blepping

Results 15 comments of blepping

it's not really possible, since you plug sigmas directly into the `SamplerCustom` node. what denoise 0.5, 20 steps is doing behind the scenes is generating 40 steps and splitting the...

started trying to test this. first the [job-iterator](https://github.com/ali1234/comfyui-job-iterator) extension has to be disabled. it also is not compatible with [rgthree's nodes](https://github.com/rgthree/rgthree-comfy) even if you disable the executor stuff in it...

just tried out the latest revision: ![image](https://github.com/comfyanonymous/ComfyUI/assets/157360029/6461f17a-b229-4b06-ad00-db45806e6981) fails due to ```plaintext ERROR:root:Failed to validate prompt for output 145: ERROR:root:* SimpleMath+ 54: ERROR:root: - Return type mismatch between linked nodes: a,...

something really strange is going on in the latest version. i ran my workflow from the broken seed initialization when add noise is turned off pull: #2841 and got ![workflow-wut](https://github.com/comfyanonymous/ComfyUI/assets/157360029/1c490fd7-ac5b-40d7-8f73-3fb4ffd602cf)...

i made a node with the MSW-MSA attention part: https://gist.github.com/blepping/02e389f660112097983684a8ea8093b1 it's a noticeable speed increase for high res generation and seems like it might also increase quality/reduce artifacts.

@patientx > I know about that but this implementation doesn't seem to work at all, neither it speeds up at normal resolutions or high resolutions nor it seem to enhance...

i did some more testing with SDXL. from what i can see, it's _probably_ not really worth using for SDXL at least from a performance standpoint. also the available blocks...

experimental implementation of the remaining parts: https://github.com/blepping/comfyui_jankhidiffusion i strongly recommend reading the README since there are more than a few gotchas.

@ppbrown > The node I'm suggesting, has nothing to do with automatic download functionality? it runs a model to convert the latents, so the node can't do anything without that...

> I think the problem is that the VAE decode ADDS information. basically the same thing. or another way to talk about it is the vae encode/decode process causes changes:...