- This PR enables two new commands in the invoke.py script
!models -- list the available models and their cache status
!switch <model> -- switch to the indicated model
Example:
invoke> !models
laion400m not loaded Latent Diffusion LAION400M model
stable-diffusion-1.4 active Stable Diffusion inference model version 1.4
waifu-1.3 cached Waifu anime model version 1.3
invoke> !switch waifu-1.3
>> Caching model stable-diffusion-1.4 in system RAM
>> Retrieving model waifu-1.3 from system RAM cache
The name and descriptions of the models are taken from
`config/models.yaml`. A future enhancement to `model_cache.py` will be
to enable new model stanzas to be added to the file
programmatically. This will be useful for the WebGUI.
More details:
- Use fast switching algorithm described in PR #948
- Models are selected using their configuration stanza name
given in models.yaml.
- To avoid filling up CPU RAM with cached models, this PR
implements an LRU cache that monitors available CPU RAM.
- The caching code allows the minimum value of available RAM
to be adjusted, but invoke.py does not currently have a
command-line argument that allows you to set it. The
minimum free RAM is arbitrarily set to 2 GB.
- Add optional description field to configs/models.yaml
Unrelated fixes:
- Added ">>" to CompViz model loading messages in order to make user experience
more consistent.
- When generating an image greater than defaults, will only warn about possible
VRAM filling the first time.
- Fixed bug that was causing help message to be printed twice. This involved
moving the import line for the web backend into the section where it is
called.
Coauthored by: @ArDiouscuros
This reverts commit 5f42d08945.
This fix was intended to solve issue #939, in which ESRGAN generates
dark images when upscaling 4X on certain GTX cards. However, the fix
apparently causes conflicts with some versions of the ESRGAN library,
and this fix will have to wait until after release of 2.0.
- txt2img2img back to using DDIM as img2img sampler; results produced
by some k* samplers are just not reliable enough for good user
experience
- img2img progress message clarifies why img2img steps taken != steps requested
- warn of potential problems when user tries to run img2img on a small init image
- img2img confirmed working with all samplers
- inpainting working on ddim & plms. Changes to k-diffusion
module seem to be needed for inpainting support.
- switched k-diffuser noise schedule to original karras schedule,
which reduces the step number needed for good results
-if readline.set_auto_history() is not implemented, as in pyreadline3, will fall
back gracefully to automatic history saving. The only issue with this is that
-!history commands will be recorded in the history.
-!fetch on missing file no longer crashes script
-!history is now one of the autocomplete commands
-.dream_history now stored in output directory rather than ~user directory.
An important limitation of the last feature is that the history is
loaded and saved to the .dream_history file in the --outdir directory
specified at script launch time. It is not swapped around when the
--outdir is changed during the session.
Add message about interpolation size
Fix crash if sampler not set to DDIM, change parameter name to hires_fix
Hi res mode fix duplicates with img2img scaling
-if readline.set_auto_history() is not implemented, as in pyreadline3, will fall
back gracefully to automatic history saving. The only issue with this is that
-!history commands will be recorded in the history.
-!fetch on missing file no longer crashes script
-!history is now one of the autocomplete commands
-.dream_history now stored in output directory rather than ~user directory.
An important limitation of the last feature is that the history is
loaded and saved to the .dream_history file in the --outdir directory
specified at script launch time. It is not swapped around when the
--outdir is changed during the session.
- normalized how filenames are written out when postprocessing invoked
- various fixes of bugs encountered during testing
- updated documentation
- updated help text
- normalized how filenames are written out when postprocessing invoked
- various fixes of bugs encountered during testing
- updated documentation
- updated help text
Build the base generator in same place and way as other generators to reduce the chance of missed arguments in the future.
Fixes crash with display in-progress images, though note the feature still doesn't work for other reasons.
- Adapted from PR #489, author Dominic Letz [https://github.com/dominicletz]
- Too many upstream changes to merge, so frankensteined it in.
- Added support for !fix syntax
- Added documentation
- The seed printed needs to be the one generated prior to the
initial noising operation. To do this, I added a new "first_seed"
argument to the image callback in dream.py.
- Closes#641
- modify strength of embiggen to reduce tiling ghosts
- normalize naming of postprocessed files (could improve more to avoid
name collisions)
- move restoration modules under ldm.dream
- supports gfpgan, esrgan, codeformer and embiggen
- To use:
dream> !fix ./outputs/img-samples/000056.292144555.png -ft gfpgan -U2 -G0.8
dream> !fix ./outputs/img-samples/000056.292144555.png -ft codeformer -G 0.8
dream> !fix ./outputs/img-samples/000056.29214455.png -U4
dream> !fix ./outputs/img-samples/000056.292144555.png -embiggen 1.5
The first example invokes gfpgan to fix faces and esrgan to upscale.
The second example invokes codeformer to fix faces, no upscaling
The third example uses esrgan to upscale 4X
The four example runs embiggen to enlarge 1.5X
- This is very preliminary work. There are some anomalies to note:
1. The syntax is non-obvious. I would prefer something like:
!fix esrgan,gfpgan
!fix esrgan
!fix embiggen,codeformer
However, this will require refactoring the gfpgan and embiggen
code.
2. Images generated using gfpgan, esrgan or codeformer all are named
"xxxxxx.xxxxxx.postprocessed.png" and the original is saved.
However, the prefix is a new one that is not related to the
original.
3. Images generated using embiggen are named "xxxxx.xxxxxxx.png",
and once again the prefix is new. I'm not sure whether the
prefix should be aligned with the original file's prefix or not.
Probably not, but opinions welcome.
Allowed values are 'auto', 'float32', 'autocast', 'float16'. If not specified or 'auto' a working precision is automatically selected based on the torch device.
