SSD-1B and LCM unable to work

#36
by andykoko - opened

@GuiyeC
Your pre-converted SSD-1B and LCM 8-bit model can work normally.
I used Guernika Model Converter v7.2 to convert 16-bit of these two models, but they didn't work properly at guernika v7.2.

SSD-1B: All the generated pictures are black.
LCM: It seems that it was not recognized as LCM when converting, and Guernika still runs it according to the ordinary SD1.5 model.

After upgrading to guernika v7.2, SDXL cannot use T2I-Adapter. The following error occurs and works normally before upgrading.

2023-10-28 19.10.08.png

Guernika org

@andykoko I'm preparing a new release to fix T2IAdapter, thanks for the report!

For SSD-1B I found that you have to use full-precision when converting, I should prepare a release to automatically do that when using those kinds of models but for now you can select it manually.
For LCM, did you convert the "SimianLuo/LCM_Dreamshaper_v7" model? I don't know what could be happening there or how can it work like a regular SD model.

Okay, I see. No wonder your pre-converted SSD-1B 8-bit model is slower than 16-bit, and the memory usage is also amazing.

I first downloaded the entire diffusers model according to the scripting method under the SimianLuo/LCM_Dreamshaper_v7 directory, and then replaced the unet with LCM_Dreamshaper_v7_4k.safetensors.
And using the previous version of model_index.json, there is no problem with the conversion process.

Forget it, let's wait for SDXL LCM. I just tasted it and got used to the picture quality of SDXL.
LCM is really good. The sampling time of each step is only half of the other schedulers, and the memory required is lower.

Guernika org

I'll try to see why 16 bit doesn't work with SSD-1B but yes, for now that's the only way I found it to work.

I also wanted to ask, do you have any suggestions/requests?
I was thinking on a major overhaul of the app, maybe have a list of jobs instead of the current Create tab, I was also thinking on redesigning the sidebar and collections, are you actually using collections? is it worth it to have it there? Does it make sense to keep the Models/Conditioning tabs in the sidebar or maybe have a joint Model library?

@GuiyeC

  • I don't have a good idea about task list, but I saw a Diffusion app with a similar design, but it should be based on electron.

2023-10-29 21.54.12.png

  • If you want to add the task list function, you can consider using a large language model to generate prompt words in batches. This is an example of using shortcuts. The effect is very good. llama.cpp also has Swift binding.
    https://github.com/czkoko/SD-AI-Prompt

  • For Models/Conditioning tabs, if you want to redesign, I think you can change it to activate it from a new window in the menu, or a small icon at the top of the current window to start a new model management window. After all, its daily usage rate is not very high.

  • For Collections, I have never created a classification. Maybe it is needed by people who rely on it at work every day. Maybe you can add the mode of viewing a single picture, so that the display effect will be better. Convenient to check picture details.
    I think the right sidebar can be kept for collections, so that it is convenient to drag pictures from the library to the ControlNet input window.

11.jpg

@GuiyeC
I simply designed an interface optimized for the canvas works. Here are some of my ideas and design drawings.

  • Put the prompt words at the bottom of the screen to make it easier for the eyes to see when typing, and add the AI prompt function to make simple prompt words more imaginative.

  • Generate 1 or 3 candidate pictures for a single picture generate, similar to Photoshop's Generative Fill.
    Candidate pictures will be previewed with small thumbnails in the lower left corner, and candidate pictures will not be automatically saved, so they will be saved manually.

  • After clicking the Queue button, the number of pictures will automatically pop up. The queuing task will not start automatically. Select to start generating on the Queue page. The pictures generated by queue will be saved directly to the gallery without affecting the pictures on the create page.

  • Pre-set the inpainting model and use it automatically when there is an inpaiting requirement.

  • Place the necessary tools such as drawing masks and setting resolutions at the top of the window.

  • External masks can be imported, and the MiDaS preprocessing model can be used to generate mask for inpainting to replace the background.

555.jpg

Guernika org

@andykoko @ZProphete hey guys! here is a small update on the new version, there is still a lot to do but I wanted to know your thoughts on this, my idea is to have the "creation bar" permanently on the right and you get to see the gallery always and the progress on top, I was thinking in maybe having a bigger "progress" on a separate screen and then maybe adding the option for a "Canvas" or "real time edition" if I get a decent enough editor.
I would love your thoughts on this, specially in the arrangement of the creation sidebar.

@GuiyeC
The design of the job queue looks good, simple and beautiful.
I found that each step of the LCM model can produce pictures with similar overall and completely different details.
I want to know if it is in the 4, 5, 6step... Will there be good quality for the decoding and saving of latent images? Or have to start a new sample from step 1?
If feasible, this can be used as a workflow to generate similar pictures.

I think we can add a help page to the menu bar, because many new users don't know how to use, convert models, and install models correctly.

Here are some suggestions for UI:

  1. For non-square-scale pictures, you may be able to set a blurred square background for them, Or a monochrome that is slightly different from the bottom background color, so that various proportions of pictures may look more neat together.

5.jpeg

  1. Aligning controls such as Stepper between the top and bottom may be more beautiful.

  2. img2img
    Scheme 1:
    Using the current design scheme, change the text buttons such as "select image" and "inpaint" to an easy-to-understand icon button to prevent text from being occlusion after the proportion of the picture box changes. The icon uses a bubble box to prompt its function. If the strength adjustment slider is placed in the picture frame, the distance may be too short to control accurately. You can test it again.

1.jpg

Scheme 2:
Merge img2img picture box with size adjustment.

2.jpg

Scheme 3: @Gomeo @ZProphete
Remove the original steps. Although adjusting it is effective, it feels negligible.
Enlarge the picture input box, more compact layout.
Text encoder can be set in the form of a historical prompt icon, The negative prompt input box is also have it.

3.jpeg

I personally prefer scheme 2 as it looks less visually complex/cluttered, the auto-shifting image ratio box is cool, but is it editable or is it just responding to the values put into the width and height pixel selection?

Hello @GuiyeC and friends,

  • I like @andykoko 's arranging (scheme 2) it's nice and compact !

  • On the other hand, I'd keep the gallery images squares, but overlay/hover some indication like a little blurred pill (like your zoom slider) with the aspect ratio inside.
    or simply go for another darker shade instead of the blur.

  • I guess the conditioning boxes will be fitted in a grid ?
    May you make them small so that it fits 3-4 in a row, their content is obvious they don't have to be very big

  • Add some roundness to the gallery images, everything else is :P
    the ram usage is cool colored feature, I always have task manager / powermetrics opened...

Great changes :)

Hello @GuiyeC , How is the new UI progressing, When can we experience it?

Hey, @andykoko
Have you experienced any intensive swap after OS update ?
I'm like getting 60gig of swap before even the smallest models starts generating.

I was like two OS update late,
I had no reason to update, stupid me.

@ZProphete My macos is currently 14.6, and I don't have the problem of excessive swap.

Sign up or log in to comment