learning rate values provided above - these got the ESR way down to 0.0037 for the example without the Ampete, and 0.008 for the version with. I think clip H gets the closest to the source tone. I can still hear some differences but it's the closest so far to getting the distinctive characteristic of the amplifier correct. The low end is still not quite there, and the real amp feels a hair smoother and less abrasive.
Thanks to all those who helped with the tweaks - starting at 0.01 and ending at 0.0037 is good going.
IMO capturing tech should be as close as it possibly can to the source, so I applaud those pushing boundaries on making that accessible to others. I'll definitely be using these tweaks going forward, the differences might be small but they're often in very important parts of the tone. Even though all those clips sound passable and good enough, many of them feel like a slightly different amp to me (if Im expecting that huge low end to punch through). I think this all just makes me want to use real amps even more on projects, all kinds of modelling are awesome for what they are but there's some magic when you turn an amp on and dial a tone in. You just get THERE faster and you have something distinct with personality at the end of it.
![[Nathan]](/data/avatars/m/0/613.jpg?1726885163){kind=avatar}
[Nathan]
Roadie
- Messages
- 668
DLC86
Roadie
- Messages
- 820
If you want you can also try using the TTSv10 input file together with the xSTD architecture, usually, besides reducing aliasing, doing that also reduces further the ESR (and ime both things can make some audible difference on this kind of tones).
In case you don't know where to find those alternative inputs, here they are: https://drive.google.com/drive/folders/13Df-GlKYTkpkx83Rwp-6QsdiQcJFGDxG?usp=drive_link
PS: mind you, training with alternative inputs on tone3000 throws off the output level of the model (unless they fixed it recently)
Nembrini
Newbie
- Messages
- 13
Where can I find NAM captures made with this? Thanks
This is still an issue. I tried training with the Wet-Dry method and the volume jump still happens. It is alleviated when the Normalization slider is checked in the Output Mode, but Raw and Calibrated modes can still blow your ears off. I gave the TONE3000 team a heads up about it.
