Dave Lage
336dd6a730
Merge 8b0a467bc0 into 2e0fcc50cb
2025-07-11 09:58:40 +08:00
Kohya S.
3e6935a07e
Merge pull request #2115 from kohya-ss/fix-flux-sampling-accelerate-error
...
Fix unwrap_model handling for None text_encoders in sample_images
2025-06-15 21:14:09 +09:00
Kohya S
fc40a279fa
Merge branch 'dev' into sd3
2025-06-15 21:05:57 +09:00
Disty0
bcd3a5a60a
Update IPEX libs
2025-06-13 16:25:16 +03:00
rockerBOO
7c83ac4369
Add avg non-zero ratio metric
2025-06-10 13:17:04 -04:00
Kohya S
bb47f1ea89
Fix unwrap_model handling for None text_encoders in sample_images function
2025-06-08 18:00:24 +09:00
rockerBOO
9629853d15
Fix wavelet loss not separating levels. Refactor loss to be spatial
2025-06-05 22:03:52 -04:00
rockerBOO
0af0302c38
Metrics, energy, loss
2025-05-19 19:15:23 -04:00
rockerBOO
346790a996
Merge branch 'sd3' into network-wavelet-loss
2025-05-19 19:10:55 -04:00
rockerBOO
984472ca09
Fix metrics
2025-05-04 18:58:54 -04:00
rockerBOO
7be3c5dce1
Refactor SWT to work properly and faster. Add SWT tests
2025-05-04 18:57:53 -04:00
rockerBOO
964bfcb576
Fix padding for small latents. Add DWT tests
2025-05-04 18:57:42 -04:00
rockerBOO
d5f8f7de1f
Add wavelet loss fn
2025-05-04 18:57:10 -04:00
rockerBOO
56dfdae7c5
Add QuaternionWaveletTransform. Update WaveletLoss
2025-05-04 18:55:50 -04:00
rockerBOO
19ce0ae61f
Add wavelet_loss_band_level_weights
2025-05-04 18:39:51 -04:00
rockerBOO
7b9e92a8cc
Fix band weights via toml. Add more logging
2025-05-04 18:39:48 -04:00
rockerBOO
20a99771bf
Add back in padding
2025-05-04 18:39:45 -04:00
rockerBOO
f553b7bf31
Add wavelet loss recording
2025-05-04 18:39:36 -04:00
rockerBOO
6d42b95e2b
Refactor transforms, fix loss calculations
...
- add full conditional_loss functionality to wavelet loss
- Transforms are separate and abstracted
- Loss now doesn't include LL except the lowest level
- ll_level_threshold allows you to control the level the ll is
used in the loss
- band weights can now be passed in
- rectified flow calculations can be bypassed for experimentation
- Fixed alpha to 1.0 with new weighted bands producing lower loss
2025-05-04 18:39:32 -04:00
rockerBOO
64422ff4a0
Suggest the right module
2025-05-04 18:39:25 -04:00
rockerBOO
837231a5c7
Add wavelet loss
2025-05-04 18:39:17 -04:00
rockerBOO
813942a967
Add wavelet loss
2025-05-04 18:38:47 -04:00
Kohya S.
7c075a9c8d
Merge pull request #2060 from saibit-tech/sd3
...
Fix: try aligning dtype of matrixes when training with deepspeed and mixed-precision is set to bf16 or fp16
2025-05-01 23:20:17 +09:00
Kohya S
64430eb9b2
Merge branch 'dev' into sd3
2025-04-29 21:30:57 +09:00
Kohya S
d8717a3d1c
Merge branch 'main' into dev
2025-04-29 21:30:33 +09:00
Kohya S
4625b34f4e
Fix mean image aspect ratio error calculation to avoid NaN values
2025-04-29 21:27:04 +09:00
Kohya S
fd3a445769
fix: revert default emb guidance scale and CFG scale for FLUX.1 sampling
2025-04-27 22:50:27 +09:00
saibit
46ad3be059
update deepspeed wrapper
2025-04-24 11:26:36 +08:00
sharlynxy
abf2c44bc5
Dynamically set device in deepspeed wrapper ( #2 )
...
* get device type from model
* add logger warning
* format
* format
* format
2025-04-23 18:57:19 +08:00
Robert
f501209c37
Merge branch 'dev/xy/align_dtype_using_mixed_precision' of github.com:saibit-tech/sd-scripts into dev/xy/align_dtype_using_mixed_precision
2025-04-22 16:19:52 +08:00
Robert
c8af252a44
refactor
2025-04-22 16:19:14 +08:00
saibit
7f984f4775
#
2025-04-22 16:15:12 +08:00
saibit
d33d5eccd1
#
2025-04-22 16:12:06 +08:00
saibit
7c61c0dfe0
Add autocast warpper for forward functions in deepspeed_utils.py to try aligning precision when using mixed precision in training process
2025-04-22 16:06:55 +08:00
Kohya S
629073cd9d
Add guidance scale for prompt param and flux sampling
2025-04-16 21:50:36 +09:00
Kohya S
06df0377f9
Merge branch 'sd3' into flux-sample-cfg
2025-04-16 21:27:08 +09:00
Kohya S.
c56dc90b26
Merge pull request #1992 from rockerBOO/flux-ip-noise-gamma
...
Add IP noise gamma for Flux
2025-04-06 21:29:26 +09:00
Kohya S.
606e6875d2
Merge pull request #2022 from LexSong/fix-resize-issue
...
Fix size parameter types and improve resize_image interpolation
2025-04-05 19:28:25 +09:00
Kohya S
f1423a7229
fix: add resize_interpolation parameter to FineTuningDataset constructor
2025-04-03 21:48:51 +09:00
Lex Song
b822b7e60b
Fix the interpolation logic error in resize_image()
...
The original code had a mistake. It used 'lanczos' when the image got smaller (width > resized_width and height > resized_height) and 'area' when it stayed the same or got bigger. This was the wrong way. 'area' is better for big shrinking.
2025-04-02 22:04:37 +08:00
Lex Song
ede3470260
Ensure all size parameters are integers to prevent type errors
2025-04-02 03:50:33 +08:00
Kohya S
b3c56b22bd
Merge branch 'dev' into sd3
2025-03-31 22:05:40 +09:00
Kohya S
583ab27b3c
doc: update license information in jpeg_xl_util.py
2025-03-31 22:02:25 +09:00
Kohya S
1f432e2c0e
use PIL for lanczos and box
2025-03-30 20:40:29 +09:00
Kohya S.
93a4efabb5
Merge branch 'sd3' into resize-interpolation
2025-03-30 19:30:56 +09:00
rockerBOO
e8b3254858
Add flux_train_utils tests for get get_noisy_model_input_and_timesteps
2025-03-20 15:01:15 -04:00
rockerBOO
16cef81aea
Refactor sigmas and timesteps
2025-03-20 14:32:56 -04:00
rockerBOO
f974c6b257
change order to match upstream
2025-03-19 14:27:43 -04:00
rockerBOO
5d5a7d2acf
Fix IP noise calculation
2025-03-19 13:50:04 -04:00
rockerBOO
1eddac26b0
Separate random to a variable, and make sure on device
2025-03-19 00:49:42 -04:00
