| --- |
| library_name: diffusers |
| tags: |
| - modular-diffusers |
| - diffusers |
| - krea2 |
| - text-to-image |
| --- |
| This is a modular diffusion pipeline built with 🧨 Diffusers' modular pipeline framework. |
|
|
| **Pipeline Type**: Krea2AutoBlocks |
|
|
| **Description**: Auto Modular pipeline for text-to-image generation using Krea 2: encode text -> core denoise (symmetric CFG) -> decode. |
|
|
| This pipeline uses a 3-block architecture that can be customized and extended. |
|
|
| ## Example Usage |
|
|
| [TODO] |
|
|
| ## Pipeline Architecture |
|
|
| This modular pipeline is composed of the following blocks: |
|
|
| 1. **text_encoder** (`Krea2TextEncoderStep`) |
| - Text encoder step that tokenizes the prompt(s) with the Krea 2 chat template, runs the Qwen3-VL text encoder, and stacks a fixed set of decoder-layer hidden states per token as the transformer's text conditioning. When `guidance_scale > 0` the negative prompt is encoded the same way for CFG. |
| 2. **denoise** (`Krea2CoreDenoiseStep`) |
| - Core denoising workflow for Krea 2 text-to-image: prepares the batch/latents/timesteps and the shared position ids, then runs the symmetric-CFG denoising loop, producing the denoised packed latents for the decoder. |
| 3. **decode** (`Krea2DecodeStep`) |
| - Step that unpacks the denoised packed latents back to the spatial grid, de-normalizes them with the VAE's per-channel statistics, and decodes them through the Qwen-Image VAE into images. |
|
|
| ## Model Components |
|
|
| 1. text_encoder (`Qwen3VLModel`): The Qwen3-VL text encoder. |
| 2. tokenizer (`AutoTokenizer`): The tokenizer paired with the text encoder. |
| 3. transformer (`Krea2Transformer2DModel`) |
| 4. scheduler (`FlowMatchEulerDiscreteScheduler`) |
| 5. vae (`AutoencoderKLQwenImage`) |
| 6. image_processor (`VaeImageProcessor`) |
|
|
| ## Configuration Parameters |
|
|
| is_distilled (default: False) |
| |
| ## Workflow Input Specification |
| |
| <details> |
| <summary><strong>text2image</strong></summary> |
| |
| - `prompt` (`str`): The prompt or prompts to guide image generation. |
| |
| </details> |
| |
| |
| ## Input/Output Specification |
| |
| **Inputs:** |
| |
| - `prompt` (`str`): The prompt or prompts to guide image generation. |
| - `negative_prompt` (`str`, *optional*): The negative prompt(s) for CFG. |
| - `guidance_scale` (`float`, *optional*, defaults to `4.5`): CFG scale; the negative prompt is only encoded when this is > 0. |
| - `max_sequence_length` (`int`, *optional*, defaults to `512`): Maximum sequence length for prompt encoding. |
| - `num_images_per_prompt` (`int`, *optional*, defaults to `1`): The number of images to generate per prompt. |
| - `latents` (`Tensor`, *optional*): Pre-generated noisy latents for image generation. |
| - `height` (`int`): The height in pixels of the generated image. |
| - `width` (`int`): The width in pixels of the generated image. |
| - `generator` (`Generator`, *optional*): Torch generator for deterministic generation. |
| - `num_inference_steps` (`int`, *optional*, defaults to `28`): The number of denoising steps. |
| - `sigmas` (`list`, *optional*): Custom sigma schedule (defaults to a linear ramp). |
| - `attention_kwargs` (`dict`, *optional*): Additional kwargs for attention processors. |
| - `output_type` (`str`, *optional*, defaults to `pil`): Output format: 'pil', 'np', 'pt'. |
|
|
| **Outputs:** |
|
|
| - `images` (`list`): Generated images. |
|
|
