24 Pull requests merged by 16 people

• [train_dreambooth_lora.py] Set LANCZOS as default interpolation mode for resizing

  #11421 merged Apr 26, 2025

• Modular Diffusers Guiders

  #11311 merged Apr 25, 2025

• Fix typos in strings and comments

  #11407 merged Apr 24, 2025

• [BUG] fixed WAN docstring

  #11226 merged Apr 24, 2025

• Fix Flux IP adapter argument in the pipeline example

  #11402 merged Apr 24, 2025

• [HiDream LoRA] optimizations + small updates

  #11381 merged Apr 24, 2025

• Kolors additional pipelines, community contrib

  #11372 merged Apr 23, 2025

• [Feature] Added Xlab Controlnet support

  #11249 merged Apr 23, 2025

• [train_dreambooth_flux] Add LANCZOS as the default interpolation mode for image resizing

  #11395 merged Apr 23, 2025

• Fix group offloading with block_level and use_stream=True

  #11375 merged Apr 23, 2025

• Update README_hidream.md

  #11386 merged Apr 23, 2025

• [HiDream] move deprecation to 0.35.0

  #11384 merged Apr 22, 2025

• Update modeling imports

  #11129 merged Apr 22, 2025

• [LoRA] add LoRA support to HiDream and fine-tuning script

  #11281 merged Apr 22, 2025

• Add stochastic sampling to FlowMatchEulerDiscreteScheduler

  #11369 merged Apr 22, 2025

• [Refactor] Minor Improvement for import utils

  #11161 merged Apr 21, 2025

• Support different-length pos/neg prompts for FLUX.1-schnell variants like Chroma

  #11120 merged Apr 21, 2025

• [cogview4][feat] Support attention mechanism with variable-length support and batch packing

  #11349 merged Apr 21, 2025

• Add Serialized Type Name kwarg in Model Output

  #10502 merged Apr 21, 2025

• [BUG] fixes in kadinsky pipeline

  #11080 merged Apr 21, 2025

• Fix Wan I2V prepare_latents dtype

  #11371 merged Apr 21, 2025

• fix issue that training flux controlnet was unstable and validation r…

  #11373 merged Apr 21, 2025

• [train_dreambooth_lora_sdxl.py] Fix the LR Schedulers when num_train_epochs is passed in a distributed training env

  #11240 merged Apr 21, 2025

• [Flux LoRAs] fix lr scheduler bug in distributed scenarios

  #11242 merged Apr 21, 2025

16 Pull requests opened by 13 people

• Add VisualCloze

  #11377 opened Apr 21, 2025

• Add StableDiffusion3InstructPix2PixPipeline

  #11378 opened Apr 22, 2025

• [docs] Memory optims

  #11385 opened Apr 22, 2025

• Add fluxfill train code

  #11393 opened Apr 23, 2025

• Fixing missing provider options argument

  #11397 opened Apr 23, 2025

• [Feature] AutoModel can load components using model_index.json

  #11401 opened Apr 23, 2025

• enable 28 GGUF test cases on XPU

  #11404 opened Apr 24, 2025

• enable group_offload cases and quanto cases on XPU

  #11405 opened Apr 24, 2025

• enable test_layerwise_casting_memory cases on XPU

  #11406 opened Apr 24, 2025

• Add LORA Hotswap Compatibility with torch.compile()

  #11409 opened Apr 24, 2025

• [Feature] Implement tiled VAE encoding/decoding for Wan model.

  #11414 opened Apr 24, 2025

• Bump h11 from 0.14.0 to 0.16.0 in /examples/server

  #11415 opened Apr 24, 2025

• Fix typos in docs and comments

  #11416 opened Apr 25, 2025

• Raise warning instead of error for block offloading with streams

  #11425 opened Apr 27, 2025

• [train_text_to_image] Better image interpolation in training scripts follow up

  #11426 opened Apr 27, 2025

• [train_text_to_image_lora] Better image interpolation in training scripts follow up

  #11427 opened Apr 27, 2025

10 Issues closed by 6 people

• train flux controlnet with train_controlnet_flux.py script get a bad result with toy-dataset(fill50k)!!

