0d7b0a4dc7
Small cleanups.
2 years ago
9225465975
Further tuning and fix mem_free_total.
2 years ago
2c096e4260
Add ipex optimize and other enhancements for Intel GPUs based on recent memory changes.
2 years ago
e9469e732d
--disable-smart-memory now disables loading model directly to vram.
2 years ago
c9b562aed1
Free more memory before VAE encode/decode.
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
39ac856a33
ReVision support: unclip nodes can now be used with SDXL.
2 years ago
76d53c4622
Add support for clip g vision model to CLIPVisionLoader.
2 years ago
e59fe0537a
Fix referenced before assignment
...
For https://github.com/BlenderNeko/ComfyUI_TiledKSampler/issues/13
2 years ago
be9c5e25bc
Fix issue with not freeing enough memory when sampling.
2 years ago
ac0758a1a4
Fix bug with lowvram and controlnet advanced node.
2 years ago
c28db1f315
Fix potential issues with patching models when saving checkpoints.
2 years ago
3aee33b54e
Add --disable-smart-memory for those that want the old behaviour.
2 years ago
2be2742711
Fix issue with regular torch version.
2 years ago
89a0767abf
Smarter memory management.
...
Try to keep models on the vram when possible.
Better lowvram mode for controlnets.
2 years ago
2c97c30256
Support small diffusers controlnet so both types are now supported.
2 years ago
53f326a3d8
Support diffusers mini controlnets.
2 years ago
58f0c616ed
Fix clip vision issue with old transformers versions.
2 years ago
ae270f79bc
Fix potential issue with batch size and clip vision.
2 years ago
a2ce9655ca
Refactor unclip code.
2 years ago
9cc12c833d
CLIPVisionEncode can now encode multiple images.
2 years ago
0cb6dac943
Remove 3m from PR #1213 because of some small issues.
2 years ago
e244b2df83
Add sgm_uniform scheduler that acts like the default one in sgm.
2 years ago
58c7da3665
Gpu variant of dpmpp_3m_sde. Note: use 3m with exponential or karras.
2 years ago
ba319a34e4
Merge branch 'dpmpp3m' of https://github.com/FizzleDorf/ComfyUI
2 years ago
3cfad03a68
dpmpp 3m + dpmpp 3m sde added
2 years ago
585a062910
Print unet config when model isn't detected.
2 years ago
c8a23ce9e8
Support for yet another lora type based on diffusers.
2 years ago
2bc12d3d22
Add --temp-directory argument to set temp directory.
2 years ago
c20583286f
Support diffuser text encoder loras.
2 years ago
cf10c5592c
Disable calculating uncond when CFG is 1.0
2 years ago
1f0f4cc0bd
Add argument to disable auto launching the browser.
2 years ago
d8e58f0a7e
Detect hint_channels from controlnet.
2 years ago
c5d7593ccf
Support loras in diffusers format.
2 years ago
1ce0d8ad68
Add CMP 30HX card to the nvidia_16_series list.
2 years ago
c99d8002f8
Make sure the pooled output stays at the EOS token with added embeddings.
2 years ago
4a77fcd6ab
Only shift text encoder to vram when CPU cores are under 8.
2 years ago
3cd31d0e24
Lower CPU thread check for running the text encoder on the CPU vs GPU.
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
c910b4a01c
Remove unused code and torchdiffeq dependency.
2 years ago
1141029a4a
Add --disable-metadata argument to disable saving metadata in files.
2 years ago
fbf5c51c1c
Merge branch 'fix_batch_timesteps' of https://github.com/asagi4/ComfyUI
2 years ago
68be24eead
Remove some prints.
2 years ago
1ea4d84691
Fix timestep ranges when batch_size > 1
2 years ago
5379051d16
Fix diffusers VAE loading.
2 years ago
727588d076
Fix some new loras.
2 years ago
4f9b6f39d1
Fix potential issue with Save Checkpoint.
2 years ago
5f75d784a1
Start is now 0.0 and end is now 1.0 for the timestep ranges.
2 years ago
7ff14b62f8
ControlNetApplyAdvanced can now define when controlnet gets applied.
2 years ago
d191c4f9ed
Add a ControlNetApplyAdvanced node.
