ru-dalle

Converts Russian language text prompts into images through a two-stage pipeline: a DalleTransformer encoder processes tokenized Russian text into a latent representation, which is then decoded by a Variational Autoencoder (VAE) into pixel-space images. The architecture uses transformer attention mechanisms for semantic understanding of Russian language nuances and supports multiple pre-trained model variants (Malevich, Emojich, Surrealist, Kandinsky) with parameter counts ranging from 1.3B to 12B, enabling trade-offs between generation speed and output quality.

Provides four distinct pre-trained model checkpoints (Malevich for general-purpose, Emojich for emoji-style, Surrealist for artistic, Kandinsky for high-quality) accessible via `get_rudalle_model()` API function. Each variant is independently trained on curated datasets emphasizing different visual styles, allowing users to select the appropriate model for their generation task without retraining. Model loading is abstracted through a registry pattern that handles checkpoint downloading, caching, and device placement (CPU/GPU).

Extends core image generation to produce sequences of images that form coherent videos through temporal consistency constraints. The VideoDALLE extension applies the generation pipeline frame-by-frame while maintaining visual continuity between frames, using techniques like optical flow guidance or latent space interpolation to ensure smooth transitions. This enables video generation from text prompts without training separate video models.

Provides infrastructure for adapting pre-trained models to specialized domains by fine-tuning on custom Russian image-text pair datasets. The fine-tuning pipeline supports both full model training and parameter-efficient methods (LoRA, adapter layers) to reduce computational requirements. Users can supply custom datasets, configure training hyperparameters, and evaluate fine-tuned models on validation sets, enabling domain-specific image generation without training from scratch.

Extends text-only generation by accepting optional image prompts that condition the generation process, allowing users to guide visual output toward specific reference images. The system encodes reference images into the same latent space as text tokens, concatenating or blending these representations before passing to the VAE decoder. This enables fine-grained control over composition, style, and content without full image-to-image translation.

Post-processes generated images through RealESRGAN (Real-ESRGAN) super-resolution model to upscale output resolution by 2x-4x with detail enhancement. The enhancement pipeline is decoupled from core generation, allowing optional application after image synthesis. RealESRGAN uses a residual dense network trained on perceptual loss to reconstruct high-frequency details, converting low-resolution VAE outputs into sharper, higher-resolution images suitable for print or display.

Filters and ranks generated images by computing semantic similarity between image content and original text prompt using ruCLIP (Russian CLIP), a vision-language model trained on Russian image-text pairs. The system encodes both the prompt and each generated image into a shared embedding space, computing cosine similarity scores to identify images most aligned with user intent. This enables cherry-picking best results from batch generations without manual review.

Provides fine-grained control over generation randomness through top-k (select from k most likely tokens) and top-p (nucleus sampling, select from smallest set of tokens with cumulative probability ≥ p) parameters passed to the DalleTransformer decoder. These sampling strategies control the trade-off between diversity (high k/p) and coherence (low k/p) during autoregressive token generation, allowing users to tune output variability without retraining models.

Allows generation of images in non-square aspect ratios (e.g., 16:9, 4:3, 1:2) by adjusting VAE decoder input dimensions and applying aspect-ratio-aware padding or cropping during latent space processing. The system supports multiple predefined aspect ratios and custom dimensions, enabling users to generate images optimized for specific display contexts (mobile, widescreen, portrait) without training aspect-ratio-specific models.

Implements a specialized tokenizer that converts Russian language text into discrete tokens compatible with the DalleTransformer encoder. The tokenizer handles Cyrillic character encoding, Russian morphology, and language-specific preprocessing (punctuation normalization, case handling) to create token sequences that preserve semantic meaning for the transformer. Tokens are mapped to learned embeddings in the transformer's vocabulary space, enabling the model to understand Russian language nuances.

Implements a Variational Autoencoder that maps latent representations (produced by DalleTransformer) into high-dimensional pixel space, reconstructing images from compressed latent codes. The VAE decoder uses transposed convolutions and upsampling layers to progressively reconstruct image details from low-resolution latent features, enabling efficient generation without pixel-space autoregression. The decoder is trained jointly with the encoder to minimize reconstruction loss, enabling lossy compression of image information into latent space.

Supports generating multiple images from the same or different prompts by iterating through input prompts and applying the generation pipeline sequentially. The system accumulates generated images in memory or writes them to disk, providing options for batch result aggregation, filtering, and ranking. While individual generation steps are sequential (no parallelization within a single batch), the API abstracts batch handling to simplify multi-image workflows.