Feature request

[Fcucgvhhhvjv]

hi I have converted 6-7 models , actually I created a colab notebooks which does it automatically . Most of the model works and give some what good result but dreamlike photo real 2 doesn't work . Also how can we load models from storage rather than building it in the app itself? The feature to use different sampler and save images . Let me know if u want the models . If you can guide me a bit I'll try to modifying the app file so loading from external storage works .

[ZTMIDGO]

If the model cannot run, you can use netron to view the input shape of the model. If it is consistent with what I use, it can usually run, but other situations cannot be ruled out. I'll be modifying the project this week to allow external selection of models

[Fcucgvhhhvjv]

I have checked with both attention slicing auto and max while converting the model but it doesn't work , model card says it was trained on 768 x 768 images and I'm running 256 or 320 maybe that's why?

Also what about the scheduler? Maybe using a scheduler like dpm will make some drastic changes. Im looking forward to it.

I have weak understanding of Java and Android app so I can't do much , I'm learning .

[ZTMIDGO]

Perhaps the model only supports input shapes with a width ratio of 1:1

[Fcucgvhhhvjv]

Can u check it when ur free ill give u the link for the model

hi , i was somehow able to generate 448x448 and 512 x512 , my device only has 6gb ram

this was made with majicmix model https://huggingface.co/Androidonnxfork/test/blob/main/Screenshot_2023-07-31-18-43-14-665_com.example.open.diffusion.jpg https://huggingface.co/Androidonnxfork/test/blob/main/Screenshot_2023-07-31-18-36-40-178_com.example.open.diffusion.jpg

[ZTMIDGO]

你可以在有空的时候检查一下吗我会给你模型的链接

你好，我以某种方式能够生成 448x448 和 512x512，我的设备只有 6GB RAM

这是用 majicmix 模型制作的 https://huggingface.co/Androidonnxfork/test/blob/main/Screenshot_2023-07-31-18-43-14-665_com.example.open.diffusion.jpg https://huggingface.co /Androidonnxfork/test/blob/main/Screenshot_2023-07-31-18-36-40-178_com.example.open.diffusion.jpg

how do you do it Usually there will be a memory overflow。

[Fcucgvhhhvjv]

I dont know myself maybe its the model? i recent increased my ram extension so i have extra 5gb maybe thats why? ill post more pictures let me create first. btw do u have discord or something else so we can chat in there , maybe lets have a discord server so it will give boost to this project .

[Fcucgvhhhvjv]

https://huggingface.co/Androidonnxfork/test/blob/main/Screenshot_2023-07-31-19-25-56-586_com.example.open.diffusion.jpg https://huggingface.co/Androidonnxfork/test/blob/main/Screenshot_2023-07-31-19-29-29-325_com.example.open.diffusion.jpg https://huggingface.co/Androidonnxfork/test/blob/main/Screenshot_2023-07-31-19-44-21-629_com.example.open.diffusion.jpg here , i can confirm it works for 512 generation .

[Fcucgvhhhvjv]

maybe u wanna see my colab notebook to see it ? can u tell me which model was the original it crashed with 512 ill try to convert and check if it it crashes at 512

[ZTMIDGO]

Nice, but might run slower. I can't use chat software like discord, because the Chinese government blocks all foreign chat software

[ZTMIDGO]

也许你想看看我的 colab 笔记本来看看它？你能告诉我哪个型号是原来的吗？它崩溃了 512 我尝试转换并检查它是否崩溃于 512

You can try this: https://huggingface.co/runwayml/stable-diffusion-v1-5

[Fcucgvhhhvjv]

yeah it is inneed slower , 20 steps take like 3-5 min on 512 x 512 . can gpu usage be implement like vulkan backend? sad to hear that , i have Instagram and whatsapp let me know what u can use .

[Fcucgvhhhvjv]

也许你想看看我的 colab 笔记本来看看它？你能告诉我哪个型号是原来的吗？它崩溃了 512 我尝试转换并检查它是否崩溃于 512

You can try this: https://huggingface.co/runwayml/stable-diffusion-v1-5

ill try that tonight .

[Fcucgvhhhvjv]

u can use any resource from this repo for your project, also here are the latest models u can use https://huggingface.co/Androidonnxfork/test/tree/main/fp16fullonnxsdquantized_in_ort , I'll be making the colab notebook work automatically so it can generate models without me having to do it regularly ,

[ZTMIDGO]

yeah it is inneed slower , 20 steps take like 3-5 min on 512 x 512 . can gpu usage be implement like vulkan backend? sad to hear that , i have Instagram and whatsapp let me know what u can use .

Hardware acceleration is currently unavailable, you can learn why here https://github.com/microsoft/onnxruntime/issues/15629

[ZTMIDGO]

u can use any resource from this repo for your project, also here are the latest models u can use https://huggingface.co/Androidonnxfork/test/tree/main/fp16fullonnxsdquantized_in_ort , I'll be making the colab notebook work automatically so it can generate models without me having to do it regularly ,

sure

[Fcucgvhhhvjv]

can models be loaded from external storage now? Also what about saving image and scheduler

[ZTMIDGO]

can models be loaded from external storage now? Also what about saving image and scheduler

Yes, models can now be selected externally

[Fcucgvhhhvjv]

thank you , ill build an app for testing . I need some help , the conversation script from Stable-Diffusion-ONNX-FP16 repo doesnt use gpu for conversion , also linux doesnt support direct ml so the conversion is super slow like 30 min per model , can you check the script or whats wrong? I tried to install onnxruntime-gpu and changed execution provoder to gpu but it didn't work in colab