v0.8.0 release checklist

[ssube]

Issues:

• add a warning or note when a model does not support inpainting #54
• ONNX acceleration for Real ESRGAN v3 #113
• add an option to remove temporary files and models after converting #122
• add i18next for internationalization/localization #127
• add an option to run face correction before upscaling #132
• use stage parameters before request parameters for chain pipelines #138
• log VRAM usage along with GC stats #141
• add a way for users to add friendly names to their extras models #144
• add a way to select a different VAE when converting diffusion models #145
• export Real ESRGAN v3 at each of the supported scales #151
• error when running Real ESRGAN x4 models with x2 scale #152
• add support for attention slicing and other CUDA optimizations #155
• run Torch/CUDA garbage collection during regular GC #156
• add support for LoRA weights #157
• add an option to turn off progress bars in server logs #158
• fix react element warnings  #160
• add a way to export state for later or debugging  #161
• any error during conversion causes remaining models to be skipped  #166
• torch JIT fails to load with various PKL file not found errors #167
• can't start new thread error after running for a while #170
• empty black output when inpainting small images  #172
• add initial support for Textual Inversion models #179
• automate the release test checklist #180
• make sure the client loads even without a server #181
• add the prompt parameter to upscaling tab #187
• blend tab sources are reversed #188
• add requirements files for each platform #193
• add eta parameter for DDIM scheduler #194
• add a way to generate a batch of images with one request #195
• error converting models with non-default models directory #221
• model_config not found error when conversion is interrupted part way through #222
• error when enabling long prompt weighting #224
• add support for UniPC multistep scheduler #226
• fix the empty/none options in the mask filter menu #228
• invalid selection: 1 error when setting tile order #229
• error running gfpgan in txt2img #233
• error running SD upscaling with txt2img #234

Merges:

• chore(deps): update dependency esbuild to ^0.17.0 #45
• load custom vae when converting models #163
• feat(build): add livereloading to esbuild #169
• fix(build): Explicitly set esbuild serve host/port #173
• fix inpaint for small images #174
• ONNX acceleration for Real ESRGAN v3 #175
• add parameter to run correction before upscaling #176
• CUDA optimizations #177
• prefer chain stage parameters over request parameters #178
• add support for textual inversion #199
• add batch size parameter #200
• feat(api): add txt2txt endpoint #204
• per device worker #205
• support half-precision models #206
• i18next and initial translations #208
• support for ckpt inversions #209
• feat: allow users to add their own labels for models (#144) #214
• allow client to load without a server #215
• platform-specific requirements #218
• track and recycle leaking workers #220
• check for index file during model conversion #223

Testing:

• txt2img
  • 256x256 with SD v1.5
    • should fail: neon blobs worked, but that's ok
    • has automation
  • 512x512 with SD v1.5
    • DEIS Multi
      • should work
    • DPM Multi
      • should work
      • has automation
    • Euler A
      • should work
    • Heun
      • should work
      • has automation
  • 512x512 with SD v2.1
    • should work
    • has automation
  • 768x768 with SD v2.1
    • should work, given sufficient memory
    • has automation
  • extra models
    • 512x512 with Knollingcase
      • should work
      • has automation
    • 512x512 with OpenJourney
      • should work
      • has automation
    • 256x256 with OpenJourney
      • should work
  • textual inversion
    • 512x512 with SD v1.5 + Autumn (embeddings)
      • should work
    • 512x512 with SD v1.5 + Minecraft (concept)
      • should work
• img2img
  • 256x256 input
    • should work: expanded into valid image
    • has automation
  • 512x512 input
    • should work
    • has automation
  • 1024x768 input
    • should work
• inpaint
  • regular inpaint
    • black mask
      • should keep all pixels, same image
      • has automation
    • white mask
      • should replace all pixels, different image
      • has automation
  • outpaint
    • 0 all sides
      • should work, run 1 tile
      • primarily a client-side test
    • 256 all sides
      • should work, run 8 tiles
      • has automation
    • 512 top and bottom, 0 left and right
      • should work, run 3 tiles
      • has automation
    • 512 left and right, 0 top and bottom
      • should work, run 3 tiles
      • has automation
• upscale
  • Real ESRGAN
    • x4 with CodeFormer
      • should work
    • x4 with GFPGAN
      • should work
    • x4 without face correction
      • should work
      • has automation
    • x2 without face correction
      • should work
      • has automation
    • x2 model and x4 scale
      • should sort of work: ignores scale and uses x2
    • x4 model and x2 scale
      • should fail: tiles
    • v3 model and x4 scale
      • should work
  • Stable Diffusion
    • using x2 scale
      • should fail: tiles
    • using x4 scale
      • should work
    • with CodeFormer
      • should work
    • with GFPGAN
      • should work
    • without face correction
      • should work
• blend
  • two 512x512 inputs
    • should work
  • two 1024x1024 inputs
    • should work
  • two different size inputs
    • 256x256 and 512x512
      • should work
    • 512x512 and 1024x1024
      • should work
• interactions
  • generate a new image
    • should request and then load an image from the server
  • delete a pending image
    • should remove a single image and leave the rest
  • delete a finished image
    • should remove a single image and leave the rest
  • copy an image to img2img
    • should switch to the img2img tab
    • should populate the image source
    • the generate button should be enabled
  • copy an image to inpaint
    • should switch to the inpaint tab
    • should populate the image source
    • the generate button should be enabled
  • copy an image to upscale
    • should switch to the upscale tab
    • should populate the image source
    • the generate button should be enabled
  • copy two images to blend
    • should switch to the blend tab
    • should populate the image sources
    • the generate button should be enabled once both sources have been populated
  • state should persist on refresh
    • loading images
    • switching tabs
    • images sources do not
• schedulers
  • DDIM
  • DDPM
  • DEIS Multi
  • DPM Multi
  • DPM Single
  • Euler A
  • Euler
  • Heun
  • iPNDM
  • KDPM2 A
  • KDPM2
  • Karras Ve
  • LMS
  • PNDM
  • UniPC Multi

Release:

• fix any late arrival bugs
• check on/add UniPCMultistepScheduler
• document ONNX_WEB_EXTRA_ARGS
• more bugs
• update release tests to include Textual Inversions
• check and fix lint
• update package versions and stage files
• run commit-and-tag-version --sign --git-tag-fallback --commit-all --release-as=minor to make VERSION
• make sure packages and images have been built
• update GH pages bundle and default version
• post release on GH
• make follow up tickets
• close milestone and checklist

[ssube]

There are a few bugs around model conversion that still need to be fixed, as well as LoRA token expansion.

[ssube]

All of the bugs that have been reported so far have been fixed, and the release test script passed on my Win10 + AMD + DirectML and Ubuntu + Nvidia + CUDA systems. Unfortunately, my Ubuntu + AMD + ROCm test machine is not cooperating, so I need to fix that or will have to release without specifically testing ROCm, which I don't like very much.

The new ONNX_WEB_EXTRA_ARGS variable still needs to be documented and I need to check if the UniPCMultistepScheduler can be added yet.

[ssube]

• the Karras Ve scheduler is still broken (TypeError: KarrasVeScheduler.step() missing 1 required positional argument: 'sample_hat')
• iPNDM is still producing chips (LMS and PNDM may as well without enough steps)
• the Generate button on the Blend tab is enabled even without the second source