10 Pull requests merged by 5 people

• Sd3 safetensors merge

  #1960 merged Mar 5, 2025

• Val loss improvement

  #1903 merged Feb 26, 2025

• Update IPEX libs

  #1953 merged Feb 26, 2025

• Move progress bar to account for sampling image first

  #1909 merged Feb 24, 2025

• Remove v-pred warning.

  #1918 merged Feb 24, 2025

• Remove position_ids for V2

  #1945 merged Feb 24, 2025

• Fix validation split and add test

  #1943 merged Feb 18, 2025

• Fix validation epoch loss to check epoch average

  #1938 merged Feb 17, 2025

• Fix sizes for validation split

  #1940 merged Feb 17, 2025

• Fix validation epoch divergence

  #1935 merged Feb 15, 2025

10 Pull requests opened by 8 people

• Support Lumina-image-2.0

  #1927 opened Feb 12, 2025

• Add resize interpolation parameter

  #1936 opened Feb 14, 2025

• PoC: Rewrite fine_tune.py as train_native.py

  #1950 opened Feb 23, 2025

• Fix sd 2.0/2.1 lora fine-tuning text encoder params mismatch issue

  #1951 opened Feb 24, 2025

• Fast image size reading support for JPEG XL

  #1955 opened Feb 26, 2025

• Remove super().assert_extra_args() call in sdxl_train

  #1961 opened Feb 28, 2025

• Fix missing text encoder attn modules

  #1964 opened Mar 1, 2025

• Fatser fix bug for SDXL super SD1.5 assert cant use 32

  #1966 opened Mar 2, 2025

• Refactor image resizing logic: simplify aspect ratio calculations and use round_to_steps consistently

  #1971 opened Mar 5, 2025

• Add LoRA-GGPO for Flux

  #1974 opened Mar 6, 2025

9 Issues closed by 8 people

• Cast FLUX model to fp8. This may take a while. You can reduce the time by using fp8 checkpoint.

  #1965 closed Mar 2, 2025

• What's the correct --pretrained_model_name_or_path for safetensors splitted to multiple file?

  #1967 closed Mar 2, 2025

• "Resume from saved training state" does not seem to work for Dreambooth.

  #1959 closed Mar 1, 2025

• TypeError: FluxNetworkTrainer.encode_images_to_latents() missing 1 required positional argument: 'images'

  #1954 closed Feb 26, 2025

• Flux stripes artifacts

  #1948 closed Feb 22, 2025

• No Gui file?

  #1946 closed Feb 20, 2025

• BUG: Validation breaks cached image sizes

  #1934 closed Feb 19, 2025

• Help with settings please for first time training a LoRA

  #1942 closed Feb 17, 2025

• Which kohya setting causes LoRAs to have the "is_dreambooth": true parameter?

  #1931 closed Feb 13, 2025

25 Issues opened by 24 people

• 解決済み

  #1972 opened Mar 5, 2025

• train_single_block_indices Parameter Appears Ineffective

  #1970 opened Mar 3, 2025

• Can I just save merged model after lora training for sd3.5-large?

  #1969 opened Mar 3, 2025

• Why are model names hardcoded in analyze_checkpoint_state?

  #1968 opened Mar 2, 2025

• --cache_text_encoder_outputs_to_disk seems broken (in sdxl_train_control_net_lllite.py)

  #1963 opened Mar 1, 2025

• flux_train_network.py not support min_snr_gamma

  #1962 opened Mar 1, 2025

• 'timestep_sampling=flux_shift' can offer quality improvements, but barely mentioned in docs

  #1958 opened Feb 27, 2025

• Looking to understand this repos branching strategy

  #1957 opened Feb 27, 2025

• controlnet训练

  #1956 opened Feb 27, 2025

• SDXL seems to not train self_attn layers in Text Encoders

  #1952 opened Feb 25, 2025

• svd_merge_lora ver.flux

  #1949 opened Feb 22, 2025

• Should we have a "train_native.py"?

  #1947 opened Feb 21, 2025

• At resuming: RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!

  #1944 opened Feb 18, 2025

• Why does the following error occur when running the latest flux_train_network.py with PyTorch 2.6.0+CUDA 12.4?

