I am pretty new to Z-Image. This is my first small step into training. So far I have focused on improving nipples. I have some other training going on now for some other aspects that I hope I can get out before to long.
In my limited work with it so far, I have found that DPM++SDE with Beta has gotten my best results. From 1 to 2.5 CFG. I have never changed the shift from the default of 6. I use 8 or 9 steps.
I use it in Forge Neo. Using the lumina setting, the ae.safetensors vae and the qwen_3_4b.safetensors text encoder.
I am calling this version 0.5 since there is more work I have to do. I welcome any comments. In one of my first tests with it I got a head cropped off image, but it doesn't seem to happen often. If you get anything weird, feel free to comment so I know what issues might need to be fixed on the next version.
Description
fp8 version of Acorn iZ-Spinning. Correct file now uploaded. I didn't realize when ComfyUI saves the model file it inflates it back up to the 11.4 GB version. Fixed now.
FAQ
Comments (24)
fp8 plz
I am working on it. Should be up before too long.
@Seeker70 why your FP8 version same size as fp16?
@Unicom When I uploaded the fp8 I hadn't noticed ComfyUI re-inflated the file back to the 11.4 GB file. When I fixed it and tried to re-upload it, it put it in the full model's version. I had to end up re-uploading both files to both versions and then delete the one's that were on the wrong version. It's fixed now.
It's a shame that people still refuse to use ComfyUI as generation standard. But why should we be surprised when most people eat fast food as christmas dinner ;-).
why would you mind other peoples food choices. you are not forced to eat it.
As soon as Comfy makes the generation speed acceptable on Mac, I will immediately consider using it.
It's a shame that ComfyUI doesn't make an interface that doesn't constantly need to be tinkered with to keep the ability to do simple things that work flawlessly in ForgeUI.
@Seeker70 You could use SwarmUI. It's a wrapper around ComfyUI that gives it a Forge-like UI.
@xto I've tried it, don't like it. I have just have much better luck with ForgeUI. I get better images in Forge than Comfy and it always just works. When using Comfy I spend way more time trying to find nodes that work than I do generating stuff. Pretty much all I use Comfy for anymore is merging or converting and sometimes video.
Gist in Bold
ComfyUI is amazing for complex tasks when customized and functional, but having breaking updates every other week makes maintenance of specifically those types of workflows a nightmare.
So many people prefer to only use more simplistic, single-task workflows for reliability, which alternative platforms usually offer a more streamlined approach for.
Considering the frequency of novelties to keep up with, ComfyUI needs to be updated/modified constantly, which inevitably results in this perpetual experimental state in lots of areas that did not have the time to settle yet, making the tinkering and adapting on the user's side a necessity as well, for better or for worse.
@RedPinkRetro Managing my ComfyUI on a regular basis is - to be honest - a huge part of the fun when working/playing with AI based image generation.
Producing as many mediocre mostly pornographic images isn't my favorite goal. Thank god there are also other platforms where you can get your inspiration from than this rotten CivitAI.
@xto nope it's not. Just a simple task - basic comparsion XYZ plot rotating multiple lora checkpoints via Promt S/R and rotating multiple prompts from styles file, variating seed for style but not for lora. Save as one big images with labeled loras and style names.
On comfy it's a hellish spagetty monster. On forge it just 1 drop down and 3 clicks.
@desm0nt I agree. I won't dispute Comfy will do some things that Forge doesn't, but when it comes to generating images, I have found nothing at all that Comfy does better than Forge. Personally for me, any times I have done direct comparisons using the same checkpoints, prompts and setting, Forge outputs have been either the same or better than the Comfy outputs.
I do use Comfy to merge Loras into checkpoints. For that I have a super simple workflow that does it very quickly. I have used it for video as well, but that has also been one the huge time wasters for me. I'm sure I spent more time trying to get a working workflow than I ever did doing the actual video generations. Seems like there are always missing nodes and most times the manager fails to find the nodes to install and I end up having to search all over for a solution to manually install a few nodes after spending more time having to find the nodes to begin with. Overall, Comfy for me is not worth all the extra time wasted. Again, I have not found any instance where Comfy actually produced a better result, so for me, Forge is still king. For video, there has gotten to be enough free or low cost sites online to generate video in a tiny fraction of the time it takes to do it locally, that I really don't even waste time anymore on local video generation. My computer only has a 3080 with 16GB VRAM, so it takes normally 20 to 30 minutes to generate a 10 second video that I can generate in about 30 seconds on TensorHub or TensorArt.
@desm0nt Did you reply to the wrong comment? I said try SwarmUI, not Comfy nodes. SwarmUI has a simple form-based grid generator (in the bottom panel, Tools > Grid Generator). It's similar to the Forge one.
Very slow model, Pussy need vae,
Not sure what you are comparing it to. For me it runs faster than Flux by a fair amount. I do have the smaller fp8 version up now as well. I don't know if it would be any faster or just save on VRAM.
Based on their file sizes and CivitAI's file info section, it seems you've mislabeled the fp16 version as fp8 and vice versa.
Edit: Thanks for the fix!
Yeah, when I uploaded the fp8 I hadn't noticed ComfyUI re-inflated the file back to the 11.4 GB file. When I fixed it and tried to re-upload it, it put it in the full model's version. I had to end up re-uploading both files to both versions and then delete the one's that were on the wrong version.
Welcome back. Unfortunately, this model is no different from the default zit.🫤
It is different, but still a work in progress. That's why this is V0.5 instead of V1. This time around I am training all my own stuff to add to the model and when I want to change one thing, I want it to affect everything else as little as possible. I am starting with nipples since that is almost always the worst thing with any new model. I will work on other areas as well as I get toward V1.
Glad to see the Acorn keeps on spinning! 🙌🏻
Merry Christmas btw! 🎄
Merry Christmas!
Eagerly awaiting fixes for "other body parts".