Crate ferrotorch_diffusion 

Stable-Diffusion model composition for ferrotorch.

Phase B.3 of real-artifact-driven development. This crate implements the VAE decoder (Phase B.3a) and the UNet2DConditionModel (Phase B.3b) of runwayml/stable-diffusion-v1-5. The encoder, the CLIP text encoder, and the scheduler are out of scope and tracked under follow-up dispatches.

VAE decoder

Mirrors vae/config.json — VaeDecoder inverts a latent [B, 4, 64, 64] into an image [B, 3, 512, 512]. See vae.

UNet2DConditionModel

Mirrors unet/config.json — UNet2DConditionModel consumes (noisy_latent [B, 4, 64, 64], timestep [B], text_embed [B, S, 768]) and returns predicted noise [B, 4, 64, 64]. See unet.

ResnetBlock2DTime (UNet flavour with time bias):

h = silu(norm1(x)); h = conv1(h)
t = silu(temb); h = h + Linear(t).view(B, out, 1, 1)
h = silu(norm2(h)); h = conv2(h)
out = h + (x if in==out else conv_shortcut(x))

Transformer2DModel (SD UNet flavour):

h = GroupNorm(x); h = proj_in (Conv2d k=1, [B, inner, H, W])
h = flatten to [B, HW, inner]; for block in blocks: h = block(h, ehs)
h = reshape back; h = proj_out (Conv2d k=1); out = h + residual

Each BasicTransformerBlock is the canonical pre-LN (self-attn → cross-attn → GEGLU FF) stack.

REQ status (per .design/ferrotorch-diffusion/lib.md)

REQ	Status	Evidence
REQ-1	SHIPPED	pub mod block at ferrotorch-diffusion/src/lib.rs:100..114 declares every sub-module; consumer: ferrotorch-diffusion/src/safetensors_loader.rs:17..21 imports six of them
REQ-2	SHIPPED	pub use block at ferrotorch-diffusion/src/lib.rs:116..139 re-exports the top-level types; consumer: ferrotorch-hub/src/registry.rs references ClipTextEncoder through the re-export
REQ-3	SHIPPED	crate-level lint attributes at ferrotorch-diffusion/src/lib.rs:6..59; consumer: cargo clippy -p ferrotorch-diffusion --lib -- -D warnings enforces it
REQ-4	SHIPPED	crate //! doc-comment at ferrotorch-diffusion/src/lib.rs:61..98; consumer: cargo doc -p ferrotorch-diffusion renders this as the crate landing page

Re-exports

pub use attention::Attention;

pub use attention::BasicTransformerBlock;

pub use attention::FeedForward;

pub use attention::Transformer2DModel;

pub use blocks::AttnBlock2D;

pub use blocks::DownEncoderBlock2D;

pub use blocks::Downsample2D;

pub use blocks::ResnetBlock2D;

pub use blocks::UNetMidBlock2D;

pub use blocks::UpDecoderBlock2D;

pub use blocks::Upsample2D;

pub use clip_text_encoder::ClipEncoder;

pub use clip_text_encoder::ClipEncoderLayer;

pub use clip_text_encoder::ClipMlp;

pub use clip_text_encoder::ClipSelfAttention;

pub use clip_text_encoder::ClipTextConfig;

pub use clip_text_encoder::ClipTextEmbeddings;

pub use clip_text_encoder::ClipTextEncoder;

pub use config::VaeDecoderConfig;

pub use pipeline::PipelineStepDump;

pub use pipeline::StableDiffusionPipeline;

pub use resnet_block_time::ResnetBlock2DTime;

pub use safetensors_loader::DropReport;

pub use safetensors_loader::load_clip_text_encoder;

pub use safetensors_loader::load_unet;

pub use safetensors_loader::load_vae_decoder;

pub use safetensors_loader::load_vae_encoder;

pub use scheduler::BetaSchedule;

pub use scheduler::DDIMConfig;

pub use scheduler::DDIMScheduler;

pub use scheduler::PredictionType;

pub use scheduler::TimestepSpacing;

pub use time_embedding::TimestepEmbedding;

pub use time_embedding::Timesteps;

pub use unet::AnyDownBlock;

pub use unet::AnyUpBlock;

pub use unet::CrossAttnDownBlock2D;

pub use unet::CrossAttnUpBlock2D;

pub use unet::DownBlock2D;

pub use unet::UNet2DConditionModel;

pub use unet::UNetMidBlock2DCrossAttn;

pub use unet::UpBlock2D;

pub use unet_config::UNet2DConditionConfig;

pub use vae::Decoder;

pub use vae::VaeDecoder;

pub use vae_encoder::DiagonalGaussianDistribution;

pub use vae_encoder::Encoder;

pub use vae_encoder::VaeEncoder;

pub use vae_encoder::VaeEncoderConfig;

Modules

attention

Multi-head attention + the Transformer2DModel wrapper used by the SD UNet’s CrossAttn blocks.

blocks

Building blocks of the Stable-Diffusion VAE decoder.

clip_text_encoder

Stable-Diffusion 1.5 CLIP text encoder (openai/clip-vit-large-patch14 — the text tower of CLIP-ViT-L/14).

config

Configuration for the Stable-Diffusion VAE decoder.

pipeline

Stable-Diffusion 1.5 end-to-end text-to-image generation pipeline.

resnet_block_time

ResnetBlock2DTime — the time-conditioned variant of the ResnetBlock2D used by the SD UNet.

safetensors_loader

Helpers that turn a path-to-safetensors into a loaded VaeDecoder.

scheduler

Deterministic DDIM scheduler matching diffusers.schedulers.DDIMScheduler for the Stable-Diffusion-1.5 sampling defaults.

time_embedding

Time-step sinusoidal positional encoding + the MLP that follows it.

unet

Stable-Diffusion UNet2DConditionModel forward pass.

unet_config

Configuration for the Stable-Diffusion UNet2DConditionModel.

vae

Stable-Diffusion VAE decoder composition.

vae_encoder

Stable-Diffusion VAE encoder composition.