  #11181 closed Apr 25, 2025

• pipeline_wan_i2v.py: minor discrepancy between arg default and docstring

  #11005 closed Apr 24, 2025

• flux img2img controlnet channels error

  #9979 closed Apr 23, 2025

• HiDream running into issues with group offloading at the block-level

  #11307 closed Apr 23, 2025

• RMSNorm in AutoencoderDC should be RMSNorm2d

  #11387 closed Apr 23, 2025

• HiDream pipeline update breaks downstream

  #11382 closed Apr 22, 2025

• [bnb] Moving a pipeline with a 8bit quantized model to CPU doesn't throw warning

  #11352 closed Apr 22, 2025

• Offloading behaviour for HiDream seems broken

  #11376 closed Apr 22, 2025

• `prepare_image` in Kandinsky pipelines doesn't support `torch.Tensor`

  #11060 closed Apr 21, 2025

• HunyuanVideoImageToVideoPipeline failures

  #11118 closed Apr 21, 2025

19 Issues opened by 17 people

• RuntimeError: group_norm in VAE decode for SDXL Masked Img2Img (even with ControlNets disabled & FP32 VAE/Latents)

  #11424 opened Apr 26, 2025

• Lora Hotswap no clear documentation

  #11423 opened Apr 26, 2025

• validation prompt argument crashes the training [HiDream][Lora]

  #11422 opened Apr 26, 2025

• How to know that "Textual inversion" file I have loaded and not turn it on?

  #11419 opened Apr 25, 2025

• How to add flux1-fill-dev-fp8.safetensors

  #11418 opened Apr 25, 2025

• attributeerror: 'distributeddataparallel' object has no attribute 'dtype'. did you mean: 'type'?

  #11417 opened Apr 25, 2025

• RuntimeError: Failed to import diffusers.pipelines.cogvideo.pipeline_cogvideox because of the following error (look up to see its traceback):

  #11413 opened Apr 24, 2025

• prs-eth/marigold-iid-appearance-v1-1 is 404

  #11412 opened Apr 24, 2025

• File "/usr/local/python/lib/python3.8/site-packages/diffusers/models/attention_processor.py", line 3074, in __call__ [rank0]: batch_size, key_tokens, _ = ( [rank0]: ValueError: too many values to unpack (expected 3)

  #11410 opened Apr 24, 2025

• set_adapters not support for compiled model

  #11408 opened Apr 24, 2025

• FluxTransformer2DModel does not have config and cannot set_default_attn_processor

  #11398 opened Apr 23, 2025

• How to convert the hidream lora trained by diffusers to a format that comfyui can load?

  #11396 opened Apr 23, 2025

• [Feature request] ostris / Flex.2-preview. Open Source 8B parameter T2I Diffusion Model with universal control and inpainting support built in.

  #11394 opened Apr 23, 2025

• Better image interpolation in training scripts follow up

  #11390 opened Apr 23, 2025

• scheduler refactor to support different `set_timesteps` and `step` method

  #11389 opened Apr 22, 2025

• extend AutoModel to be able to load transformer models & custom diffusers models

  #11388 opened Apr 22, 2025

• HiDream LoRA support

  #11383 opened Apr 22, 2025

• [Feature request] Integrate SkyReels-V2 support in diffusers

  #11374 opened Apr 21, 2025

• convert_hunyuan_video_to_diffusers.py not working - invalid load key, '\xd8'.

  #11370 opened Apr 20, 2025

38 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.

