Flux
A Black Forest Labs image-generation model family that turns text prompts into images using rectified-flow transformer backbones in latent space.
Flux is Black Forest Labs' image-generation model family. Public releases describe rectified-flow transformers that map text prompts to images in latent space, which is a different backbone choice than the U-Net denoisers common in early Stable Diffusion pipelines.
At a glance
- Family
- flux
- Source type
- Open weights
- Released
- August 2024
- Authors
- Black Forest Labs
- Source
- black-forest-labs/flux
- Modalities
- Text, Image, Multimodal
What It Is
Flux is an image-generation model family from Black Forest Labs (BFL), announced in August 2024 with the FLUX.1 suite and followed by later releases such as FLUX.2. Public materials frame Flux as a text-to-image system built around rectified flow and transformer-style denoising rather than as a chat or language model. The family ships through official inference repositories, Hugging Face model cards, and BFL-hosted APIs, so readers can inspect weights or run generation without treating Flux as a closed black box.Inputs And Outputs
Flux takes natural-language prompts as its primary input and produces raster images as its main output. Public FLUX.1 materials focus on text-to-image generation, while later FLUX.2 repositories describe additional editing-oriented workflows on top of the same broad image-generation stack. The immediate user-visible result is a finished image file rather than an embedding, caption, or token stream, which is why Flux belongs on the image-generation path even though prompts are text.Architecture
At a high level, Flux follows a latent image-generation pattern. A text encoder turns the prompt into conditioning features, a rectified-flow transformer backbone updates noisy latent representations over many steps, and a decoder renders pixels from the cleaned latents. Classifier-free guidance can steer how strongly the model follows the prompt during sampling. Compared with classic Stable Diffusion-era stacks that often paired a convolutional U-Net denoiser with CLIP text embeddings, Flux's public descriptions emphasize a diffusion transformer backbone and rectified flow training instead of treating a U-Net as the main denoising network.Generated
ImageImage
DecoderRectified Flow
Latent UpdatesDiffusion
TransformerCross-
AttentionClassifier-Free
GuidanceText
ConditioningText
PromptImage
Decoder to Generated
ImageRectified Flow
Latent Updates to Image
DecoderDiffusion
Transformer to Rectified Flow
Latent UpdatesCross-
Attention to Diffusion
TransformerClassifier-Free
Guidance to Rectified Flow
Latent UpdatesText
Conditioning to Cross-
AttentionText
Conditioning to Diffusion
TransformerText
Prompt to Text
Conditioning
Important Modules
The diffusion transformer block is the named module on this page because it carries most of the iterative latent update work. Cross-attention inside that stack is what lets prompt conditioning steer image structure while the model refines latents step by step.Training
Public FLUX.1 materials describe rectified-flow training with guidance distillation so the dev release can sample with fewer guidance-heavy steps at inference time. That training objective differs from next-token language modeling and from purely score-matching diffusion setups, even though Flux still belongs to the broader diffusion-model family because it learns to transform noisy latents into clean images.Training regimes
Linked papers
- Latent Diffusion Models
- Diffusion Transformers