Commit Graph

79 Commits

Author SHA1 Message Date
Jaret Burkett
93ea955d7c Added refiner fine tuning. Works, but needs some polish. 2023-11-05 17:15:03 -07:00
Jaret Burkett
8a9e8f708f Added base for using guidance during training. Still not working right. 2023-11-05 04:03:32 -07:00
Jaret Burkett
d35733ac06 Added support for training ssd-1B. Added support for saving models into diffusers format. We can currently save in safetensors format for ssd-1b, but diffusers cannot load it yet. 2023-11-03 05:01:16 -06:00
Jaret Burkett
a899ec91c8 Added some split prompting started code, adamw8bit, replacements improving, learnable snr gos. A lot of good stuff. 2023-11-01 06:52:21 -06:00
Jaret Burkett
436a09430e Added flat snr gamma vs min. Fixes timestep timing 2023-10-29 15:41:55 -06:00
Jaret Burkett
298001439a Added gradient accumulation finally 2023-10-28 13:14:29 -06:00
Jaret Burkett
6f3e0d5af2 Improved lorm extraction and training 2023-10-28 08:21:59 -06:00
Jaret Burkett
0a79ac9604 Added lorm. WIP 2023-10-26 18:23:51 -06:00
Jaret Burkett
002279cec3 Allow short and long caption combinations like form the new captioning system. Merge the network into the model before inference and reextract when done. Doubles inference speed on locon models during inference. allow splitting a batch into individual components and run them through alone. Basicallt gradient accumulation with single batch size. 2023-10-24 16:02:07 -06:00
Jaret Burkett
07bf7bd7de Allow augmentations and targeting different loss types fron the config file 2023-10-18 03:04:57 -06:00
Jaret Burkett
da6302ada8 added a method to apply multipliers to noise and latents prior to combining 2023-10-17 06:09:16 -06:00
Jaret Burkett
38e441a29c allow flipping for point of interesting autocropping. allow num repeats. Fixed some bugs with new free u 2023-10-12 21:02:47 -06:00
Jaret Burkett
63ceffae24 Massive speed increases and ram optimizations 2023-10-10 06:07:55 -06:00
Jaret Burkett
1d3de678aa fixed bug with trigger word embedding. Allow control images to load from the dataloader or legacy way 2023-10-09 06:21:49 -06:00
Jaret Burkett
cac8754399 Allow for training loras on onle one text encoder for sdxl 2023-10-06 08:11:56 -06:00
Jaret Burkett
f73402473b Bug fixes. Added some functionality to help with private extensions 2023-10-05 07:09:34 -06:00
Jaret Burkett
579650eaf8 Fixed big issue with bucketing dataloader and added random cripping to a point of interest 2023-10-02 18:31:08 -06:00
Jaret Burkett
8d9450ad7c Compatability fixes 2023-09-29 14:07:37 -06:00
Jaret Burkett
abf7cd221d allow setting adapter weight in prompts 2023-09-24 06:51:54 -06:00
Jaret Burkett
e5153d87c9 Fixed issues with dataloader bucketing. Allow using standard base image for t2i adapters. 2023-09-24 05:19:57 -06:00
Jaret Burkett
830e87cb87 Added IP adapter training. Not functioning correctly yet 2023-09-24 02:39:43 -06:00
Jaret Burkett
19255cdc7c Bugfixes. Added small augmentations to dataloader. Will switch to abluminations soon though. Added ability to adjust step count on start to override what is in the file 2023-09-20 05:30:10 -06:00
Jaret Burkett
0f105690cc Added some further extendability for plugins 2023-09-19 05:41:44 -06:00
Jaret Burkett
61badf85a7 t2i training working from what I can tell at least 2023-09-17 15:56:43 -06:00
Jaret Burkett
27f343fc08 Added base setup for training t2i adapters. Currently untested, saw something else shiny i wanted to finish sirst. Added content_or_style to the training config. It defaults to balanced, which is standard uniform time step sampling. If style or content is passed, it will use cubic sampling for timesteps to favor timesteps that are beneficial for training them. for style, favor later timesteps. For content, favor earlier timesteps. 2023-09-16 08:30:38 -06:00
