Suraj Patil

[POC] allow custom height, width in StableDiffusionPipeline.

1

This PR shows how to generate images with custom height and width, for example can be used to generate recatngular images. ```python pipe(prompt, height=512, width=256, guidance_scale=7.5)["sample"] ```

[WIP] Latnet (and Stable) diffusion text2img conversion script.

1

This PR adds conversion script for text2img latent diffusion models from CompVis.

dalle mega

2

# What does this PR do? This PR adds the DalleMega model from [dalle-mini](https://github.com/borisdayma/dalle-mini) for text-2-image generation. The VQGAN model required for converting the tokens to image is in this...

simplyfy AttentionBlock

1

This simplifies `AttentionBlock` by always making q,k,v a 3D tensors like we do in `CrossAttention`. This way we can also leverage sliced attention and xformers attention in this block.

[generate] return past_key_values

11

# What does this PR do? Allows returning `past_key_values` from `generate` when `use_cache=True`. Like other returned values, `past_key_values` are also returned as `Tuple`, one element per generated token. Fixes #17016

How to handle gradient accumulation with multiple models ?

9

To do gradient accumulation with `accelerate` we wrap the model in `accelerator.accumulate` context. But what would be the right way to achieve this when multiple models are involved ? For...

[WIP] Add inpainting training script

6

# What does this PR do? Add in-painting training script.

[don't merge] open-sora test pipeline

3

# What does this PR do? A test implementation for open-sora. ```bash pip install git+https://github.com/huggingface/diffusers.git@st-dit ``` ```py import torch from diffusers import OpenSoraPipeline from diffusers.utils import export_to_gif pipe = OpenSoraPipeline.from_pretrained("diffusers/open-sora-pipe-test",...