I tried the A6000 before but ran out of memory without optimisations
I tried the A6000 before but ran out of memory without optimisations
Fin that case
Fyou can enable optimizations and upscale bigger
For you can get A100 with 80GB
Bok will do thats
B*thanks
Fyou are welcome
EFor SUPIR_v13, do we need to reinstall or just git pull?
EAlso can you make a tutorial on what each parameter settings do? Or share best settings you've tested.
EAlso, the Terms of use in the Gradio UI says it's no-commercial use. But the license of SUPIR on Github is MIT open-source License. Is it no-commercial use?
ARoger that, thank you!
BI've been trying this new SUPIR from the Patreon and haven't been able to upscale a single image... It takes forever even on my 4090 and then stalls out each time.
B[Tiled VAE]: Executing Encoder Task Queue: 100%|█████████████████████████████████| 17472/17472 [00:56<00:00, 308.84it/s]
[Tiled VAE]: Done in 57.083s, max VRAM alloc 10934.813 MB
C:\Users\xxxx\Deep\SUPIR\SUPIR\venv\lib\site-packages\torch\nn\functional.py:5476: UserWarning: 1Torch was not compiled with flash attention. (Triggered internally at ..\aten\src\ATen\native\transformers\cuda\sdp_utils.cpp:263.)
attn_output = scaled_dot_product_attention(q, k, v, attn_mask, dropout_p, is_causal)
[Tiled VAE]: Done in 57.083s, max VRAM alloc 10934.813 MB
C:\Users\xxxx\Deep\SUPIR\SUPIR\venv\lib\site-packages\torch\nn\functional.py:5476: UserWarning: 1Torch was not compiled with flash attention. (Triggered internally at ..\aten\src\ATen\native\transformers\cuda\sdp_utils.cpp:263.)
attn_output = scaled_dot_product_attention(q, k, v, attn_mask, dropout_p, is_causal)
![Dazzastrous [4090]](https://cdn.discordapp.com/avatars/871389071686135838/c4a5f4d81d9eeb18d856295d692e2c07.webp?size=40)
Dhow big is your input image
Dmight be too high resolution
B3072x4080
BToo big?
Dyup
Dthats the sort of resolution you want to be upscaling too
Hwhen try install one click , i got error : Traceback (most recent call last):
File "X:\03_General\huy\supir\SUPIR\gradio_demo.py", line 5, in <module>
from gradio_imageslider import ImageSlider
File "C:\Users\AZPC\AppData\Local\Programs\Python\Python310\lib\site-packages\gradio_imageslider__init__.py", line 1, in <module>
from .imageslider import ImageSlider
File "C:\Users\AZPC\AppData\Local\Programs\Python\Python310\lib\site-packages\gradio_imageslider\imageslider.py", line 16, in <module>
from gradio.events import Events
ImportError: cannot import name 'Events' from 'gradio.events' (C:\Users\AZPC\AppData\Local\Programs\Python\Python310\lib\site-packages\gradio\events.py)
Press any key to continue . . .
File "X:\03_General\huy\supir\SUPIR\gradio_demo.py", line 5, in <module>
from gradio_imageslider import ImageSlider
File "C:\Users\AZPC\AppData\Local\Programs\Python\Python310\lib\site-packages\gradio_imageslider__init__.py", line 1, in <module>
from .imageslider import ImageSlider
File "C:\Users\AZPC\AppData\Local\Programs\Python\Python310\lib\site-packages\gradio_imageslider\imageslider.py", line 16, in <module>
from gradio.events import Events
ImportError: cannot import name 'Events' from 'gradio.events' (C:\Users\AZPC\AppData\Local\Programs\Python\Python310\lib\site-packages\gradio\events.py)
Press any key to continue . . .
BYeah that’s way too big. SUPIR shines with really small images
BDo you have any tips to help improve consistency between images, when batch upscaling?
DIdeogram 1.0 , text is really good along with prompt following, kinda like SD3 but released now
AHas anyone manually installed supir?
AI'm a bit lost at this point
For users who can connect to huggingface, please setting LLAVA_CLIP_PATH, SDXL_CLIP1_PATH, SDXL_CLIP2_CKPT_PTH in CKPT_PTH.py as None. These CLIPs will be downloaded automatically.
For users who can connect to huggingface, please setting LLAVA_CLIP_PATH, SDXL_CLIP1_PATH, SDXL_CLIP2_CKPT_PTH in CKPT_PTH.py as None. These CLIPs will be downloaded automatically.
