bdsqlsz

/

qinglong_controlnet-lllite

like 232

use comfyui or webui, not support diffusers now.
https://huggingface.co/kohya-ss/controlnet-lllite/discussions/3

if i wanted to write a python script for upscaling, should I look at how the model weights are being loaded in comfyui then?

if i wanted to write a python script for upscaling, should I look at how the model weights are being loaded in comfyui then?

search controllllite nodes in comfyui

bdsqlsz, I've been using this model in sdwebui-forge, and I just want to check in on what your intended usage is.

Model: bdsqlsz_controlllite_xl_tile_realistic [12b261fe]

At first, I assumed it was to be used in conjunction with a tiled diffusion script (like previous SD1.5 Tile controlnet models.)

However, if I enable this along with Multidiffusion (AKA "TiledDiffusion with Tiled VAE"), I get this error regardless of any settings I've tried:

    Traceback (most recent call last):
      File "C:\0_SD\stable-diffusion-webui-forge\modules\call_queue.py", line 57, in f
        res = list(func(*args, **kwargs))
    TypeError: 'NoneType' object is not iterable

I have only had successful results (No error) when I do a simple Img2Img resize with this controlnet model (No tile diffusion script enabled).

I've been using settings like Resize by: 3.0, Denoise strength: 0.3 with pretty nice results IMO, but usually a bit grainy compared to output I'm accustomed to from 1.5 tiled upscaling.
I just want to that I'm actually using the model as intended, or if the results I'm getting are a limitation of the model.

I've downloaded someone else's Tile controlnet model (tplanetSDXLControlnet_Tile_v10F16 [6c558c4d]) which functions without errors using a tiled diffusion script, like all other Tile models I've used before.

Please advise, thanks!

bdsqlsz, I've been using this model in sdwebui-forge, and I just want to check in on what your intended usage is.

Model: bdsqlsz_controlllite_xl_tile_realistic [12b261fe]

At first, I assumed it was to be used in conjunction with a tiled diffusion script (like previous SD1.5 Tile controlnet models.)

However, if I enable this along with Multidiffusion (AKA "TiledDiffusion with Tiled VAE"), I get this error regardless of any settings I've tried:

    Traceback (most recent call last):
      File "C:\0_SD\stable-diffusion-webui-forge\modules\call_queue.py", line 57, in f
        res = list(func(*args, **kwargs))
    TypeError: 'NoneType' object is not iterable

I have only had successful results (No error) when I do a simple Img2Img resize with this controlnet model (No tile diffusion script enabled).

I've been using settings like Resize by: 3.0, Denoise strength: 0.3 with pretty nice results IMO, but usually a bit grainy compared to output I'm accustomed to from 1.5 tiled upscaling.
I just want to that I'm actually using the model as intended, or if the results I'm getting are a limitation of the model.

I've downloaded someone else's Tile controlnet model (tplanetSDXLControlnet_Tile_v10F16 [6c558c4d]) which functions without errors using a tiled diffusion script, like all other Tile models I've used before.

Please advise, thanks!

Can you give me your result images?

Because of the structural differences, the controllite tile is not used in the same way as a normal controlnet

Surely - just wanted to be sure that this isn't intended for using with tiled methods (Multidiffusion / SD Ultimate Upscaler / etc)

And I want to restate that I do like the results, there's just a bit more random noise pattern (unsure how else to describe) compared to the SD 1.5 Tiled controlnet methods.

Thanks!

Original 1024

After Img2img with your Tile model

Original 1024

After Img2img with your Tile model (and some inpainting on the paws)

Surely - just wanted to be sure that this isn't intended for using with tiled methods (Multidiffusion / SD Ultimate Upscaler / etc)

And I want to restate that I do like the results, there's just a bit more random noise pattern (unsure how else to describe) compared to the SD 1.5 Tiled controlnet methods.

Thanks!

Original 1024

After Img2img with your Tile model

Original 1024

After Img2img with your Tile model (and some inpainting on the paws)

Sorry I'm in hospital right now it's hard to demo on a computer, in short lllite's controls rely more on the prompts than the model itself, so it has less impact on the lora and base model compared to controlnet.

So its upsampler is relatively cleaner in terms of supplemental detail and requires additional noise to be added to the zoom to get more detailed performance.

I’m sorry to hear that (necessity of hospital visit). Thank you for your responses.

I’m mainly just curious to know your own personal usage with your tile model?

-Do you use A1111, Forge, ComfyUI?

-Do you use a tiled upscaling method with this model? (Multidiffusion, SD Ultimate, etc)?
Or, just a normal upscale where the image is not divided into tiles and merged together?

-Do you apply your tile model with any other extension than good old ControlNet?

Many thanks, and I hope you or your loved one gets better soon

I’m sorry to hear that (necessity of hospital visit). Thank you for your responses.

I’m mainly just curious to know your own personal usage with your tile model?

-Do you use A1111, Forge, ComfyUI?

-Do you use a tiled upscaling method with this model? (Multidiffusion, SD Ultimate, etc)?
Or, just a normal upscale where the image is not divided into tiles and merged together?

-Do you apply your tile model with any other extension than good old ControlNet?

Many thanks, and I hope you or your loved one gets better soon

1、i use A1111 forge mainly,sometimes comfyui with animation...

2、simple upscaling with T2I highfix is useful in a1111 and forge.
i recommend Multidiffusion if you need more details.

3、most time i just use highfix with tile,i use it with change clothing more.

1、i use A1111 forge mainly,sometimes comfyui with animation...

2、simple upscaling with T2I highfix is useful in a1111 and forge.
i recommend Multidiffusion if you need more details.

3、most time i just use highfix with tile,i use it with change clothing more.

Correct me if I'm wrong, otherwise I'm very thankful for your insights:

-It sounds like you don't use the Tile model for big tiled upscaling - only to suppress duplicates when using HR Fix.

-You recommend T2I highfix for tiled upscaling methods (not the Tile controlnet model)

Thanks!

1、i use A1111 forge mainly,sometimes comfyui with animation...

2、simple upscaling with T2I highfix is useful in a1111 and forge.
i recommend Multidiffusion if you need more details.

3、most time i just use highfix with tile,i use it with change clothing more.

Correct me if I'm wrong, otherwise I'm very thankful for your insights:

-It sounds like you don't use the Tile model for big tiled upscaling - only to suppress duplicates when using HR Fix.

-You recommend T2I highfix for tiled upscaling methods (not the Tile controlnet model)

Thanks!

Sorry for the somewhat late reply, I just had surgery yesterday and am still recovering.

I think it's because the structure of controlllite causes it works before UltraUpscaleSD as well as mutil's tiling process, so it's used to zoom in directly rather than tiling zoom out.

Actually, tiling upscale is SD1.5 ways to save VRAM.
SDXL recommend use latent upscale and add noise to make more details.