ba07cb748e
Use faster manual cast for fp8 in unet.
2023-12-11 18:24:44 -05:00
9ac0b487ac
Make --gpu-only put intermediate values in GPU memory instead of cpu.
2023-12-08 02:35:45 -05:00
983ebc5792
Use smart model management for VAE to decrease latency.
2023-11-28 04:58:51 -05:00
c45d1b9b67
Add a function to load a unet from a state dict.
2023-11-27 17:41:29 -05:00
871cc20e13
Support SVD img2vid model.
2023-11-23 19:41:33 -05:00
410bf07771
Make VAE memory estimation take dtype into account.
2023-11-22 18:17:19 -05:00
6a491ebe27
Allow model config to preprocess the vae state dict on load.
2023-11-21 16:29:18 -05:00
cd4fc77d5f
Add taesd and taesdxl to VAELoader node.
...
They will show up if both the taesd_encoder and taesd_decoder or taesdxl
model files are present in the models/vae_approx directory.
2023-11-21 12:54:19 -05:00
0cf4e86939
Add some command line arguments to store text encoder weights in fp8.
...
Pytorch supports two variants of fp8:
--fp8_e4m3fn-text-enc (the one that seems to give better results)
--fp8_e5m2-text-enc
2023-11-17 02:56:59 -05:00
0a6fd49a3e
Print leftover keys when using the UNETLoader.
2023-11-07 22:15:55 -05:00
2455aaed8a
Allow model or clip to be None in load_lora_for_models.
2023-11-01 20:27:20 -04:00
2a134bfab9
Fix checkpoint loader with config.
2023-10-27 22:13:55 -04:00
6ec3f12c6e
Support SSD1B model and make it easier to support asymmetric unets.
2023-10-27 14:45:15 -04:00
430a8334c5
Fix some potential issues.
2023-10-18 19:48:36 -04:00
d44a2de49f
Make VAE code closer to sgm.
2023-10-17 15:18:51 -04:00
9a55dadb4c
Refactor code so model can be a dtype other than fp32 or fp16.
2023-10-13 14:41:17 -04:00
72188dffc3
load_checkpoint_guess_config can now optionally output the model.
2023-10-06 13:48:18 -04:00
9bfec2bdbf
Fix quality loss due to low precision
2023-10-04 15:40:59 +02:00
71713888c4
Print missing VAE keys.
2023-09-28 00:54:57 -04:00
7d401ed1d0
Add ldm format support to UNETLoader.
2023-09-11 16:36:50 -04:00
7931ff0fd9
Support SDXL inpaint models.
2023-09-01 15:22:52 -04:00
fe4c07400c
Fix "Load Checkpoint with config" node.
2023-08-29 23:58:32 -04:00
1300a1bb4c
Text encoder should initially load on the offload_device not the regular.
2023-08-28 15:08:45 -04:00
f92074b84f
Move ModelPatcher to model_patcher.py
2023-08-28 14:51:31 -04:00
c77f02e1c6
Move controlnet code to comfy/controlnet.py
2023-08-25 17:33:04 -04:00
15a7716fa6
Move lora code to comfy/lora.py
2023-08-25 17:11:51 -04:00
ec96f6d03a
Move text_projection to base clip model.
2023-08-24 23:43:48 -04:00
51dde87e97
Try to free enough vram for control lora inference.
2023-08-24 17:20:54 -04:00
cc44ade79e
Always shift text encoder to GPU when the device supports fp16.
2023-08-23 21:45:00 -04:00
00c0b2c507
Initialize text encoder to target dtype.
2023-08-23 21:01:15 -04:00
f081017c1a
Save memory by storing text encoder weights in fp16 in most situations.
...
Do inference in fp32 to make sure quality stays the exact same.
2023-08-23 01:08:51 -04:00
afcb9cb1df
All resolutions now work with t2i adapter for SDXL.
2023-08-22 16:23:54 -04:00
85fde89d7f
T2I adapter SDXL.
2023-08-22 14:40:43 -04:00
cf5ae46928
Controlnet/t2iadapter cleanup.
2023-08-22 01:06:26 -04:00
763b0cf024
Fix control lora not working in fp32.
2023-08-21 20:38:31 -04:00
199d73364a
Fix ControlLora on lowvram.
2023-08-21 00:54:04 -04:00
d08e53de2e
Remove autocast from controlnet code.
2023-08-20 21:47:32 -04:00
c9b562aed1
Free more memory before VAE encode/decode.
2023-08-19 12:13:13 -04:00
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
c28db1f315
Fix potential issues with patching models when saving checkpoints.
2023-08-17 11:07:08 -04:00
89a0767abf
Smarter memory management.
...
Try to keep models on the vram when possible.
Better lowvram mode for controlnets.
2023-08-17 01:06:34 -04:00
53f326a3d8
Support diffusers mini controlnets.
2023-08-16 12:28:01 -04:00
c8a23ce9e8
Support for yet another lora type based on diffusers.
2023-08-11 13:04:21 -04:00
c20583286f
Support diffuser text encoder loras.
2023-08-10 20:28:28 -04:00
d8e58f0a7e
Detect hint_channels from controlnet.
2023-08-06 14:08:59 -04:00
c5d7593ccf
Support loras in diffusers format.
2023-08-05 01:40:24 -04:00
4b957a0010
Initialize the unet directly on the target device.
2023-07-29 14:51:56 -04:00
727588d076
Fix some new loras.
