[Leeloo Please JPG]
DO ME A FAVOR if you like what I'm doing...
Please consider to test my trained models with different topics and compare it with another models in order to see if my goodies are adding anything of value to the SD world. I will be glad to get a constructive feedback about what is GOOD and what is BAD about my models.
So, by adding the REVIEW with comments you will help me to get energy and motivation to move on.
Thanks in Advance!!!
!!!NotaBene!!! Load my example images to ComfyUI.
I use complex and simplistic ComfyUI workflows with upscaling and without. To get a tips for prompts you can press a "NN Nodes" button in image resources and paste it to ComfyUI.
Also, I often repeat some popular prompts from version to version in order to show the difference, so you can compare and get an idea of what's new.
MAIN GOAL
Main focus is a balance between photorealism, creativity, bokehlicious rendering with shallow depth of field, anatomy improvements, emotions, artistic and analog style.
PS: if you like what I'm doing you can Buy me a Coffee
Description
BugFix after feedback by @Raken
This is RC8 + CinEro LoraXL (Epoch 32) + CinEro v1.3 elemental merge.
PS: I use Fine Tuning via LoCon training both U-Net and CLIP because have only 24GB of VRAM which don't let me to train CLIP with U-Net directly.
FAQ
Comments (18)
Model Toolkit says this checkpoint has a ton of junk data:
Model is 6.46 GB. Model type identified as VAE-v1-BROKEN. Model components are: VAE-v1-SD. Contains 6.31 GB of junk data! Model will be pruned to 159.56 MBThat's some issue with tooling. This is the SDXL model and it can't have such small size. I believe VAE-v1-SD detection is falsy. It is not a VAE and not a v1.
Sorry, but either "Model Tookit" is junk, or you are using it wrong. As @homoludens said, this is an SDXL model and 6.31GB is the right model size.
I am not sure what the point of your comment is to begin with.
@Nonsensei What are you on about? I just loaded SDXL 1.0 base into toolkit and it shows things just fine:
Model is 6.46 GB. Model type identified as SD-XL. Model components are: UNET-XL-SD, VAE-v1-SD, CLIP-XL-SD, CLIP-XL-AUX-SD.
Contains no junk data.
Model is unaltered, nothing to be done.
Have at it if you want to try on your own: https://github.com/arenasys/stable-diffusion-webui-model-toolkit
@i860 Maybe Kohya_SS or some ComfyUI custom node affects some metadata section of checkpoint?
But I have no idea what goes wrong and don't see a strict need of investigation so far.
Is there a real live issue appeared with my model in A1111 WebUI? Some consequences of metadata discrepancy?
@Nonsensei I can expect some issues during saving of the checkpoint. It is just a plain poorly structured array of bytes. Maybe my tolling don't write all metadata or writes it at wrong offsets.. donno
@i860 I read the description of https://github.com/arenasys/stable-diffusion-webui-model-toolkit and the point of the toolkit is that "Many models being distributed are quite bloated, most of their size being redundant or useless data."
But the regular size of a SDXL model IS 6.4GiB, so it cannot be any smaller than that. Since the model is working, it is clear that there are some bugs in "WebUI model tookit". This means that the toolkit has parsed this model incorrect. That it can parse base SDXL correctly means nothing, since that is the first test case that the program has to pass. Just because the program works in once case does not mean that it is bug free.
So if this is a concern for you (I don't know why, but I am sure there is a reason), please file a bug report against "WebUI model tookit", rather than trying to get @homoludens to fix anything.
Trust me, I am a programmer, and I know how difficult it is to write code that parse data files correctly. The author of "Model Toolkit" will thank you for filing this bug report.
@Nonsensei "Trust me! I'm an engineer!" :D
I'm also programmer 🍻
@homoludens I know that model creators don't want to piss off anyone because they don't want people to leave a "revenge review". So sometimes I volunteer to the outspoken one 😅.
I don't know model makers like you personally, but I know a lot of work, passion and creativity is involved, and I am grateful that you share your work without asking much in return.
So a big thank you from this regular Joe user, and keep up the good work! 👍😁🙏
@Nonsensei Yes I don't want offense anyone from commenters. But honestly I can expect that some python modules I use don't write some important bits of information during training or polishing my model.
The chain is long and bug can be anywhere. If some one will show that the bug is causing severe troubles I should investigate the root cause.
@homoludens Yes, as programmers we know bugs can be everywhere 😅.
But my money is on the bug being in "Model Toolkit" because writing parsers that handle edge cases in a robust fashion is hard work.
@Nonsensei agree
SDXL pruned to 159 MB? yeah thats BS
@colinw2292823 It is a BUG in ComfyUI merge tools. I almost sure. A1111 webui many extensions generates errors on interaction with SAFETENSORS produced by ComfyUI mergers. In any case I should say thanks to Topic Starter for highlighting the problem.
@homoludens This is correct.
It looks like you have switched to generating images in .webp format. I can understand that the file sizes are less byte hungry, but I really appreciate being able to see your workflows (and prompts of course). But thank you for your work all the same.
Workflow should be embedded. I've used a Saving node which (as claimed by author) should embed the workflow. Anyway, I have only one workflow with Saving node and it didn't drag e into using it everywhere.
Thanks for feed back that WEBP is not good...
Thanks for your reply.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.