SHARK Stable Release Tracker (12/2023)

[monorimet]

Adding a list of TODOs:

• [x] MultiControlNet

  • [x] resize input image(s) to desired output size
  • [x] verify correctness of pipeline structure -- are we using the right unet weights?
  • [x] enable "controlnet t2i adaptation", i.e. t2i with controlnet adapters ( ? )
  • [x] zoedepth import doesn't work
  • [x] stencil canvas isn't centered in image editor

• [ ] SDXL

  • [ ] LoRAs
  • [ ] vulkan issues with SDXL-1.0 1024x1024
  • [ ] controlnet
  • [ ] verify guidance scale implem.

• [x] fix nod logo

• [x] fix UX regressions caused by gradio version update

[countradooku]

can we have sdxl turbo too added?

[monorimet]

can we have sdxl turbo too added?

Done. https://github.com/nod-ai/SHARK/releases/tag/20231205.1064

[countradooku]

i get this on rx 7900xtx on default settings on sd xl turbo

[countradooku]

@monorimet

[one-lithe-rune]

@monorimet Now I've got my LoRA changes vaguely into shape, I'll take a look the nod-ai logo and the gradio update regressions.

EDIT: /em looks at time, sometime tomorrow if no-one has picked them up.

[countradooku]

@monorimet it works nevermind

[mergmann]

When switching to sdxl-turbo, it disables the positive prompt, leaving only the negative one. Instead, I have to write the prompt while in sdxl-base mode, then switch to turbo.

[mergmann]

The sdxl-turbo model works mostly fine for me, but trying to run sdxl-base results in SystemExit: SharkifyStableDiffusionModel.get_unet_sdxl.<locals>.UnetModel.forward() missing 2 required positional arguments: 'time_ids' and 'guidance_scale' OS: EndeavourOS/Arch Linux GPU: RX 7900 XT (using amdvlk as mesa/vulkan-radeon is broken) Settings: 512x512 50 steps

[monorimet]

When switching to sdxl-turbo, it disables the positive prompt, leaving only the negative one. Instead, I have to write the prompt while in sdxl-base mode, then switch to turbo.

There's something goofy about the inputs/guidance scale for the SDXL implementation. This is actually intended behavior as a quick'n'dirty workaround but it's not a long term solution. I'll add a task for it.