2023-07-25 16:39:15 -04:00
5f75d784a1
Start is now 0.0 and end is now 1.0 for the timestep ranges.
2023-07-24 18:38:17 -04:00
7ff14b62f8
ControlNetApplyAdvanced can now define when controlnet gets applied.
2023-07-24 17:50:49 -04:00
22f29d66ca
Try to fix memory issue with lora.
2023-07-22 21:38:56 -04:00
12a6e93171
Del the right object when applying lora.
2023-07-22 11:25:49 -04:00
78e7958d17
Support controlnet in diffusers format.
2023-07-21 22:58:16 -04:00
09386a3697
Fix issue with lora in some cases when combined with model merging.
2023-07-21 21:27:27 -04:00
58b2364f58
Properly support SDXL diffusers unet with UNETLoader node.
2023-07-21 14:38:56 -04:00
0115018695
Print errors and continue when lora weights are not compatible.
2023-07-20 19:56:22 -04:00
3ded1a3a04
Refactor of sampler code to deal more easily with different model types.
2023-07-17 01:22:12 -04:00
5f57362613
Lower lora ram usage when in normal vram mode.
2023-07-16 02:59:04 -04:00
490771b7f4
Speed up lora loading a bit.
2023-07-15 13:25:22 -04:00
50b1180dde
Fix CLIPSetLastLayer not reverting when removed.
2023-07-15 01:41:21 -04:00
6fb084f39d
Reduce floating point rounding errors in loras.
2023-07-15 00:53:00 -04:00
91ed2815d5
Add a node to merge CLIP models.
2023-07-14 02:41:18 -04:00
6ad0a6d7e2
Don't patch weights when multiplier is zero.
2023-07-09 17:46:56 -04:00
a9a4ba7574
Fix merging not working when model2 of model merge node was a merge.
2023-07-08 22:31:10 -04:00
e7bee85df8
Add arguments to run the VAE in fp16 or bf16 for testing.
2023-07-06 23:23:46 -04:00
ddc6f12ad5
Disable autocast in unet for increased speed.
2023-07-05 21:58:29 -04:00
af7a49916b
Support loading unet files in diffusers format.
2023-07-05 17:38:59 -04:00
acf95191ff
Properly support SDXL diffusers loras for unet.
2023-07-04 21:15:23 -04:00
c3e96e637d
Pass device to CLIP model.
2023-07-03 16:09:37 -04:00
2c4e0b49b7
Switch to fp16 on some cards when the model is too big.
2023-07-02 10:00:57 -04:00
1c1b0e7299
--gpu-only now keeps the VAE on the device.
2023-07-01 15:22:40 -04:00
3b6fe51c1d
Leave text_encoder on the CPU when it can handle it.
2023-07-01 14:38:51 -04:00
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.
2023-07-01 13:28:07 -04:00
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.
2023-07-01 12:37:23 -04:00
5a9ddf94eb
LoraLoader node now caches the lora file between executions.
2023-06-29 23:40:51 -04:00
62db11683b
Move unet to device right after loading on highvram mode.
2023-06-29 20:43:06 -04:00
2c7c14de56
Support for SDXL text encoder lora.
2023-06-28 02:22:49 -04:00
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.
2023-06-26 12:22:27 -04:00
b72a7a835a
Support loras based on the stability unet implementation.
2023-06-26 02:56:11 -04:00
20f579d91d
Add DualClipLoader to load clip models for SDXL.
...
Update LoadClip to load clip models for SDXL refiner.
2023-06-25 01:40:38 -04:00
b7933960bb
Fix CLIPLoader node.
2023-06-24 13:56:46 -04:00
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
8607c2d42d
Move latent scale factor from VAE to model.
2023-06-23 02:33:31 -04:00
30a3861946
Fix bug when yaml config has no clip params.
2023-06-23 01:12:59 -04:00
9e37f4c7d5
Fix error with ClipVision loader node.
2023-06-23 01:08:05 -04:00
9f83b098c9
Don't merge weights when shapes don't match and print a warning.
2023-06-22 19:08:31 -04:00
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
51581dbfa9
Fix last commits causing an issue with the text encoder lora.
2023-06-20 19:44:39 -04:00
8125b51a62
Keep a set of model_keys for faster add_patches.
2023-06-20 19:08:48 -04:00
45beebd33c
Add a type of model patch useful for model merging.
2023-06-20 17:34:11 -04:00
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
fb4bf7f591
This is not needed anymore and causes issues with alphas_cumprod.
2023-06-18 03:18:25 -04:00
f7edcfd927
Add a --gpu-only argument to keep and run everything on the GPU.
...
Make the CLIP model work on the GPU.
2023-06-15 15:38:52 -04:00
6b774589a5
Set model to fp16 before loading the state dict to lower ram bump.
2023-06-14 12:48:02 -04:00
388567f20b
sampler_cfg_function now uses a dict for the argument.
...
This means arguments can be added without issues.
2023-06-13 16:10:36 -04:00
ff9b22d79e
Turn on safe load for a few models.
2023-06-13 10:12:03 -04:00
f0a2b81cd0
Cleanup: Remove a bunch of useless files.
2023-06-13 02:19:08 -04:00
f8c5931053
Split the batch in VAEEncode if there's not enough memory.
2023-06-12 00:21:50 -04:00
c069fc0730
Auto switch to tiled VAE encode if regular one runs out of memory.
2023-06-11 23:25:39 -04:00
de142eaad5
Simpler base model code.
2023-06-09 12:31:16 -04:00
comfyanonymous
City