...
The controlnet can be applied to the positive or negative prompt only by
connecting it correctly.
2 years ago
0240946ecf
Add a way to set which range of timesteps the cond gets applied to.
2 years ago
22f29d66ca
Try to fix memory issue with lora.
2 years ago
67be7eb81d
Nodes can now patch the unet function.
2 years ago
12a6e93171
Del the right object when applying lora.
2 years ago
78e7958d17
Support controlnet in diffusers format.
2 years ago
09386a3697
Fix issue with lora in some cases when combined with model merging.
2 years ago
58b2364f58
Properly support SDXL diffusers unet with UNETLoader node.
2 years ago
0115018695
Print errors and continue when lora weights are not compatible.
2 years ago
4760c29380
Merge branch 'fix-AttributeError-module-'torch'-has-no-attribute-'mps'' of https://github.com/KarryCharon/ComfyUI
2 years ago
0b284f650b
Fix typo.
2 years ago
e032ca6138
Fix ddim issue with older torch versions.
2 years ago
18885f803a
Add MX450 and MX550 to list of cards with broken fp16.
2 years ago
9ba440995a
It's actually possible to torch.compile the unet now.
2 years ago
51d5477579
Add key to indicate checkpoint is v_prediction when saving.
2 years ago
ff6b047a74
Fix device print on old torch version.
2 years ago
9871a15cf9
Enable --cuda-malloc by default on torch 2.0 and up.
...
Add --disable-cuda-malloc to disable it.
2 years ago
55d0fca9fa
--windows-standalone-build now enables --cuda-malloc
2 years ago
1679abd86d
Add a command line argument to enable backend:cudaMallocAsync
2 years ago
3a150bad15
Only calculate randn in some samplers when it's actually being used.
2 years ago
ee8f8ee07f
Fix regression with ddim and uni_pc when batch size > 1.
2 years ago
3ded1a3a04
Refactor of sampler code to deal more easily with different model types.
2 years ago
5f57362613
Lower lora ram usage when in normal vram mode.
2 years ago
490771b7f4
Speed up lora loading a bit.
2 years ago
50b1180dde
Fix CLIPSetLastLayer not reverting when removed.
2 years ago
6fb084f39d
Reduce floating point rounding errors in loras.
2 years ago
91ed2815d5
Add a node to merge CLIP models.
2 years ago
b2f03164c7
Prevent the clip_g position_ids key from being saved in the checkpoint.
...
This is to make it match the official checkpoint.
2 years ago
46dc050c9f
Fix potential tensors being on different devices issues.
2 years ago
3e2309f149
fix mps miss import
2 years ago
606a537090
Support SDXL embedding format with 2 CLIP.
2 years ago
6ad0a6d7e2
Don't patch weights when multiplier is zero.
2 years ago
d5323d16e0
latent2rgb matrix for SDXL.
2 years ago
0ae81c03bb
Empty cache after model unloading for normal vram and lower.
2 years ago
d3f5998218
Support loading clip_g from diffusers in CLIP Loader nodes.
2 years ago
a9a4ba7574
Fix merging not working when model2 of model merge node was a merge.
2 years ago
bb5fbd29e9
Merge branch 'condmask-fix' of https://github.com/vmedea/ComfyUI
2 years ago
e7bee85df8
Add arguments to run the VAE in fp16 or bf16 for testing.
2 years ago
608fcc2591
Fix bug with weights when prompt is long.
2 years ago
ddc6f12ad5
Disable autocast in unet for increased speed.
2 years ago
603f02d613
Fix loras not working when loading checkpoint with config.
2 years ago
af7a49916b
Support loading unet files in diffusers format.
2 years ago
e57cba4c61
Add gpu variations of the sde samplers that are less deterministic
...
but faster.
2 years ago
f81b192944
Add logit scale parameter so it's present when saving the checkpoint.
2 years ago
acf95191ff
Properly support SDXL diffusers loras for unet.
2 years ago
c61a95f9f7
Fix size check for conditioning mask
...
The wrong dimensions were being checked, [1] and [2] are the image size.
not [2] and [3]. This results in an out-of-bounds error if one of them
actually matches.
2 years ago
8d694cc450
Fix issue with OSX.