Context: #526
Deprecated --full_precision / -F
Tested on both cuda and cpu by calling scripts/dream.py without arguments and checked the auto configuration worked. With --precision=auto/float32/autocast/float16 it performs as expected, either working or failing with a reasonable error. Also checked Img2Img.
- modify strength of embiggen to reduce tiling ghosts
- normalize naming of postprocessed files (could improve more to avoid
name collisions)
- move restoration modules under ldm.dream
- supports gfpgan, esrgan, codeformer and embiggen
- To use:
dream> !fix ./outputs/img-samples/000056.292144555.png -ft gfpgan -U2 -G0.8
dream> !fix ./outputs/img-samples/000056.292144555.png -ft codeformer -G 0.8
dream> !fix ./outputs/img-samples/000056.29214455.png -U4
dream> !fix ./outputs/img-samples/000056.292144555.png -embiggen 1.5
The first example invokes gfpgan to fix faces and esrgan to upscale.
The second example invokes codeformer to fix faces, no upscaling
The third example uses esrgan to upscale 4X
The four example runs embiggen to enlarge 1.5X
- This is very preliminary work. There are some anomalies to note:
1. The syntax is non-obvious. I would prefer something like:
!fix esrgan,gfpgan
!fix esrgan
!fix embiggen,codeformer
However, this will require refactoring the gfpgan and embiggen
code.
2. Images generated using gfpgan, esrgan or codeformer all are named
"xxxxxx.xxxxxx.postprocessed.png" and the original is saved.
However, the prefix is a new one that is not related to the
original.
3. Images generated using embiggen are named "xxxxx.xxxxxxx.png",
and once again the prefix is new. I'm not sure whether the
prefix should be aligned with the original file's prefix or not.
Probably not, but opinions welcome.
* Support color correction for img2img and inpainting, avoiding the shift to magenta seen when running images through img2img repeatedly.
* Fix docs for color correction
* add --init_color to prompt reconstruction
* For best results, the --init_color option should point to the *very first* image used in the sequence of img2img operations. Otherwise color correction will skew towards cyan.
Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com>
* Feature complete for #266 with exception of several small deviations:
1. initial image and model weight hashes use full sha256 hash rather than first 8 digits
2. Initialization parameters for post-processing steps not provided
3. Uses top-level "images" tags for both a single image and a grid of images. This change was suggested in a comment.
* Added scripts/sd_metadata.py to retrieve and print metadata from PNG files
* New ldm.dream.args.Args class is a namespace like object which holds all defaults and can be modified during exection to hold current settings.
* Modified dream.py and server.py to accommodate Args class.
* Update dream.py. k_euler_a and k_dpm_2_a M1 fix
Make results reproducible (so runs with the same seed produce the same result).
Implements fix by @wbowling referenced in https://github.com/lstein/stable-diffusion/issues/397#issuecomment-1240679294
* Update dream.py. Remove import torch from dream.py
* generate.py: k_euler_a and k_dpm_2_a M1 fix#579
Co-authored-by: Lincoln Stein <lincoln.stein@gmail.com>
* revert inadvertent change of conda env name (#528)
* Refactor generate.py and dream.py
* config file path (models.yaml) is parsed inside Generate() to simplify
API
* Better handling of keyboard interrupts in file loading mode vs
interactive
* Removed oodles of unused variables.
* move nonfunctional inpainting out of the scripts directory
* fix ugly ddim tqdm formatting
When running on just cpu (intel), a call to torch.layer_norm would error with RuntimeError: expected scalar type BFloat16 but found Float
Fix buggy device handling in model.py.
Tested with scripts/dream.py --full_precision on just cpu on intel laptop. Works but slow at ~10s/it.
* Add Embiggen automation
* Make embiggen_tiles masking more intelligent and count from one (at least for the user), rewrite sections of Embiggen README, fix various typos throughout README
* drop duplicate log message
* This moves the call to half() before model.to(device) to avoid GPU
copy of full model. Improves speed and reduces memory usage dramatically
* This fix contributed by @mh-dm (Mihai)
* start refactoring -not yet functional
* first phase of refactor done - not sure weighted prompts working
* Second phase of refactoring. Everything mostly working.
* The refactoring has moved all the hard-core inference work into
ldm.dream.generator.*, where there are submodules for txt2img and
img2img. inpaint will go in there as well.
* Some additional refactoring will be done soon, but relatively
minor work.
* fix -save_orig flag to actually work
* add @neonsecret attention.py memory optimization
* remove unneeded imports
* move token logging into conditioning.py
* add placeholder version of inpaint; porting in progress
* fix crash in img2img
* inpainting working; not tested on variations
* fix crashes in img2img
* ported attention.py memory optimization #117 from basujindal branch
* added @torch_no_grad() decorators to img2img, txt2img, inpaint closures
* Final commit prior to PR against development
* fixup crash when generating intermediate images in web UI
* rename ldm.simplet2i to ldm.generate
* add backward-compatibility simplet2i shell with deprecation warning
* add back in mps exception, addresses @vargol comment in #354
* replaced Conditioning class with exported functions
* fix wrong type of with_variations attribute during intialization
* changed "image_iterator()" to "get_make_image()"
* raise NotImplementedError for calling get_make_image() in parent class
* Update ldm/generate.py
better error message
Co-authored-by: Kevin Gibbons <bakkot@gmail.com>
* minor stylistic fixes and assertion checks from code review
* moved get_noise() method into img2img class
* break get_noise() into two methods, one for txt2img and the other for img2img
* inpainting works on non-square images now
* make get_noise() an abstract method in base class
* much improved inpainting
Co-authored-by: Kevin Gibbons <bakkot@gmail.com>
