comfyanonymous
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.
2023-10-11 20:38:48 -04:00
comfyanonymous
fff491b032
Model patches can now know which batch is positive and negative.
2023-09-27 12:04:07 -04:00
comfyanonymous
afa2399f79
Add a way to set output block patches to modify the h and hsp.
2023-09-22 20:26:47 -04:00
comfyanonymous
1938f5c5fe
Add a force argument to soft_empty_cache to force a cache empty.
2023-09-04 00:58:18 -04:00
Simon Lui
2da73b7073
Revert changes in comfy/ldm/modules/diffusionmodules/util.py, which is unused.
2023-09-02 20:07:52 -07:00
Simon Lui
4a0c4ce4ef
Some fixes to generalize CUDA specific functionality to Intel or other GPUs.
2023-09-02 18:22:10 -07:00
comfyanonymous
0e3b641172
Remove xformers related print.
2023-09-01 02:12:03 -04:00
comfyanonymous
bed116a1f9
Remove optimization that caused border.
2023-08-29 11:21:36 -04:00
comfyanonymous
1c794a2161
Fallback to slice attention if xformers doesn't support the operation.
2023-08-27 22:24:42 -04:00
comfyanonymous
d935ba50c4
Make --bf16-vae work on torch 2.0
2023-08-27 21:33:53 -04:00
comfyanonymous
cf5ae46928
Controlnet/t2iadapter cleanup.
2023-08-22 01:06:26 -04:00
comfyanonymous
b80c3276dc
Fix issue with gligen.
2023-08-18 16:32:23 -04:00
comfyanonymous
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.
2023-08-18 11:59:51 -04:00
comfyanonymous
2b13939044
Remove some useless code.
2023-07-30 14:13:33 -04:00
comfyanonymous
95d796fc85
Faster VAE loading.
2023-07-29 16:28:30 -04:00
comfyanonymous
4b957a0010
Initialize the unet directly on the target device.
2023-07-29 14:51:56 -04:00
comfyanonymous
9ba440995a
It's actually possible to torch.compile the unet now.
2023-07-18 21:36:35 -04:00
comfyanonymous
ddc6f12ad5
Disable autocast in unet for increased speed.
2023-07-05 21:58:29 -04:00
comfyanonymous
103c487a89
Cleanup.
2023-07-02 11:58:23 -04:00
comfyanonymous
78d8035f73
Fix bug with controlnet.
2023-06-24 11:02:38 -04:00
comfyanonymous
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.
2023-06-24 03:30:22 -04:00
comfyanonymous
fa28d7334b
Remove useless code.
2023-06-23 12:35:26 -04:00
comfyanonymous
f87ec10a97
Support base SDXL and SDXL refiner models.
...
Large refactor of the model detection and loading code.
2023-06-22 13:03:50 -04:00
comfyanonymous
9fccf4aa03
Add original_shape parameter to transformer patch extra_options.
2023-06-21 13:22:01 -04:00
comfyanonymous
8883cb0f67
Add a way to set patches that modify the attn2 output.
...
Change the transformer patches function format to be more future proof.
2023-06-18 22:58:22 -04:00
comfyanonymous
ae43f09ef7
All the unet weights should now be initialized with the right dtype.
2023-06-15 18:42:30 -04:00
comfyanonymous
7bf89ba923
Initialize more unet weights as the right dtype.
2023-06-15 15:00:10 -04:00
comfyanonymous
e21d9ad445
Initialize transformer unet block weights in right dtype at the start.
2023-06-15 14:29:26 -04:00
comfyanonymous
21f04fe632
Disable default weight values in unet conv2d for faster loading.
2023-06-14 19:46:08 -04:00
comfyanonymous
9d54066ebc
This isn't needed for inference.
2023-06-14 13:05:08 -04:00
comfyanonymous
6971646b8b
Speed up model loading a bit.
...
Default pytorch Linear initializes the weights which is useless and slow.
2023-06-14 12:09:41 -04:00
comfyanonymous
274dff3257
Remove more useless files.
2023-06-13 02:22:19 -04:00
comfyanonymous
f0a2b81cd0
Cleanup: Remove a bunch of useless files.
2023-06-13 02:19:08 -04:00
comfyanonymous
b8636a44aa
Make scaled_dot_product switch to sliced attention on OOM.
2023-05-20 16:01:02 -04:00
comfyanonymous
797c4e8d3b
Simplify and improve some vae attention code.
2023-05-20 15:07:21 -04:00
BlenderNeko
d9e088ddfd
minor changes for tiled sampler
2023-05-12 23:49:09 +02:00
comfyanonymous
cb1551b819
Lowvram mode for gligen and fix some lowvram issues.
2023-05-05 18:11:41 -04:00
comfyanonymous
bae4fb4a9d
Fix imports.
2023-05-04 18:10:29 -04:00
comfyanonymous
ba8a4c3667
Change latent resolution step to 8.
2023-05-02 14:17:51 -04:00
comfyanonymous
66c8aa5c3e
Make unet work with any input shape.
2023-05-02 13:31:43 -04:00
comfyanonymous
5282f56434
Implement Linear hypernetworks.
...
Add a HypernetworkLoader node to use hypernetworks.
2023-04-23 12:35:25 -04:00
comfyanonymous
6908f9c949
This makes pytorch2.0 attention perform a bit faster.
2023-04-22 14:30:39 -04:00
comfyanonymous
3696d1699a
Add support for GLIGEN textbox model.
2023-04-19 11:06:32 -04:00
comfyanonymous
73c3e11e83
Fix model_management import so it doesn't get executed twice.
2023-04-15 19:04:33 -04:00
EllangoK
e5e587b1c0
seperates out arg parser and imports args
2023-04-05 23:41:23 -04:00
comfyanonymous
e46b1c3034
Disable xformers in VAE when xformers == 0.0.18
2023-04-04 22:22:02 -04:00
comfyanonymous
539ff487a8
Pull latest tomesd code from upstream.
2023-04-03 15:49:28 -04:00
comfyanonymous
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.
2023-04-01 23:19:15 -04:00
comfyanonymous
0d972b85e6
This seems to give better quality in tome.
2023-03-31 18:36:18 -04:00
comfyanonymous
18a6c1db33
Add a TomePatchModel node to the _for_testing section.
...
Tome increases sampling speed at the expense of quality.
2023-03-31 17:19:58 -04:00