2 years ago
c3e96e637d
Pass device to CLIP model.
2 years ago
5e6bc824aa
Allow passing custom path to clip-g and clip-h.
2 years ago
dc9d1f31c8
Improvements for OSX.
2 years ago
103c487a89
Cleanup.
2 years ago
2c4e0b49b7
Switch to fp16 on some cards when the model is too big.
2 years ago
6f3d9f52db
Add a --force-fp16 argument to force fp16 for testing.
2 years ago
1c1b0e7299
--gpu-only now keeps the VAE on the device.
2 years ago
ce35d8c659
Lower latency by batching some text encoder inputs.
2 years ago
3b6fe51c1d
Leave text_encoder on the CPU when it can handle it.
2 years ago
b6a60fa696
Try to keep text encoders loaded and patched to increase speed.
...
load_model_gpu() is now used with the text encoder models instead of just
the unet.
2 years ago
97ee230682
Make highvram and normalvram shift the text encoders to vram and back.
...
This is faster on big text encoder models than running it on the CPU.
2 years ago
5a9ddf94eb
LoraLoader node now caches the lora file between executions.
2 years ago
9920367d3c
Fix embeddings not working with --gpu-only
2 years ago
62db11683b
Move unet to device right after loading on highvram mode.
2 years ago
4376b125eb
Remove useless code.
2 years ago
89120f1fbe
This is unused but it should be 1280.
2 years ago
2c7c14de56
Support for SDXL text encoder lora.
2 years ago
fcef47f06e
Fix bug.
2 years ago
8248babd44
Use pytorch attention by default on nvidia when xformers isn't present.
...
Add a new argument --use-quad-cross-attention
2 years ago
9b93b920be
Add CheckpointSave node to save checkpoints.
...
The created checkpoints contain workflow metadata that can be loaded by
dragging them on top of the UI or loading them with the "Load" button.
Checkpoints will be saved in fp16 or fp32 depending on the format ComfyUI
is using for inference on your hardware. To force fp32 use: --force-fp32
Anything that patches the model weights like merging or loras will be
saved.
The output directory is currently set to: output/checkpoints but that might
change in the future.
2 years ago
b72a7a835a
Support loras based on the stability unet implementation.
2 years ago
c71a7e6b20
Fix ddim + inpainting not working.
2 years ago
4eab00e14b
Set the seed in the SDE samplers to make them more reproducible.
2 years ago
cef6aa62b2
Add support for TAESD decoder for SDXL.
2 years ago
20f579d91d
Add DualClipLoader to load clip models for SDXL.
...
Update LoadClip to load clip models for SDXL refiner.
2 years ago
b7933960bb
Fix CLIPLoader node.
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
8607c2d42d
Move latent scale factor from VAE to model.
2 years ago
30a3861946
Fix bug when yaml config has no clip params.
2 years ago
9e37f4c7d5
Fix error with ClipVision loader node.
2 years ago
9f83b098c9
Don't merge weights when shapes don't match and print a warning.
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
51581dbfa9
Fix last commits causing an issue with the text encoder lora.
2 years ago
8125b51a62
Keep a set of model_keys for faster add_patches.
2 years ago
45beebd33c
Add a type of model patch useful for model merging.
2 years ago
036a22077c
Fix k_diffusion math being off by a tiny bit during txt2img.
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
cd930d4e7f
pop clip vision keys after loading them.
2 years ago
c9e4a8c9e5
Not needed anymore.
2 years ago
fb4bf7f591
This is not needed anymore and causes issues with alphas_cumprod.
2 years ago
45be2e92c1
Fix DDIM v-prediction.
2 years ago
e6e50ab2dd
Fix an issue when alphas_comprod are half floats.
2 years ago
ae43f09ef7
All the unet weights should now be initialized with the right dtype.
2 years ago
f7edcfd927
Add a --gpu-only argument to keep and run everything on the GPU.
...
Make the CLIP model work on the GPU.
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
bb1f45d6e8
Properly disable weight initialization in clip models.
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
fa2cca056c
Don't initialize CLIPVision weights to default values.
2 years ago
6b774589a5
Set model to fp16 before loading the state dict to lower ram bump.