  #1941 opened Feb 17, 2025

• SDXL LoRA Training Stuck at Step 0

  #1939 opened Feb 16, 2025

• Bug❓: --cache_latents prevents regular checkpoint saving in train_network.py 😥

  #1937 opened Feb 15, 2025

• WARNING Unsuppored keys for sd-scripts ！

  #1933 opened Feb 14, 2025

• Train flux api

  #1932 opened Feb 14, 2025

• use multigpu(8*A800 80G) train flux_train.py OOM problem

  #1930 opened Feb 13, 2025

• The lora+ setting and prodigy cannot be used together.

  #1929 opened Feb 12, 2025

• Can the algorithm for automatic scaling of ARB buckets be specified?

  #1928 opened Feb 12, 2025

• Separate Learning Rates for UNET and TE with Prodigy

  #1926 opened Feb 9, 2025

• Multi GPU need a fix : to perform barrier as devices used by this process are currently unknown. This can potentially cause a hang if this rank to GPU mapping is incorrect.Specify device_ids in barrier() to force use of a particular device,or call

  #1925 opened Feb 7, 2025

• Support Lumina 2.0

  #1924 opened Feb 7, 2025

• Why have the options to define the single and double blocks_to_swap been removed?

  #1923 opened Feb 7, 2025

23 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• --save_state doesn't produce anything?

  #1921 commented on Feb 10, 2025 • 0 new comments

• Has anyone already done finetuning a flux model（not lora)? Why is the image quality of the model generated by my finetune so poor?

  #1681 commented on Feb 12, 2025 • 0 new comments

• Have any guide for training flux Fill lora, or fine-turn flux Fill?

  #1889 commented on Feb 15, 2025 • 0 new comments

• flux_train.py when deepspeed enable RuntimeError: mat1 and mat2 must have the same dtype, but got Float and BFloat16

  #1871 commented on Feb 17, 2025 • 0 new comments

• zero3 train flux, accelerate unwrap_model got tensor(size(0))

  #1782 commented on Feb 17, 2025 • 0 new comments

• error when use deepspeed for FLUX.1 fine-tuning

  #1591 commented on Feb 19, 2025 • 0 new comments

• Flux Dedistilled / fluxdev2pro support ?

  #1702 commented on Feb 20, 2025 • 0 new comments

• what's the value for --noise_offset

  #211 commented on Feb 20, 2025 • 0 new comments

• Can not generate full body with face lora model.

  #1916 commented on Feb 21, 2025 • 0 new comments

• [Feature] svd_merge_lora ver.flux

  #1698 commented on Feb 22, 2025 • 0 new comments

• why my flux dreambooth version can't inference right image but a noise Map

  #1579 commented on Feb 22, 2025 • 0 new comments

• Multi-gpus(RTX3090-24GB) Flux finetuning

  #1791 commented on Feb 23, 2025 • 0 new comments

• Flux textual_inversion

  #1588 commented on Feb 26, 2025 • 0 new comments

• Old LoRA load ControlNet raise _IncompatibleKeys

  #579 commented on Feb 28, 2025 • 0 new comments

• fused_backward_pass in prodigy-plus-schedule-free [SOLVED VIA EXTERNAL SOLUTION ✅]

  #1834 commented on Mar 1, 2025 • 0 new comments

• Support FLUX series models

  #1445 commented on Mar 2, 2025 • 0 new comments

• support SD3

  #1374 commented on Mar 5, 2025 • 0 new comments

• Add VRAM free after sample images

  #1396 commented on Feb 9, 2025 • 0 new comments

• New format for caching latent and Text Encoder outputs, new dataset metadata

  #1784 commented on Feb 19, 2025 • 0 new comments

• support fused_back_pass for prodigy-plus-schedule-free

  #1867 commented on Feb 25, 2025 • 0 new comments

• Add rank_stabilized for networks

  #1870 commented on Feb 11, 2025 • 0 new comments

• Fix for sdxl_gen_img.py unable to load in Google Colab Env and Distributed Inference Version (accel_sdxl_gen_img.py)

  #1901 commented on Mar 4, 2025 • 0 new comments

• Optimize Latent Caching Speed with VAE Optimizations

  #1910 commented on Feb 15, 2025 • 0 new comments