f97f6a3b4b
feature: use refiners library for generation
...
BREAKING CHANGE
- stable diffusion 1.5 + inpainting working
- self-attention guidance working. improves image generation quality
- tile-mode working
- inpainting self-attention guidance working
disable/broken features:
- sd 1.4, 2.0, 2.1
- most of the samplers
- pix2pix edit
- most of the controlnets
- memory management
- python 3.8 support
wip
11 months ago
558d3388e5
style: speed up linting and autoformatting. fix lints
1 year ago
8243ed616d
fix: pydantic models for http server working now. Fixes #380
1 year ago
7c2004bfcc
feature/fix: migrate to pydantic 2.3
...
- test: add schema tests/fuzzer and fixes
- fix default prompt. add tests
- fix outpaint and controlnet defaults
- fix init image strength defaults
1 year ago
82d74c6b49
feature: switch to pydantic models
...
- allow prompt re-use by deferring random seed
1 year ago
38ac0b7f54
tests: fix tests
1 year ago
df25936d6f
feature: automatic use of inpainting
...
feature disabled since controlnet inpainting doesn't work great. Was disabled by setting `inpaint_method="finetune",`
1 year ago
d32e1060cd
feature: multi-controlnet support at the command line
...
add controlnet option for edit demo
1 year ago
bcaa000d35
fix: model logging
1 year ago
926692ad03
tests: "prime" the controlnets
...
Trying to get things working on m1. doesn't fix everything
1 year ago
eca97a25a0
tests: adjust tests to pass
1 year ago
4c77fd376b
feature: improvements to memory management
...
not thoroughly tested on low-memory devices
1 year ago
6db296aa37
tests: faster tests
1 year ago
c082ea523f
tests: update tests
...
- controlnet version changes + graphics card change
1 year ago
db8d3b74ea
feature: add "detail" control mode.
...
Add details to an image. Great for upscaling an image.
1 year ago
476a81a967
feature: add "shuffle" control mode
...
Image is generated from elements of control image. similar to style transfer
1 year ago
832adf27bc
feature: update controlnets to 1.1
...
- smaller filesizes for the controlnet weights
- fix: works with other 1.5 based models
- enables more control modes getting added
1 year ago
95d3d08d27
fix: better input image path handling
...
throw exception for non-existent images
2 years ago
37d6642c83
fix: fix model downloads that were broken
...
by [library change in transformers 4.27.0](8f3b4a1d5b
2 years ago
3b777b98d8
fix: improved large images using composition
2 years ago
40da2b0357
tests: fix tests
2 years ago
70c58467c0
fix: img2img was broken for all samplers except ddim,plms
...
img2img was broken for all samplers except plms and ddim when init image strength was >~0.25. Been this way for a while. whoops
2 years ago
e3b23584ed
refactor: split up command line tool
2 years ago
b261c62d4e
docs: update docs
2 years ago
54c3ad51d6
feature: controlnet
2 years ago
bcd761b8a6
tests: better coverage of cli commands
...
add "empty" model that just uses empty state for models
2 years ago
7b684b6e5c
test: add smoke test for edit command
2 years ago
1563e0b871
test: add some autoencoder tests
...
the fold-unfold encoding/decoding looks like it's slower and has worse seams than the sliced feathering approach
2 years ago
b93b6a4d7c
perf: use silu instead of nonlinearity for speedup
2 years ago
003a512dc8
perf: mps perf improvement
...
attention slicing wasn't working
2 years ago
ea1d4baafe
refactor: cleanup image generation code
2 years ago
8a97213622
feature: sliced latent decoding
...
allows generation of bigger images. tile seams can be noticeable occasionally despite the feathering
2 years ago
9eacf5e7ed
perf: improve startup time by doing some imports lazily ( #233 )
...
just running `aimg --help` or `aimg --version` was very slow due to all the imports being brought in eagerly
Before changes `aimg --help`
`2.24s user 4.05s system 184% cpu 3.416 total`
After changes:
`0.04s user 0.02s system 8% cpu 0.625 total`
Used `PYTHONPROFILEIMPORTTIME=1 aimg --help` to find time consuming imports.
Also switched to using `scripts` instead of `entrypoints` since the scripts are much faster.
Made duplicate SAMPLER_TYPE_OPTIONS that can be loaded without loading all the samplers themselves.
Likely a breaking change - not sure.
2 years ago
9ee09ac842
feature: add compilation animations ( #224 )
...
- add generation/compare gifs
2 years ago
542e4fbd55
fix: tile mode was broken since latest perf improvements ( #220 )
2 years ago
7deabaae62
tests: update expected images for tests ( #215 )
2 years ago
8855823293
fix: rename files for windows compat
2 years ago
5b56f05da7
feature: print version
...
- fix: default config wasn't being selected for custom models
2 years ago
81f294216b
feature: 🎉 outpainting
2 years ago
24e10f9e5f
fix: don't report a safety issue when a black image is generated
2 years ago
5cc73f6087
feature: finetuning
...
- feature: finetuning your own image models
- feature: image prep command. crops to face or other interesting parts of photo
- fix: back-compat for hf_hub_download
- feature: add prune-ckpt command
- feature: allow specification of model config file
2 years ago
9e0a9e2c29
tests: update
2 years ago
eb40842078
test: more flexible embedding test
2 years ago
1381c7fed4
lint: new ruff linter
2 years ago
da0f1e1ee6
fix: make sure randomness is generated on cpu for consistency
2 years ago
1f7403155e
fix: more thorough cleaning of memory when switching models
...
also cleanup up some test failures
2 years ago
f30823e0b5
feature: Stable Diffusion 2.1
2 years ago
2f8edc1530
feature: negative prompting
...
- feature: negative prompting. `--negative-prompt` or `ImaginePrompt(..., negative_prompt="ugly, deformed, extra arms, etc")`
- feature: a default negative prompt is added to all generations. Images in SD-2.0 don't look bad anymore. Images in 1.5 look improved as well.
2 years ago
554fc1d0ec
tests: update images
...
- probably the k-diffusin update fixed some bugs
- add missing images
2 years ago
257752887d
perf: add back memory efficiency improvements
...
Removed these in a rush to get SD-2.0 out.
2 years ago
Bryce