• Add VidTok AutoEncoders

  #11261 commented on Apr 21, 2025 • 17 new comments

• Attention Dispatcher

  #11368 commented on Apr 22, 2025 • 8 new comments

• Add generic support for Intel Gaudi accelerator (hpu device)

  #11328 commented on Apr 24, 2025 • 8 new comments

• [docs] Model cards

  #11112 commented on Apr 23, 2025 • 3 new comments

• feat: pipeline-level quantization config

  #11130 commented on Apr 27, 2025 • 3 new comments

• Cosmos

  #10660 commented on Apr 26, 2025 • 0 new comments

• Support for control-lora

  #10686 commented on Apr 22, 2025 • 0 new comments

• [tests] add tests to check for graph breaks, recompilation, cuda syncs in pipelines during torch.compile()

  #11085 commented on Apr 27, 2025 • 0 new comments

• Add gradient checkpointing support for AutoencoderKLWan

  #11105 commented on Apr 21, 2025 • 0 new comments

• Hunyuan Video adjustments

  #11140 commented on Apr 25, 2025 • 0 new comments

• [Quantization] Add TRT as a Backend

  #11173 commented on Apr 25, 2025 • 0 new comments

• First Block Cache

  #11180 commented on Apr 22, 2025 • 0 new comments

• [modular diffusers] more refactor

  #11235 commented on Apr 25, 2025 • 0 new comments

• [LoRA] improve LoRA fusion tests

  #11274 commented on Apr 24, 2025 • 0 new comments

• [Tests] Enable more general testing for `torch.compile()` with LoRA hotswapping

  #11322 commented on Apr 25, 2025 • 0 new comments

• [LoRA] parse metadata from LoRA and save metadata

  #11324 commented on Apr 21, 2025 • 0 new comments

• [docs] Adapters

  #11331 commented on Apr 23, 2025 • 0 new comments

• Pipeline deprecations

  #11354 commented on Apr 21, 2025 • 0 new comments

• Update License year; Add missing license for HiDream implementation

  #11358 commented on Apr 25, 2025 • 0 new comments

• bug while using cogvideox image to video pipeline

  #11133 commented on Apr 20, 2025 • 0 new comments

• [Community] Help us fix the LR schedulers when `num_train_epochs` is passed in a distributed training env

  #8384 commented on Apr 21, 2025 • 0 new comments

• CogvideoX VAE decoder consumes significantly more memory in the latest version

  #10091 commented on Apr 22, 2025 • 0 new comments

• [core] introduce auto classes for T2V and I2V

  #10486 commented on Apr 22, 2025 • 0 new comments

• Civit AI flux model razor-8step-rapid-real not working with diffusers single file

  #11127 commented on Apr 22, 2025 • 0 new comments

• Implement caching on LTX and WAN video models

  #11134 commented on Apr 22, 2025 • 0 new comments

• Error in loading Lora

  #11224 commented on Apr 22, 2025 • 0 new comments

• [Feature request] LTX-Video v0.9.6 15x faster inference than non-distilled model.

  #11359 commented on Apr 22, 2025 • 0 new comments

• The density_for_timestep_sampling and loss_weighting for SD3 Training！！！

  #9056 commented on Apr 22, 2025 • 0 new comments

• [New Feature] Flux PuLID

  #10141 commented on Apr 23, 2025 • 0 new comments

• Flux FIll lora / dreambooth finetune support

  #11145 commented on Apr 23, 2025 • 0 new comments

• hello i cant generate beautiful anime women

  #11345 commented on Apr 23, 2025 • 0 new comments

• TypeError: BnB4BitDiffusersQuantizer.create_quantized_param() got an unexpected keyword argument 'dtype'

  #11149 commented on Apr 24, 2025 • 0 new comments

• Run FLUX-controlnet zero3 training failed: 'weight' must be 2-D

  #10969 commented on Apr 25, 2025 • 0 new comments

• APG: Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion Models

  #9585 commented on Apr 25, 2025 • 0 new comments

• HunyuanVideo pipe.transformer.compile(): torch._dynamo hit config.recompile_limit (8)

  #10932 commented on Apr 26, 2025 • 0 new comments

• flux controlnet train ReadMe have a bug

  #11321 commented on Apr 27, 2025 • 0 new comments

• Wan Video Vae RemoteVae

  #11326 commented on Apr 27, 2025 • 0 new comments

• make multi controlnet more robust

  #9920 commented on Apr 21, 2025 • 0 new comments