RPG-DiffusionMaster

Leverages multimodal large language models (GPT-4 or local models via mllm.py) to analyze and refine user-provided text prompts, enriching them with additional detail, clarity, and structural information before passing to the diffusion pipeline. The system uses templated prompt engineering to guide MLLMs toward consistent, parseable outputs that enhance semantic richness while maintaining user intent.

Decomposes image generation into spatially-aware regions by using MLLMs to analyze the recaptioned prompt and generate region-specific sub-prompts along with split ratios that define how the image canvas should be divided. The planning phase (via mllm.py's get_params_dict()) parses MLLM output into structured region definitions, enabling precise control over object placement and attribute binding across different image areas without retraining the diffusion model.

Supports generating multiple images from different prompts while maintaining consistent regional decomposition strategies (e.g., same split ratios, same region count) across the batch. The MLLM planning phase can be run once and reused, or run per-prompt with constraints to maintain consistency, enabling efficient batch processing without per-image planning overhead.

Implements two specialized diffusion pipeline classes (RegionalDiffusionPipeline for SD v1.4/1.5/2.0/2.1 and RegionalDiffusionXLPipeline for SDXL) that extend the standard diffusers library pipelines to support region-specific prompt conditioning. During the diffusion sampling loop, different prompts are applied to different spatial regions of the latent representation, enabling fine-grained control over content generation in each region while maintaining global coherence through a base prompt and cross-region attention mechanisms.

Provides a unified Python interface (mllm.py) that abstracts over multiple MLLM backends — GPT-4 (via OpenAI API) and local models (via transformers/ollama) — allowing users to swap backends without changing downstream code. The abstraction handles API communication, response parsing, and parameter extraction, exposing a single get_params_dict() function that returns consistent structured outputs regardless of backend choice.

Implements an iterative composition refinement loop (IterComp) that generates an initial image, analyzes it with an MLLM to identify composition issues, and regenerates with refined regional prompts and split ratios. Each iteration feeds the previous image back to the MLLM for visual analysis, enabling multi-step optimization of spatial layout, object placement, and attribute binding without manual intervention or retraining.

Integrates ControlNet models (edge detection, pose, depth, etc.) as optional auxiliary conditioning inputs to the regional diffusion pipeline, allowing users to provide structural constraints (edge maps, pose skeletons, depth maps) that guide generation while regional prompts control semantic content. The integration preserves regional decomposition while adding structural priors, enabling generation that respects both spatial layout and visual structure.

Uses hand-crafted prompt templates (embedded in mllm.py and RPG.py) to guide MLLMs toward generating structured, parseable outputs with consistent formatting. Templates specify the desired output format (e.g., 'split_ratio: [0.3, 0.7]', 'region_1_prompt: ...'), enabling reliable extraction of parameters via regex or string parsing without requiring MLLM function calling or JSON schema enforcement.

Enables generation of images containing multiple distinct entities (e.g., 'a red cat and a blue dog') by decomposing the scene into per-entity regions with independent prompts that specify entity-specific attributes. Each region's prompt is isolated from others, preventing attribute confusion where properties intended for one entity bleed into another. The regional diffusion pipeline applies region-specific guidance to enforce attribute binding without cross-region interference.

Achieves spatial control and multi-region generation without modifying or fine-tuning the underlying diffusion model weights. Instead, it adapts pre-trained SD/SDXL models by modifying the inference-time conditioning mechanism (regional prompt injection into the UNet forward pass) and using MLLM-guided planning to structure the generation process. This enables high-quality generation with off-the-shelf models without the computational cost or data requirements of fine-tuning.

Provides a single Python API (RPG.py) that abstracts over multiple Stable Diffusion architectures (v1.4/1.5/2.0/2.1 and SDXL) with different pipeline implementations (RegionalDiffusionPipeline and RegionalDiffusionXLPipeline) but identical user-facing interfaces. Users specify model architecture once and the framework automatically selects the correct pipeline, enabling seamless model switching without code changes.