Gazingstars123

/

Anima-2.9B

Like 375

1. If i understand your question correctly, the fix is on the training side, not the model side. Character LoRAs need style-diverse datasets (as many different artists as they can find for that character) and explicit artist tags in the captions. Also: stack at ~0.6–0.8 weight rather than 1.0.
2. It's already supported natively in Forge Neo
3. I won't say no, but it's a not an easy job, and the capability to actually train them depend on two things: time and money. So it can either takes very long or very fast depending on people, and if they are willing to support the project. And it's something that has to be decided today, not later.
4. Upscaling works a bit better, but images do start getting noisy at around 2.5MP or above
5. Sorry man, I don't understand this question

1. If i understand your question correctly, the fix is on the training side, not the model side. Character LoRAs need style-diverse datasets (as many different artists as they can find for that character) and explicit artist tags in the captions. Also: stack at ~0.6–0.8 weight rather than 1.0.
2. It's already supported natively in Forge Neo
3. I won't say no, but it's a not an easy job, and the capability to actually train them depend on two things: time and money. So it can either takes very long or very fast depending on people, and if they are willing to support the project. And it's something that has to be decided today, not later.
4. Upscaling works a bit better, but images do start getting noisy at around 2.5MP or above
5. Sorry man, I don't understand this question

Hi again, regarding my stupid questions:
Questions 1 and 5: Thanks for the advice. I was just working with models that have a consistent style, and I was looking for ways to ensure that the Lora style would be applied correctly and fully to such a model, rather than distorted.

5. Sorry man, I don't understand this question

Will Lora be backwards compatible with the regular Base version if I train on 2.9?

No, old LoRA from 2B can be mapped, but 2.9B LoRA can't.

No, old LoRA from 2B can be mapped, but 2.9B LoRA can't.

Got it, thanks.

No, old LoRA from 2B can be mapped, but 2.9B LoRA can't.

I have another stupid question: is the limitation on using Anima base and Anima 2.9B related to the technical structure of the architecture, or is some layer data lost during fine-tuning/merging, which is why training on custom checkpoints is unavailable?

No, old LoRA from 2B can be mapped, but 2.9B LoRA can't.

I received a response from the developers of Anima, I came across that your training trainer is the same as kohya_ss I didn't train lore on custom checkpoints:
Yeah that's a key prefix error. I downloaded the model you're trying to use and all of the prefixes use model.diffusion_model, including the one in the error which is "model.diffusion_model.x_embedder.proj.1.weight". The trainer is likely expecting just "net.x_embedder.proj.1.weight" as the prefix for all the keys (it looks like I had it the other way around with how Comfy saves, my bad). You can definitely just ask any AI to write a short python script to rename the key prefixes in the checkpoint to "net" instead which will fix this.

No, old LoRA from 2B can be mapped, but 2.9B LoRA can't.

Off-topic
A question arose, isn't it possible to launch training in your trainer in a queue, rather than manually each one?
For example, label the data, provide all the parameters and simply run a batch of sequential training?

No, old LoRA from 2B can be mapped, but 2.9B LoRA can't.

Off-topic
A question arose, isn't it possible to launch training in your trainer in a queue, rather than manually each one?
For example, label the data, provide all the parameters and simply run a batch of sequential training?

It's not supported yet.

No, old LoRA from 2B can be mapped, but 2.9B LoRA can't.

I received a response from the developers of Anima, I came across that your training trainer is the same as kohya_ss I didn't train lore on custom checkpoints:
Yeah that's a key prefix error. I downloaded the model you're trying to use and all of the prefixes use model.diffusion_model, including the one in the error which is "model.diffusion_model.x_embedder.proj.1.weight". The trainer is likely expecting just "net.x_embedder.proj.1.weight" as the prefix for all the keys (it looks like I had it the other way around with how Comfy saves, my bad). You can definitely just ask any AI to write a short python script to rename the key prefixes in the checkpoint to "net" instead which will fix this.

also, my trainer can load any checkpoint prefix now.

No, old LoRA from 2B can be mapped, but 2.9B LoRA can't.

I received a response from the developers of Anima, I came across that your training trainer is the same as kohya_ss I didn't train lore on custom checkpoints:
Yeah that's a key prefix error. I downloaded the model you're trying to use and all of the prefixes use model.diffusion_model, including the one in the error which is "model.diffusion_model.x_embedder.proj.1.weight". The trainer is likely expecting just "net.x_embedder.proj.1.weight" as the prefix for all the keys (it looks like I had it the other way around with how Comfy saves, my bad). You can definitely just ask any AI to write a short python script to rename the key prefixes in the checkpoint to "net" instead which will fix this.

also, my trainer can load any checkpoint prefix now.