2 years ago
0c7cad404c
Don't initialize clip weights to default values.
2 years ago
6971646b8b
Speed up model loading a bit.
...
Default pytorch Linear initializes the weights which is useless and slow.
2 years ago
388567f20b
sampler_cfg_function now uses a dict for the argument.
...
This means arguments can be added without issues.
2 years ago
ff9b22d79e
Turn on safe load for a few models.
2 years ago
735ac4cf81
Remove pytorch_lightning dependency.
2 years ago
2b14041d4b
Remove useless code.
2 years ago
274dff3257
Remove more useless files.
2 years ago
f0a2b81cd0
Cleanup: Remove a bunch of useless files.
2 years ago
f8c5931053
Split the batch in VAEEncode if there's not enough memory.
2 years ago
c069fc0730
Auto switch to tiled VAE encode if regular one runs out of memory.
2 years ago
c64ca8c0b2
Refactor unCLIP noise augment out of samplers.py
2 years ago
de142eaad5
Simpler base model code.
2 years ago
23cf8ca7c5
Fix bug when embedding gets ignored because of mismatched size.
2 years ago
0e425603fb
Small refactor.
2 years ago
a3a713b6c5
Refactor previews into one command line argument.
...
Clean up a few things.
2 years ago
3e17971acb
preview method autodetection
2 years ago
d5a28fadaa
Add latent2rgb preview
2 years ago
48f7ec750c
Make previews into cli option
2 years ago
b4f434ee66
Preview sampled images with TAESD
2 years ago
fed0a4dd29
Some comments to say what the vram state options mean.
2 years ago
0a5fefd621
Cleanups and fixes for model_management.py
...
Hopefully fix regression on MPS and CPU.
2 years ago
700491d81a
Implement global average pooling for controlnet.
2 years ago
67892b5ac5
Refactor and improve model_management code related to free memory.
2 years ago
499641ebf1
More accurate total
2 years ago
b5dd15c67a
System stats endpoint
2 years ago
5c38958e49
Tweak lowvram model memory so it's closer to what it was before.
2 years ago
94680732d3
Empty cache on mps.
2 years ago
03da8a3426
This is useless for inference.
2 years ago
eb448dd8e1
Auto load model in lowvram if not enough memory.
2 years ago
b9818eb910
Add route to get safetensors metadata:
...
/view_metadata/loras?filename=lora.safetensors
2 years ago
a532888846
Support VAEs in diffusers format.
2 years ago
0fc483dcfd
Refactor diffusers model convert code to be able to reuse it.
2 years ago
eb4bd7711a
Remove einops.
2 years ago
87ab25fac7
Do operations in same order as the one it replaces.
2 years ago
2b1fac9708
Merge branch 'master' of https://github.com/BlenderNeko/ComfyUI
2 years ago
e1278fa925
Support old pytorch versions that don't have weights_only.
2 years ago
8b4b0c3188
vecorized bislerp
2 years ago
b8ccbec6d8
Various improvements to bislerp.
2 years ago
34887b8885
Add experimental bislerp algorithm for latent upscaling.
...
It's like bilinear but with slerp.
2 years ago
6cc450579b
Auto transpose images from exif data.
2 years ago
dc198650c0
sample_dpmpp_2m_sde no longer crashes when step == 1.
2 years ago
069657fbf3
Add DPM-Solver++(2M) SDE and exponential scheduler.
...
exponential scheduler is the one recommended with this sampler.
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
ef815ba1e2
Switch default scheduler to normal.
2 years ago
68d12b530e
Merge branch 'tiled_sampler' of https://github.com/BlenderNeko/ComfyUI
2 years ago
3a1f47764d
Print the torch device that is used on startup.
2 years ago
1201d2eae5
Make nodes map over input lists ( #579 )
...
* allow nodes to map over lists
* make work with IS_CHANGED and VALIDATE_INPUTS
* give list outputs distinct socket shape
* add rebatch node
* add batch index logic
* add repeat latent batch
* deal with noise mask edge cases in latentfrombatch
2 years ago
19c014f429
comment out annoying print statement
2 years ago
d9e088ddfd
minor changes for tiled sampler
2 years ago
f7c0f75d1f
Auto batching improvements.
