Version 0.11.1

What's Changed

• Resize and crop conditioning image in the same way as input image by @woct0rdho in #2368
• fix: correctly reuse disk-cached latents in FineTuningDataset by @kohya-ss in #2380
• Fix image cropping alignment and cache latents in FineTuningDataset by @kohya-ss in #2381
• Add timestep distribution visualization and loss weighting functionality by @kohya-ss in #2384
• feat(anima): add image-only 2D Qwen-Image VAE option by @kohya-ss in #2382
• Add per-block torch.compile support for Anima training by @kohya-ss in #2379
• feat: Anima LLLite inpainting support by @kohya-ss in #2378
• Add version 0.11.1 changes to changelog by @kohya-ss in #2387
• Add timestep visualization, loss weighting, and Anima VAE support by @kohya-ss in #2386

Full Changelog: v0.11.0...v0.11.1

Version 0.11.0

What's Changed

• Refactor for ai agents by @kohya-ss in #2372
• Merge refactoring for AI agents and humans by @kohya-ss in #2377

Full Changelog: v0.10.6...v0.11.0

Version 0.10.6

What's Changed

• Fix metadata handling for images without extensions by @kohya-ss in #2327
• Fix dataset image size handling and improve debug dataset functionality by @kohya-ss in #2328
• Fix broken links to the Dataset Configuration Guide in multiple docum… by @kohya-ss in #2329
• Add AGENTS.md to .gitignore by @kohya-ss in #2330
• doc: add Anima ControlNet-LLLite training guide to README by @kohya-ss in #2331
• Fix: Anima学習時の--base_weightsのバグを修正する by @u-haru in #2359
• Merge dev into main: dataset image size fixes and documentation updates by @kohya-ss in #2373
• chore: update README for 0.10.6 by @kohya-ss in #2375

Full Changelog: v0.10.5...v0.10.6

Version 0.10.5

What's Changed

• fix: improve numerical stability by conditionally using float32 in Anima by @kohya-ss in #2302
• fix: improve numerical stability by conditionally using float32 in Anima with fp16 training by @kohya-ss in #2303
• update ipex by @WhitePr in #2307
• fix: replace deprecated CLIPFeatureExtractor with CLIPImageProcessor for transformers 5.x compatibility by @marcus165090-spec in #2315
• Update for supporting transformers>=5 by @kohya-ss in #2316
• Add inpainting training and sampling support for SD1.5 and SDXL by @allanoepping in #2309
• Add inpainting training and sampling support for SD1.5 and SDXL by @kohya-ss in #2318
• Feat anima lllite by @kohya-ss in #2317
• Support transformers 5 and add ControlNet-LLLite for Anima by @kohya-ss in #2325

New Contributors

• @WhitePr made their first contribution in #2307
• @marcus165090-spec made their first contribution in #2315
• @allanoepping made their first contribution in #2309

Full Changelog: v0.10.2...v0.10.5

Version 0.10.2

What's Changed

• Fix bug and optimization for Lumina model by @duongve13112002 in #2262
• fix: str is not "no" to str != "no" by @kohya-ss in #2267
• merge sd3 to main by @kohya-ss in #2268
• Fix the LoRA dropout issue in the Anima model and implement support for network_reg_alphas during LoRA training. by @duongve13112002 in #2272
• Multi-resolution dataset for SD1/SDXL by @woct0rdho in #2269
• feat: Stability with fp16 for anima by @kohya-ss in #2277
• feat: backward compatibility for SD/SDXL latent cache by @kohya-ss in #2276
• feat: Support LoKr/LoHa for SDXL and Anima by @kohya-ss in #2275
• fix: rename character_tags to img_character_tags to fix UnboundLocalError by @kohya-ss in #2280
• Add skip_image_resolution to deduplicate multi-resolution dataset by @woct0rdho in #2273
• docs: add skip_image_resolution to config README by @kohya-ss in #2288
• Bump pytorch-optimizer into 3.10.0 by @kozistr in #2283
• fix: Anima validation dataset not working with Text Encoder output cache by @kohya-ss in #2291
• Log d*eff_lr for ProdigyPlusScheduleFree, and some cleanup of LR logging by @woct0rdho in #2289
• Use torch.svd_lowrank for large matrices in resize_lora.py by @woct0rdho in #2240
• Fix the 'off by 1' problem in dynamically resized LoRA rank by @woct0rdho in #2194
• Add --svd_lowrank_niter option to resize_lora.py by @kohya-ss in #2296
• feat: SD1.x/2.x と SDXL 向けの LECO 学習スクリプトを追加 by @umisetokikaze in #2285
• Fix/leco cleanup by @kohya-ss in #2294
• fix: AdaLN modulation to use float32 for numerical stability in fp16 by @kohya-ss in #2297
• Improve clarity of README table of contents and change history by @kohya-ss in #2299
• doc: update change history in README files to include LECO training by @kohya-ss in #2300
• Print verbose info while resizing LoRA is running by @woct0rdho in #2301
• Merge development changes into main by @kohya-ss in #2298

