Forum Replies Created
-
Posts
-
Thanks for the feedback! It’s great that you find good value in the membership.
I aim to spread the knowledge while getting somewhat compensated (and justified) for my time. I’m glad to find a solution that works both ways.
I agree it’s not entirely straightforward. I usually deal with memory issues as they arise. This can be done using a more memory-efficient version of the model (fp8, fp4), using a smaller image size, and unloading models from memory, etc.
The HiRes fix function often results in memory issues. I am not sure what’s wrong with the implementation. Maybe a flux workflow with increasing batch size or image size is a better way to test.
Welcome, Jessica! Thank you for sharing your beautiful work. It is a great remix of “What is life”!
I’m sorry I didn’t notice your message awaiting my moderation.
This is a good suggestion. Instructions for A1111 and Forge should be interchangeable. I will probably add new ComfyUI ones.
Thanks for the suggestions.
When I wrote the course, I tried to keep it agnostic to how to use A1111 (local/colab/online service). So I didn’t write much about instructions specific to the Colab notebook. I will add them to lessons that require additional extensions.
Several factors determine the relationship between model size and the required VRAM.
- Not all model parts need to be in the memory simultaneously. For example, The CLIP model can be unloaded after processing the prompt, and the VAE is required only after sampling. So, the VRAM required is smaller than the model size.
-
A model’s size is measured by the number of parameters. A parameter can be represented in different precisions in a GPU, e.g., FP32 (high), FP16, and FP8 (low). The lower the precision, the smaller the size of a GPU, but the quality may be reduced.
Optimizations like these enable fixing large models in limited VRAMs.
I use all cloud services in my day job. There’s no hardware maintenance and no noise when running a heavy job. I am fortunate enough that I was allowed to keep the VM running 24/7. I won’t enjoy it as much if I need to shut it down every day.
Some online services, like Think Diffusion, auto-shutdown after a specific amount of time. This is not a bad way to control costs.
If you are on the fence, using an online service is not a bad idea. It is probably cheaper than owning one if you are a casual user.
The newer architectures have some new optimization techniques and can be faster in training and using models.
SD models use GPU differently from gaming applications. A GPU card’s FLOPS number (floating number operations per second) is a good gauge for performance.
4090 is for sure faster than 3090 but they should generate the same image with the same setting. The only difference is how long you wait.
I would only consider 24GB+ VRAM if I buy a GPU card now. Consider it an investment to future-proof your system. A slower card means you need to wait longer. A low-VRAM card means you cannot run certain models at all. (Or you need to jump through hoops to do it)
But if you are happy with the toolset now – SD 1.5, SDXL, Flux, getting a 16GB card is not a bad idea to save some money.