Turbo node vs Lora node, beta vs normal (Turbo 500)
These are examples using the Turbo Lora node vs normal Lora node, as well as Ksampler vs DualClock sampler vs Turbo sampler. (interesting to note, the DualClock seems to not care if it's using beta or simple, it gets processed the same way)
The model in use is the unpruned INT8 FL2VA model. (though I have great success using the turbo lora with the pruned one, so I don't understand why it's advertised as not working, I found the turbo results more favourable on the pruned one when using the same settings)
Turbo Node | Turbo Sampler | Normal (first) vs Beta (second) scheduler
Lora Node | Turbo Sampler | Normal (first) vs Beta (second) scheduler
Turbo Node | DualClock Sampler | Normal (first) vs Beta (second) scheduler
Lora Node | Dualclock Sampler | Normal (first) vs Beta (second) scheduler
Turbo Node | KSampler | Normal (first) vs Beta (second) scheduler
Lora Node | KSampler | Normal (first) vs Beta (second) scheduler
Although the beta sampler came out very overblown on the ksampler here it's good on the turbo sampler, I seem to consistently get better results with Beta over Normal which tends to smear and overfit more. The turbo node seems very bad as of now.
Yes, there is a stupid precision bug in turbo lora loader, fixing, thanks for probing and pointing out!
These are examples using the Turbo Lora node vs normal Lora node, as well as Ksampler vs DualClock sampler vs Turbo sampler. (interesting to note, the DualClock seems to not care if it's using beta or simple, it gets processed the same way)
The model in use is the unpruned INT8 FL2VA model. (though I have great success using the turbo lora with the pruned one, so I don't understand why it's advertised as not working, I found the turbo results more favourable on the pruned one when using the same settings)
Although the beta sampler came out very overblown on the ksampler here it's good on the turbo sampler, I seem to consistently get better results with Beta over Normal which tends to smear and overfit more. The turbo node seems very bad as of now.
Hey, thank you for the wonderful comparison.
Its not that it doesn't work with pruned version, it just gives errors on terminal because it can't find some keys which are intentionally left or removed from pruned versions.
I found that results are good with pruned as well.
Which combo you suggest? for the best quality. I appreciate your help <3
These are examples using the Turbo Lora node vs normal Lora node, as well as Ksampler vs DualClock sampler vs Turbo sampler. (interesting to note, the DualClock seems to not care if it's using beta or simple, it gets processed the same way)
The model in use is the unpruned INT8 FL2VA model. (though I have great success using the turbo lora with the pruned one, so I don't understand why it's advertised as not working, I found the turbo results more favourable on the pruned one when using the same settings)Although the beta sampler came out very overblown on the ksampler here it's good on the turbo sampler, I seem to consistently get better results with Beta over Normal which tends to smear and overfit more. The turbo node seems very bad as of now.
Hey, thank you for the wonderful comparison.
Its not that it doesn't work with pruned version, it just gives errors on terminal because it can't find some keys which are intentionally left or removed from pruned versions.
I found that results are good with pruned as well.Which combo you suggest? for the best quality. I appreciate your help <3
No turbo node, Euler+Beta, Turbo sampler optional, better on this example but I have experienced otherwise as well, but it's not a downgrade.
These are examples using the Turbo Lora node vs normal Lora node, as well as Ksampler vs DualClock sampler vs Turbo sampler. (interesting to note, the DualClock seems to not care if it's using beta or simple, it gets processed the same way)
The model in use is the unpruned INT8 FL2VA model. (though I have great success using the turbo lora with the pruned one, so I don't understand why it's advertised as not working, I found the turbo results more favourable on the pruned one when using the same settings)Although the beta sampler came out very overblown on the ksampler here it's good on the turbo sampler, I seem to consistently get better results with Beta over Normal which tends to smear and overfit more. The turbo node seems very bad as of now.
Hey, thank you for the wonderful comparison.
Its not that it doesn't work with pruned version, it just gives errors on terminal because it can't find some keys which are intentionally left or removed from pruned versions.
I found that results are good with pruned as well.Which combo you suggest? for the best quality. I appreciate your help <3
No turbo node, Euler+Beta, Turbo sampler optional, better on this example but I have experienced otherwise as well, but it's not a downgrade.
sorry for the noob question, but when you say "Beta", what exactly do you mean? is it a node or a sampler type? and could you just show it to me?
EDIT: sorry never mind it's the schedular.
sorry for the noob question, but when you say "Beta", what exactly do you mean? is it a node or a sampler type? and could you just show it to me?
No question is noob, you're just new to it
These are examples using the Turbo Lora node vs normal Lora node, as well as Ksampler vs DualClock sampler vs Turbo sampler. (interesting to note, the DualClock seems to not care if it's using beta or simple, it gets processed the same way)
The model in use is the unpruned INT8 FL2VA model. (though I have great success using the turbo lora with the pruned one, so I don't understand why it's advertised as not working, I found the turbo results more favourable on the pruned one when using the same settings)Although the beta sampler came out very overblown on the ksampler here it's good on the turbo sampler, I seem to consistently get better results with Beta over Normal which tends to smear and overfit more. The turbo node seems very bad as of now.
Hey, thank you for the wonderful comparison.
Its not that it doesn't work with pruned version, it just gives errors on terminal because it can't find some keys which are intentionally left or removed from pruned versions.
I found that results are good with pruned as well.Which combo you suggest? for the best quality. I appreciate your help <3
No turbo node, Euler+Beta, Turbo sampler optional, better on this example but I have experienced otherwise as well, but it's not a downgrade.
sorry for the noob question, but when you say "Beta", what exactly do you mean? is it a node or a sampler type? and could you just show it to me?
Either of these two (for H3 using the turbo lora he has specifically made the "MiniMax-H3 Turbo Sampler (4-step)" node but I am otherwise be refering to the "KSamplerSelect" node.
sorry for the noob question, but when you say "Beta", what exactly do you mean? is it a node or a sampler type? and could you just show it to me?
No question is noob, you're just new to it
many thanks and respects.