New Contributors

• @umisetokikaze made their first contribution in #2285

Full Changelog: v0.10.1...v0.10.2

Version 0.10.1

What's Changed

• Fix AttributeError for UNet2DConditionModel with newer diffusers versions by @cgcalatrava in #2255
• Support Anima model by @duongve13112002 in #2260
• Add/modify some implementation for anima by @kohya-ss in #2261
• feat: Anima support by @kohya-ss in #2263
• Release 0.10.1 by @kohya-ss in #2264

New Contributors

• @cgcalatrava made their first contribution in #2255
• @duongve13112002 made their first contribution in #2260

Full Changelog: v0.10.0...v0.10.1

Version 0.10.0

What's Changed

• Fix AttributeError: 'T5EncoderModel' object has no attribute 'text_model', while loading T5 model in GPU. by @fireicewolf in #1452
• Fix AttributeError: 'FluxNetworkTrainer' object has no attribute 'sample_prompts_te_outputs' by @fireicewolf in #1461
• Flux の LoRA マージ機能を修正 by @exveria1015 in #1469
• Flux merge lora by @kohya-ss in #1482
• make guidance_scale keep float in args by @Akegarasu in #1525
• fix: text_encoder_conds referenced before assignment by @Akegarasu in #1529
• Sd3 freeze x_block by @sdbds in #1417
• Add Flux_Shift for solving the problem of multi-resolution training blurry by @sdbds in #1541
• Critical typo fix by @cocktailpeanut in #1592
• Improve wandb logging by @p1atdev in #1576
• sd3 schedule free opt by @kohya-ss in #1605
• fix: backward compatibility for text_encoder_lr by @Akegarasu in #1620
• Retain alpha in pil_resize for --alpha_mask by @emcmanus in #1619
• Support Sdxl Controlnet by @sdbds in #1648
• ControlNet training for SDXL by @kohya-ss in #1655
• fix: fix some distributed training error in windows by @Akegarasu in #1686
• Fast image sizes by @kohya-ss in #1660
• Caching latents and Text Encoder outputs with multiple GPUs by @kohya-ss in #1690
• Differential Output Preservation loss for LoRA by @kohya-ss in #1710
• SD3.5 Large support by @kohya-ss in #1719
• Dango/timesteps fix by @Dango233 in #1768
• Update README.md by @kohya-ss in #1769
• fix: sort order when getting image size from cache file by @feffy380 in #1770
• Support avif get image size by @sdbds in #1774
• Improve block swap speed and apply to LoRA by @kohya-ss in #1779
• Implement pseudo Huber loss for Flux and SD3 by @recris in #1808
• Flux huber loss by @kohya-ss in #1815
• Allow unknown schedule-free optimizers to continue to module loader by @rockerBOO in #1811
• Add pytest testing by @rockerBOO in #1812
• Add Flux ControlNet by @minux302 in #1813
• Workflow tests fixes and documentation by @rockerBOO in #1817
• Unify controlnet parameters name and change scripts name. by @sdbds in #1821
• Prevent git credentials from leaking into other actions by @rockerBOO in #1828
• Added RAdamScheduleFree support by @nhamanasu in #1830
• Validation loss by @rockerBOO in #1864
• Val loss by @kohya-ss in #1899
• Fix validation epoch divergence by @rockerBOO in #1935
• Fix sizes for validation split by @rockerBOO in #1940
• Fix validation epoch loss to check epoch average by @rockerBOO in #1938
• Fix validation split and add test by @rockerBOO in #1943
• Remove position_ids for V2 by @yidiq7 in #1945
• Remove v-pred warning. by @tsukimiya in #1918
• Move progress bar to account for sampling image first by @rockerBOO in #1909
• Update IPEX libs by @Disty0 in #1953
• Val loss improvement by @kohya-ss in #1903
• Sd3 safetensors merge by @kohya-ss in #1960
• Support pytorch_optimizer by @gesen2egee in #1985
• Fatser fix bug for SDXL super SD1.5 assert cant use 32 by @sdbds in #1966
• Add resize interpolation parameter by @rockerBOO in #1936
• Resize interpolation small fix by @kohya-ss in #2018
• Add LoRA-GGPO for Flux by @rockerBOO in #1974
• Fast image size reading support for JPEG XL by @Disty0 in #1955
• fix: add resize_interpolation parameter to FineTuningDataset constructor by @kohya-ss in #2026
• Fix size parameter types and improve resize_image interpolation by @LexSong in #2022
• Fix resize PR link by @rockerBOO in #2028
• Add IP noise gamma for Flux by @rockerBOO in #1992
• Using --vae_batch_size to set batch size for dynamic latent generation by @DKnight54 in #2015
• fix: update hf_hub_download parameters to fix wd14 tagger regression by @GlenCarpenter in #2053
• Add CFG for sampling in training with FLUX.1 by @kohya-ss in #2064
• Create FUNDING.yml by @kohya-ss in #2065
• Quick fix flux sampling scales by @kohya-ss in #2066
• Fix mean image aspect ratio error calculation to avoid NaN values by @kohya-ss in #2070
• Fix: try aligning dtype of matrixes when training with deepspeed and mixed-precision is set to bf16 or fp16 by @sharlynxy in #2060
• Deepspeed readme by @kohya-ss in #2074
• Add pythonpath to pytest.ini by @rockerBOO in #2072
• Fix mean grad norms by @rockerBOO in #2073
• Add English docs with Japanese details by @kohya-ss in #2086
• Update workflows to read-all instead of write-all by @rockerBOO in #2088
• Add tests for syntax checking training scripts by @rockerBOO in #2108
• Update IPEX libs by @Disty0 in #2121
• Fix unwrap_model handling for None text_encoders in sample_images by @kohya-ss in #2115
• Support Lumina-image-2.0 by @sdbds in #1927
• Lumina: Change to 3 by @rockerBOO in #2146
• Add prompt guidance files for Claude and Gemini by @kohya-ss in #2147
• feat: add .claude and .gemini to .gitignore by @kohya-ss in #2148
• Fix alphas cumprod after add_noise for DDIMScheduler by @rockerBOO in #2153
• Feature lumina image by @kohya-ss in #2138
• experimental: Documentation Updates with Gemini by @kohya-ss in #2042
• Feature Chroma support by @kohya-ss in #2157
• Force t5 attn mask for chroma training by @kohya-ss in #2165
• Update pytorch-optimizer to 3.7.0 by @kozistr in #2160
• Fix validation documentation by @rockerBOO in #2167
• Update model spec to 1.0.1 by @rockerBOO in #2168
• fix model type handling in analyze_state_dict_state function for SD3 by @kohya-ss in #2174
• Resize ControlLoRA by @woct0rdho in #2175
• doc: add sponsor logo and annoucements by @kohya-ss in #2184
• gitignore: add CLAUDE.md, GEMINI.md, MagicMock and related files by @kohya-ss in #2185
• Feat: Update libraries, remove warnings by @kohya-ss in #2178
• fix: chroma LoRA training without Text Encode caching by @kohya-ss in #2191
• Fix validation dataset documentation to not use subsets by @rockerBOO in #2196
• Doc update for latest features by @kohya-ss in #2192
• feat: Speeding up loading .safetensors files by @kohya-ss in #2200
• feat: support HunyuanImage-2.1 by @kohya-ss in https://github.com/kohya-ss/sd-scripts/pull/...