Jaret Burkett
17e4fe40d7 Prevent lycoris network moduels if not training that part of network. Skew timesteps to favor later steps. It performs better 2023-09-14 15:13:24 -06:00
Jaret Burkett
569d7464d5 implemented device placement preset system more places. Vastly improved speed on setting network multiplier and activating network. Fixed timing issues on progress bar 2023-09-14 08:31:54 -06:00
Jaret Burkett
4e945917df added dropout to LoRA networks 2023-09-13 15:23:07 -06:00
Jaret Burkett
ae70200d3c Bug fixes, speed improvements, compatability adjustments withdiffusers updates 2023-09-13 07:03:53 -06:00
Jaret Burkett
d74dd636ee Memory optimizations. Default to using cudamalloc when torch 2.0 for mem allocation 2023-09-12 04:30:23 -06:00
Jaret Burkett
34bfeba229 Massive speed increase. Added latent caching both to disk and to memory 2023-09-10 08:54:49 -06:00
Jaret Burkett
626ed2939a bug fixes 2023-09-09 15:04:44 -06:00
Jaret Burkett
2128ac1e08 fixed issue with embed name, save whole config to dir instead of just process so it can be easily shared. Only make one config, no timesteps 2023-09-09 12:24:08 -06:00
Jaret Burkett
be804c9cf5 Save embeddings as their trigger to match auto and comfy style loading. Also, FINALLY found why gradients were wonkey and fixed it. The root problem is dropping out of network state before backward pass. 2023-09-09 12:02:07 -06:00
Jaret Burkett
408c50ead1 actually got gradient checkpointing working, again, again, maybe 2023-09-09 11:27:42 -06:00
Jaret Burkett
b01ab5d375 FINALLY fixed gradient checkpointing issue. Big batches baby. 2023-09-08 15:21:46 -06:00
Jaret Burkett
ce4f9fe02a Bug fixes and improvements to token injection 2023-09-08 06:10:59 -06:00
Jaret Burkett
92a086d5a5 Fixed issue with token replacements 2023-09-07 13:42:39 -06:00
Jaret Burkett
3feb663a51 Switched to new bucket system that matched sdxl trained buckets. Fixed requirements. Updated embeddings to work with sdxl. Added method to train lora with an embedding at the trigger. Still testing but works amazingly well from what I can see 2023-09-07 13:06:18 -06:00
Jaret Burkett
64a5441832 Fully tested and now supporting locon on sdxl. If you have the ram 2023-09-04 14:05:10 -06:00
Jaret Burkett
a4c3507a62 Added LoCON from LyCORIS 2023-09-04 08:48:07 -06:00
Jaret Burkett
fa8fc32c0a Corrected key saving and loading to better match kohya 2023-09-04 00:22:34 -06:00
Jaret Burkett
22ed539321 Allow special args for schedulers 2023-09-03 20:38:44 -06:00
Jaret Burkett
2a40937b4f reworked samplers. Trying to find what is wrong with diffusers sampling is sdxl 2023-09-03 07:56:09 -06:00
Jaret Burkett
836fee47a6 Fixed some mismatched weights by adjusting tolerance. The mismatch ironically made the models better lol 2023-08-29 15:20:03 -06:00
Jaret Burkett
14ff51ceb4 fixed issues with converting and saving models. Cleaned keys. Improved testing for cycle load saving. 2023-08-29 12:31:19 -06:00
Jaret Burkett
714854ee86 Hude rework to move the batch to a DTO to make it far more modular to the future ui 2023-08-29 10:22:19 -06:00
Jaret Burkett
bd758ff203 Cleanup and small bug fixes 2023-08-29 05:45:49 -06:00
Jaret Burkett
a008d9e63b Fixed issue with loadin models after resume function added. Added additional flush if not training text encoder to clear out vram before grad accum 2023-08-28 17:56:30 -06:00
Jaret Burkett
e866c75638 Built base interfaces for a DTO to handle batch infomation transports for the dataloader 2023-08-28 12:43:31 -06:00