Sorry if I'm bothering you with my stupid questions, I just want to learn so much from someone who knows how to do it...
And I have another question: if I, say, install additional graphics cards and enable the multi-card mode, will the training be accelerated or does the mode simply allow multiple cards to be trained simultaneously?

No, old LoRA from 2B can be mapped, but 2.9B LoRA can't.

I received a response from the developers of Anima, I came across that your training trainer is the same as kohya_ss I didn't train lore on custom checkpoints:
Yeah that's a key prefix error. I downloaded the model you're trying to use and all of the prefixes use model.diffusion_model, including the one in the error which is "model.diffusion_model.x_embedder.proj.1.weight". The trainer is likely expecting just "net.x_embedder.proj.1.weight" as the prefix for all the keys (it looks like I had it the other way around with how Comfy saves, my bad). You can definitely just ask any AI to write a short python script to rename the key prefixes in the checkpoint to "net" instead which will fix this.

also, my trainer can load any checkpoint prefix now.

Sorry if I'm bothering you with my stupid questions, I just want to learn so much from someone who knows how to do it...
And I have another question: if I, say, install additional graphics cards and enable the multi-card mode, will the training be accelerated or does the mode simply allow multiple cards to be trained simultaneously?

There are many different multi-gpu modes, but generally, with two gpus, you can train twice-ish as fast per epoch, i.e twice amount of images can be seen by the model per training step.

No, old LoRA from 2B can be mapped, but 2.9B LoRA can't.

I received a response from the developers of Anima, I came across that your training trainer is the same as kohya_ss I didn't train lore on custom checkpoints:
Yeah that's a key prefix error. I downloaded the model you're trying to use and all of the prefixes use model.diffusion_model, including the one in the error which is "model.diffusion_model.x_embedder.proj.1.weight". The trainer is likely expecting just "net.x_embedder.proj.1.weight" as the prefix for all the keys (it looks like I had it the other way around with how Comfy saves, my bad). You can definitely just ask any AI to write a short python script to rename the key prefixes in the checkpoint to "net" instead which will fix this.

also, my trainer can load any checkpoint prefix now.

Sorry if I'm bothering you with my stupid questions, I just want to learn so much from someone who knows how to do it...
And I have another question: if I, say, install additional graphics cards and enable the multi-card mode, will the training be accelerated or does the mode simply allow multiple cards to be trained simultaneously?

There are many different multi-gpu modes, but generally, with two gpus, you can train twice-ish as fast per epoch, i.e twice amount of images can be seen by the model per training step.

Sorry if I'm still going off-topic, but I wanted to clarify whether it would be better if I expand my video memory, for example, with a Tesla V100 32GB (2018 server accelerator), or would it be better to buy a modern video card like a 5060 ti 16GB for the same price for neural networks?

I'm currently torn between building a small server out of a Tesla V100 32GB and building a second server out of a graphics card weaker than my 5070 Ti, purely for parallel processing. I've been thinking about either adding an additional graphics card (V100 32GB or 5060 Ti 16GB) to my PC, but I don't know what to do yet, and I have no one to ask... 😞

No, old LoRA from 2B can be mapped, but 2.9B LoRA can't.

I received a response from the developers of Anima, I came across that your training trainer is the same as kohya_ss I didn't train lore on custom checkpoints:
Yeah that's a key prefix error. I downloaded the model you're trying to use and all of the prefixes use model.diffusion_model, including the one in the error which is "model.diffusion_model.x_embedder.proj.1.weight". The trainer is likely expecting just "net.x_embedder.proj.1.weight" as the prefix for all the keys (it looks like I had it the other way around with how Comfy saves, my bad). You can definitely just ask any AI to write a short python script to rename the key prefixes in the checkpoint to "net" instead which will fix this.

also, my trainer can load any checkpoint prefix now.

Sorry if I'm bothering you with my stupid questions, I just want to learn so much from someone who knows how to do it...
And I have another question: if I, say, install additional graphics cards and enable the multi-card mode, will the training be accelerated or does the mode simply allow multiple cards to be trained simultaneously?

There are many different multi-gpu modes, but generally, with two gpus, you can train twice-ish as fast per epoch, i.e twice amount of images can be seen by the model per training step.

Sorry if I'm still going off-topic, but I wanted to clarify whether it would be better if I expand my video memory, for example, with a Tesla V100 32GB (2018 server accelerator), or would it be better to buy a modern video card like a 5060 ti 16GB for the same price for neural networks?

I'm currently torn between building a small server out of a Tesla V100 32GB and building a second server out of a graphics card weaker than my 5070 Ti, purely for parallel processing. I've been thinking about either adding an additional graphics card (V100 32GB or 5060 Ti 16GB) to my PC, but I don't know what to do yet, and I have no one to ask... 😞

you can join our discord, link in the front page.