1a4bd9e9a6
Refactor the attention functions.
...
There's no reason for the whole CrossAttention object to be repeated when
only the operation in the middle changes.
1 year ago
fff491b032
Model patches can now know which batch is positive and negative.
1 year ago
afa2399f79
Add a way to set output block patches to modify the h and hsp.
1 year ago
1938f5c5fe
Add a force argument to soft_empty_cache to force a cache empty.
1 year ago
2da73b7073
Revert changes in comfy/ldm/modules/diffusionmodules/util.py, which is unused.
2 years ago
4a0c4ce4ef
Some fixes to generalize CUDA specific functionality to Intel or other GPUs.
2 years ago
0e3b641172
Remove xformers related print.
2 years ago
bed116a1f9
Remove optimization that caused border.
2 years ago
1c794a2161
Fallback to slice attention if xformers doesn't support the operation.
2 years ago
d935ba50c4
Make --bf16-vae work on torch 2.0
2 years ago
cf5ae46928
Controlnet/t2iadapter cleanup.
2 years ago
b80c3276dc
Fix issue with gligen.
2 years ago
d6e4b342e6
Support for Control Loras.
...
Control loras are controlnets where some of the weights are stored in
"lora" format: an up and a down low rank matrice that when multiplied
together and added to the unet weight give the controlnet weight.
This allows a much smaller memory footprint depending on the rank of the
matrices.
These controlnets are used just like regular ones.
2 years ago
2b13939044
Remove some useless code.
2 years ago
95d796fc85
Faster VAE loading.
2 years ago
4b957a0010
Initialize the unet directly on the target device.
2 years ago
9ba440995a
It's actually possible to torch.compile the unet now.
2 years ago
ddc6f12ad5
Disable autocast in unet for increased speed.
2 years ago
103c487a89
Cleanup.
2 years ago
78d8035f73
Fix bug with controlnet.
2 years ago
05676942b7
Add some more transformer hooks and move tomesd to comfy_extras.
...
Tomesd now uses q instead of x to decide which tokens to merge because
it seems to give better results.
2 years ago
fa28d7334b
Remove useless code.
2 years ago
f87ec10a97
Support base SDXL and SDXL refiner models.
...
Large refactor of the model detection and loading code.
2 years ago
9fccf4aa03
Add original_shape parameter to transformer patch extra_options.
2 years ago
8883cb0f67
Add a way to set patches that modify the attn2 output.
...
Change the transformer patches function format to be more future proof.
2 years ago
ae43f09ef7
All the unet weights should now be initialized with the right dtype.
2 years ago
7bf89ba923
Initialize more unet weights as the right dtype.
2 years ago
e21d9ad445
Initialize transformer unet block weights in right dtype at the start.
2 years ago
21f04fe632
Disable default weight values in unet conv2d for faster loading.
2 years ago
9d54066ebc
This isn't needed for inference.
2 years ago
6971646b8b
Speed up model loading a bit.
...
Default pytorch Linear initializes the weights which is useless and slow.
2 years ago
274dff3257
Remove more useless files.
2 years ago
f0a2b81cd0
Cleanup: Remove a bunch of useless files.
2 years ago
b8636a44aa
Make scaled_dot_product switch to sliced attention on OOM.
2 years ago
797c4e8d3b
Simplify and improve some vae attention code.
2 years ago
d9e088ddfd
minor changes for tiled sampler
2 years ago
cb1551b819
Lowvram mode for gligen and fix some lowvram issues.
2 years ago
bae4fb4a9d
Fix imports.
2 years ago
ba8a4c3667
Change latent resolution step to 8.
2 years ago
66c8aa5c3e
Make unet work with any input shape.
2 years ago
5282f56434
Implement Linear hypernetworks.
...
Add a HypernetworkLoader node to use hypernetworks.
2 years ago
6908f9c949
This makes pytorch2.0 attention perform a bit faster.
2 years ago
3696d1699a
Add support for GLIGEN textbox model.
2 years ago
73c3e11e83
Fix model_management import so it doesn't get executed twice.
2 years ago
e5e587b1c0
seperates out arg parser and imports args
2 years ago
e46b1c3034
Disable xformers in VAE when xformers == 0.0.18
2 years ago
539ff487a8
Pull latest tomesd code from upstream.
2 years ago
809bcc8ceb
Add support for unCLIP SD2.x models.
...
See _for_testing/unclip in the UI for the new nodes.
unCLIPCheckpointLoader is used to load them.
unCLIPConditioning is used to add the image cond and takes as input a
CLIPVisionEncode output which has been moved to the conditioning section.
2 years ago
0d972b85e6
This seems to give better quality in tome.
2 years ago
18a6c1db33
Add a TomePatchModel node to the _for_testing section.
...
Tome increases sampling speed at the expense of quality.
2 years ago
comfyanonymous
Simon Lui
BlenderNeko
EllangoK