EI used ChatGPT to explain some new parameters in SUPIR:
--edm_steps: Number of steps for the EDM Sampling Scheduler. This determines how many iterations the sampling process will take. A higher number generally means more refinement but takes longer.
--s_stage1: Strength control for the first stage of processing. A negative value indicates that this stage is not used.
--s_churn: A hyperparameter of the EDM process, influencing the variability or diversity in the generated samples.
--s_noise: Another hyperparameter for EDM, affecting the amount of noise introduced during the diffusion process, which can impact the clarity or smoothness of the output.
--s_cfg: Classifier-free guidance scale for prompts. This parameter adjusts the influence of textual prompts on the generation process, affecting how closely the output matches the given textual description.
--s_stage2: Strength control for the second stage of processing. Adjusts the intensity of modifications in the second stage.
linear_CFG and --linear_s_stage2: These settings allow for a linear increase of the s_cfg and s_stage2 parameters from their start points (--spt_linear_CFG and --spt_linear_s_stage2) to their specified values through the process, providing a way to dynamically adjust the guidance strength.
--spt_linear_CFG and --spt_linear_s_stage2: Define the starting points for the linear increase of s_cfg and s_stage2, respectively.
--ae_dtype and --diff_dtype: Specify the data types (precision) used during inference for the autoencoder and diffusion processes, respectively, affecting memory usage and computational speed.
--edm_steps: Number of steps for the EDM Sampling Scheduler. This determines how many iterations the sampling process will take. A higher number generally means more refinement but takes longer.
--s_stage1: Strength control for the first stage of processing. A negative value indicates that this stage is not used.
--s_churn: A hyperparameter of the EDM process, influencing the variability or diversity in the generated samples.
--s_noise: Another hyperparameter for EDM, affecting the amount of noise introduced during the diffusion process, which can impact the clarity or smoothness of the output.
--s_cfg: Classifier-free guidance scale for prompts. This parameter adjusts the influence of textual prompts on the generation process, affecting how closely the output matches the given textual description.
--s_stage2: Strength control for the second stage of processing. Adjusts the intensity of modifications in the second stage.
linear_CFG and --linear_s_stage2: These settings allow for a linear increase of the s_cfg and s_stage2 parameters from their start points (--spt_linear_CFG and --spt_linear_s_stage2) to their specified values through the process, providing a way to dynamically adjust the guidance strength.
--spt_linear_CFG and --spt_linear_s_stage2: Define the starting points for the linear increase of s_cfg and s_stage2, respectively.
--ae_dtype and --diff_dtype: Specify the data types (precision) used during inference for the autoencoder and diffusion processes, respectively, affecting memory usage and computational speed.
Fi think if you dont make a SaaS out of it they wont care
Fjust git pull
ECool
Fyes too big 
Fhello. when you were installing did you see any error on cmd?
Fcan you reinstall into a new folder and share entire install cmd logs with me?
D@Furkan Gözükara SECourses do you know how to make SUPIR work on ComfyUI with reduced memory? I’ve been trying to run the ComfyUI node but it runs out of VRAM - I’m using Colab T4 which has 15gb of VRAM
Fsadly no idea
Fi am not really expert at comfyui
DWhat is it that you changed on your version? Is it that you’re loading the SUPIR models as fp16 instead? Is that what the half weights thing is?
Fwe changed SUPIR model weights precision
Fand added tiled VAE
Fboth from official repo
DFrom FP32 -> FP16?
Fyep
Fhalf
BJust FYI, I upscale a pano 2048x1024 to 4096x2048 (2x) and it was eating over 20 gigs of Vram. That's the limit my 3090 can do. It did add a lot of details. Some area didn't look good so I did another version of 2x with Topaz Gigapixels. I then cherry pick masked in area I like best from the two with Photoshop. You can see the final pano here "https://sapphiregreenearth.com/stabledifussionwaterfall2/"
Thinking back, maybe Magnific can be justified for their high price, considering the amount of Vram they need time the amount of servers they use.
Thinking back, maybe Magnific can be justified for their high price, considering the amount of Vram they need time the amount of servers they use.
Sapphire Green Earthstaff
Sapphire Green Earth is dedicated to creating a convenient and affordable path to the use of top green , natural and organic products.
ADeepfaking the young rock tv show actor into actual Dwayne Johnson haha
BCan't wait for metadata in PNG for InstantID GUI. You rock Dr.
BNewbie questionif one have windows 10 x64 which of these two apply
Boh i have intel cpu
Fyep i will add hopefully soon
Fi plan to add both metadata