...
Try batching when cond sizes don't match with smart padding.
2 years ago
314e526c5c
Not needed anymore because sampling works with any latent size.
2 years ago
c6e34963e4
Make t2i adapter work with any latent resolution.
2 years ago
a1f12e370d
Merge branch 'autostart' of https://github.com/EllangoK/ComfyUI
2 years ago
6fc4917634
Make maximum_batch_area take into account python2.0 attention function.
...
More conservative xformers maximum_batch_area.
2 years ago
678f933d38
maximum_batch_area for xformers.
...
Remove useless code.
2 years ago
8e03c789a2
auto-launch cli arg
2 years ago
cb1551b819
Lowvram mode for gligen and fix some lowvram issues.
2 years ago
af9cc1fb6a
Search recursively in subfolders for embeddings.
2 years ago
6ee11d7bc0
Fix import.
2 years ago
bae4fb4a9d
Fix imports.
2 years ago
fcf513e0b6
Refactor.
2 years ago
a74e176a24
Merge branch 'tiled-progress' of https://github.com/pythongosssss/ComfyUI
2 years ago
5eeecf3fd5
remove unused import
2 years ago
8912623ea9
use comfy progress bar
2 years ago
908dc1d5a8
Add a total_steps value to sampler callback.
2 years ago
fdf57325f4
Merge remote-tracking branch 'origin/master' into tiled-progress
2 years ago
27df74101e
reduce duplication
2 years ago
93c64afaa9
Use sampler callback instead of tqdm hook for progress bar.
2 years ago
06ad35b493
added progress to encode + upscale
2 years ago
ba8a4c3667
Change latent resolution step to 8.
2 years ago
66c8aa5c3e
Make unet work with any input shape.
2 years ago
9c335a553f
LoKR support.
2 years ago
d3293c8339
Properly disable all progress bars when disable_pbar=True
2 years ago
a2e18b1504
allow disabling of progress bar when sampling
2 years ago
071011aebe
Mask strength should be separate from area strength.
2 years ago
870fae62e7
Merge branch 'condition_by_mask_node' of https://github.com/guill/ComfyUI
2 years ago
af02393c2a
Default to sampling entire image
...
By default, when applying a mask to a condition, the entire image will
still be used for sampling. The new "set_area_to_bounds" option on the
node will allow the user to automatically limit conditioning to the
bounds of the mask.
I've also removed the dependency on torchvision for calculating bounding
boxes. I've taken the opportunity to fix some frustrating details in the
other version:
1. An all-0 mask will no longer cause an error
2. Indices are returned as integers instead of floats so they can be
used to index into tensors.
2 years ago
056e5545ff
Don't try to get vram from xpu or cuda when directml is enabled.
2 years ago
2ca934f7d4
You can now select the device index with: --directml id
...
Like this for example: --directml 1
2 years ago
3baded9892
Basic torch_directml support. Use --directml to use it.
2 years ago
e214c917ae
Add Condition by Mask node
...
This PR adds support for a Condition by Mask node. This node allows
conditioning to be limited to a non-rectangle area.
2 years ago
5a971cecdb
Add callback to sampler function.
...
Callback format is: callback(step, x0, x)
2 years ago
aa57136dae
Some fixes to the batch masks PR.
2 years ago
c50208a703
Refactor more code to sample.py
2 years ago
7983b3a975
This is cleaner this way.
2 years ago
0b07b2cc0f
gligen tuple
2 years ago
c8c9926eeb
Add progress to vae decode tiled
2 years ago
d9b1595f85
made sample functions more explicit
2 years ago
5818539743
add docstrings
2 years ago
8d2de420d3
Merge branch 'master' of https://github.com/BlenderNeko/ComfyUI
2 years ago
2a09e2aa27
refactor/split various bits of code for sampling
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
907010e082
Remove some useless code.
2 years ago
96b57a9ad6
Don't pass adm to model when it doesn't support it.
2 years ago
3696d1699a
Add support for GLIGEN textbox model.
2 years ago
comfyanonymous
Simon Lui
Alexopus
FizzleDorf
asagi4
KarryCharon
mara
space-nuko
BlenderNeko
EllangoK
pythongosssss
Jacob Segal