Version 0.9.1

What's Changed

• Fix missing text encoder attn modules by @Nekotekina in #1964

New Contributors

• @Nekotekina made their first contribution in #1964

Full Changelog: v0.9.0...v0.9.1

Version 0.9.0

See #1879 for details.

Version 0.8.8

Oct 27, 2024 / 2024-10-27:

• svd_merge_lora.py VRAM usage has been reduced. However, main memory usage will increase (32GB is sufficient).

Oct 26, 2024 / 2024-10-26:

• Fixed a bug in svd_merge_lora.py, sdxl_merge_lora.py, and resize_lora.py where the hash value of LoRA metadata was not correctly calculated when the save_precision was different from the precision used in the calculation. See issue #1722 for details. Thanks to JujoHotaru for raising the issue.

Sep 13, 2024 / 2024-09-13:

• sdxl_merge_lora.py now supports OFT. Thanks to Maru-mee for the PR #1580.
• svd_merge_lora.py now supports LBW. Thanks to terracottahaniwa. See PR #1575 for details.
• sdxl_merge_lora.py also supports LBW.
• See LoRA Block Weight by hako-mikan for details on LBW.

Jun 23, 2024 / 2024-06-23:

• Fixed cache_latents.py and cache_text_encoder_outputs.py not working.

Oct 27, 2024 / 2024-10-27:

• svd_merge_lora.py のVRAM使用量を削減しました。ただし、メインメモリの使用量は増加します（32GBあれば十分です）。

Oct 26, 2024 / 2024-10-26:

• svd_merge_lora.py、sdxl_merge_lora.py、resize_lora.pyで、保存時の精度が計算時の精度と異なる場合、LoRAメタデータのハッシュ値が正しく計算されない不具合を修正しました。詳細は issue #1722 をご覧ください。問題提起していただいた JujoHotaru 氏に感謝します。

Sep 13, 2024 / 2024-09-13:

• sdxl_merge_lora.py で OFT がサポートされました。PR #1580 Maru-mee 氏に感謝します。
• svd_merge_lora.py で LBW がサポートされました。PR #1575 terracottahaniwa 氏に感謝します。
• sdxl_merge_lora.py でも LBW がサポートされました。
• LBW の詳細は hako-mikan 氏の LoRA Block Weight をご覧ください。

Jun 23, 2024 / 2024-06-23:

• cache_latents.py および cache_text_encoder_outputs.py が動作しなくなっていたのを修正しました。