Open wardensc2 opened 7 months ago
Hi there, I'm currently looking into it, the closest I see is that ToMe is applied here although its not clear where it imports from.
https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/dev/modules/sd_models.py#L866
I think that is token merge in Cross attention optimization in Settings but it can't get much speed or reduce Vram like your node, I already try but the result is bad, the image is uglier
People already update your node into Forge but still no one write an extension for A1111 yet sadly
forge added model patcher similar to comfyUI, can't just copy them directly
People already update your node into Forge but still no one write an extension for A1111 yet sadly
I'm not sure if it's work 100% correctly, but you can try https://github.com/AG-w/sd-webui-todo/
People already update your node into Forge but still no one write an extension for A1111 yet sadly
I'm not sure if it's work 100% correctly, but you can try https://github.com/AG-w/sd-webui-todo/
Thank you so much
we're also exploring it over here https://github.com/ethansmith2000/ImprovedTokenMerge/issues/2 the current method im using lets us circumvent having to use the setup a1111 uses for ToMe but also still unsure on some other parts
People already update your node into Forge but still no one write an extension for A1111 yet sadly
I'm not sure if it's work 100% correctly, but you can try https://github.com/AG-w/sd-webui-todo/
Hi AG-w I try to run it on A1111 so far it only work in some resolution, I try to render 1920x1080 A1111 will stop running even when I disable this extension. Thank for your effort it seem that we need to wait for ethansmith2000 to solve this problem
I was implementing ToDo for kohya/sd-scripts (I found it greatly speeds up training). The code was self-contained enough that I was able to turn it into an extension. https://github.com/feffy380/sd-webui-token-downsampling
Instead of individual downsample factors for different depths, I've opted to add a single max_depth slider which determines which layers are affected (basically tomesd's max_downsample
argument but renamed to avoid confusion). I found the quality/speed tradeoff terrible at deeper layers so I didn't feel the granularity was necessary.
There's also the option to disable ToDo after a certain percentage of steps. I found this helps with details, though you sacrifice some speed. You're essentially using the original model as a refiner for the ToDo model.
Awesome to hear! The reason why I set multiple depths is because for 2048x2048 I found using 4 and 2 for the top layer and one below it respectively works pretty well.
Ah fair enough, I hadn't gone that high. I'll see about adding sliders to override the downsample factor for individual layers.
Hi ethan
Can you support us write an extension for A1111 or Forge. Your plugin not only improve render speed, it also reduce about 30% or more Vram when upscale and really necessary for us to create high resolution image
Thank you so much