Stable Diffusion Latent Upscaler (#2059)
* Modify UNet2DConditionModel
- allow skipping mid_block
- adding a norm_group_size argument so that we can set the `num_groups` for group norm using `num_channels//norm_group_size`
- allow user to set dimension for the timestep embedding (`time_embed_dim`)
- the kernel_size for `conv_in` and `conv_out` is now configurable
- add random fourier feature layer (`GaussianFourierProjection`) for `time_proj`
- allow user to add the time and class embeddings before passing through the projection layer together - `time_embedding(t_emb + class_label))`
- added 2 arguments `attn1_types` and `attn2_types`
* currently we have argument `only_cross_attention`: when it's set to `True`, we will have a to the
`BasicTransformerBlock` block with 2 cross-attention , otherwise we
get a self-attention followed by a cross-attention; in k-upscaler, we need to have blocks that include just one cross-attention, or self-attention -> cross-attention;
so I added `attn1_types` and `attn2_types` to the unet's argument list to allow user specify the attention types for the 2 positions in each block; note that I stil kept
the `only_cross_attention` argument for unet for easy configuration, but it will be converted to `attn1_type` and `attn2_type` when passing down to the down blocks
- the position of downsample layer and upsample layer is now configurable
- in k-upscaler unet, there is only one skip connection per each up/down block (instead of each layer in stable diffusion unet), added `skip_freq = "block"` to support
this use case
- if user passes attention_mask to unet, it will prepare the mask and pass a flag to cross attention processer to skip the `prepare_attention_mask` step
inside cross attention block
add up/down blocks for k-upscaler
modify CrossAttention class
- make the `dropout` layer in `to_out` optional
- `use_conv_proj` - use conv instead of linear for all projection layers (i.e. `to_q`, `to_k`, `to_v`, `to_out`) whenever possible. note that when it's used to do cross
attention, to_k, to_v has to be linear because the `encoder_hidden_states` is not 2d
- `cross_attention_norm` - add an optional layernorm on encoder_hidden_states
- `attention_dropout`: add an optional dropout on attention score
adapt BasicTransformerBlock
- add an ada groupnorm layer to conditioning attention input with timestep embedding
- allow skipping the FeedForward layer in between the attentions
- replaced the only_cross_attention argument with attn1_type and attn2_type for more flexible configuration
update timestep embedding: add new act_fn gelu and an optional act_2
modified ResnetBlock2D
- refactored with AdaGroupNorm class (the timestep scale shift normalization)
- add `mid_channel` argument - allow the first conv to have a different output dimension from the second conv
- add option to use input AdaGroupNorm on the input instead of groupnorm
- add options to add a dropout layer after each conv
- allow user to set the bias in conv_shortcut (needed for k-upscaler)
- add gelu
adding conversion script for k-upscaler unet
add pipeline
* fix attention mask
* fix a typo
* fix a bug
* make sure model can be used with GPU
* make pipeline work with fp16
* fix an error in BasicTransfomerBlock
* make style
* fix typo
* some more fixes
* uP
* up
* correct more
* some clean-up
* clean time proj
* up
* uP
* more changes
* remove the upcast_attention=True from unet config
* remove attn1_types, attn2_types etc
* fix
* revert incorrect changes up/down samplers
* make style
* remove outdated files
* Apply suggestions from code review
* attention refactor
* refactor cross attention
* Apply suggestions from code review
* update
* up
* update
* Apply suggestions from code review
* finish
* Update src/diffusers/models/cross_attention.py
* more fixes
* up
* up
* up
* finish
* more corrections of conversion state
* act_2 -> act_2_fn
* remove dropout_after_conv from ResnetBlock2D
* make style
* simplify KAttentionBlock
* add fast test for latent upscaler pipeline
* add slow test
* slow test fp16
* make style
* add doc string for pipeline_stable_diffusion_latent_upscale
* add api doc page for latent upscaler pipeline
* deprecate attention mask
* clean up embeddings
* simplify resnet
* up
* clean up resnet
* up
* correct more
* up
* up
* improve a bit more
* correct more
* more clean-ups
* Update docs/source/en/api/pipelines/stable_diffusion/latent_upscale.mdx
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* Update docs/source/en/api/pipelines/stable_diffusion/latent_upscale.mdx
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* add docstrings for new unet config
* Update src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_latent_upscale.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* Update src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_latent_upscale.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* # Copied from
* encode the image if not latent
* remove force casting vae to fp32
* fix
* add comments about preconditioning parameters from k-diffusion paper
* attn1_type, attn2_type -> add_self_attention
* clean up get_down_block and get_up_block
* fix
* fixed a typo(?) in ada group norm
* update slice attention processer for cross attention
* update slice
* fix fast test
* update the checkpoint
* finish tests
* fix-copies
* fix-copy for modeling_text_unet.py
* make style
* make style
* fix f-string
* Update src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_latent_upscale.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* fix import
* correct changes
* fix resnet
* make fix-copies
* correct euler scheduler
* add missing #copied from for preprocess
* revert
* fix
* fix copies
* Update docs/source/en/api/pipelines/stable_diffusion/latent_upscale.mdx
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
* Update docs/source/en/api/pipelines/stable_diffusion/latent_upscale.mdx
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
* Update docs/source/en/api/pipelines/stable_diffusion/latent_upscale.mdx
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
* Update docs/source/en/api/pipelines/stable_diffusion/latent_upscale.mdx
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
* Update src/diffusers/models/cross_attention.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
* Update src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_latent_upscale.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
* Update src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_latent_upscale.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
* clean up conversion script
* KDownsample2d,KUpsample2d -> KDownsample2D,KUpsample2D
* more
* Update src/diffusers/models/unet_2d_condition.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
* remove prepare_extra_step_kwargs
* Update src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_latent_upscale.py
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
* Update src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_latent_upscale.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* fix a typo in timestep embedding
* remove num_image_per_prompt
* fix fasttest
* make style + fix-copies
* fix
* fix xformer test
* fix style
* doc string
* make style
* fix-copies
* docstring for time_embedding_norm
* make style
* final finishes
* make fix-copies
* fix tests
---------
Co-authored-by: yiyixuxu <yixu@yis-macbook-pro.lan>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Pedro Cuenca <pedro@huggingface.co>
