black-forest-labs/flux-tencent models

flux-tencent by Black Forest Labs — AI Model Family

The flux-tencent family from Black Forest Labs represents a specialized collaboration with Tencent researchers, building on the renowned FLUX architecture through advanced techniques like Srpo (likely referencing Tencent's MixGRPO framework for enhanced preference alignment). This family addresses key challenges in AI image generation, such as improving prompt adherence, visual quality, and human preference alignment in diffusion models. It includes four optimized models across Text to Image and Image to Image categories: Tencent | Flux 1 | Srpo | Image to Image, Tencent | Flux 1 | Srpo | Text to Image, Tencent | Flux | Srpo | Image to Image, and Tencent | Flux | Srpo | Text to Image. These models leverage Tencent's innovations in training efficiency and adaptive denoising to deliver superior results for creative and professional workflows.

Drawing from Black Forest Labs' FLUX base—known for its 12B parameter scale, 57 DiT blocks, and support for high-fidelity 1024x1024 generation—this family integrates Tencent's expertise in human-aligned training, making it ideal for tasks requiring precise control and consistency.

flux-tencent Capabilities and Use Cases

The flux-tencent family excels in both Text to Image and Image to Image generation, with Flux 1 variants offering refined precision and base Flux models providing broader versatility. All models support core FLUX specs like 1024x1024 resolution, 28 inference timesteps via EDM sampler, and CFG scales around 3.5 for balanced guidance.

• Tencent | Flux 1 | Srpo | Text to Image: Generates photorealistic or artistic images directly from text prompts. Use case: Marketing teams creating campaign visuals. Example prompt: "A futuristic cityscape at dusk with neon lights reflecting on rainy streets, cyberpunk style, ultra-detailed." This model shines in complex scene composition with strong semantic alignment.

• Tencent | Flux 1 | Srpo | Image to Image: Transforms existing images based on text instructions, ideal for iterative editing. Use case: Graphic designers refining sketches into polished renders. Start with a base image of a landscape and prompt: "Convert this forest path into a magical enchanted woodland with glowing mushrooms and fireflies, high fantasy art."

• Tencent | Flux | Srpo | Text to Image: Handles diverse styles from the core Flux architecture, optimized for speed and creativity. Use case: Concept artists brainstorming ideas quickly for games or films.

• Tencent | Flux | Srpo | Image to Image: Enables seamless inpainting, outpainting, or style transfers. Use case: E-commerce pros updating product photos, like altering backgrounds on apparel shots.

These models integrate powerfully in pipelines: Use a Text to Image model to generate an initial asset, then chain to Image to Image for refinements—such as adding elements or adjusting lighting—creating end-to-end workflows for production-ready outputs. Technical highlights include support for 1024x1024 images, multi-step denoising (up to 50 steps for precision), and efficient attention mechanisms like aiter for scalable inference.

What Makes flux-tencent Stand Out

flux-tencent distinguishes itself through Tencent's MixGRPO framework, which integrates sliding-window mixed ODE-SDE sampling for up to 71% training speedup in human preference alignment, directly enhancing generation quality on FLUX baselines. This results in superior prompt adherence and visual consistency, as seen in benchmarks where FLUX variants outperform leaders like Kolors in HPSv3 categories for aesthetics, semantics, and appeal.

Key strengths include:

• Adaptive interaction denoising: Dynamically modulates text-image interactions across transformer blocks and timesteps, boosting quality without heavy fine-tuning.
• Efficiency and scalability: Lightweight inference (e.g., 6-hour training on 8 H100s for FLUX), sparse enhancement on crucial blocks, and high-res support up to 2K in related FLUX evolutions.
• Human preference optimization: Srpo techniques ensure outputs align closely with user intent, reducing iterations.

Ideal for professional creators (e.g., advertisers, game devs), researchers testing diffusion advancements, and developers building AI pipelines needing reliable, high-quality image synthesis. Compared to standard FLUX, flux-tencent offers finer control and faster alignment, bridging research-grade performance with practical speed.

Access flux-tencent Models via each::labs API

each::labs is the premier platform for accessing the full flux-tencent family through a unified, developer-friendly API at eachlabs.ai. Seamlessly integrate all four models—Tencent | Flux 1 | Srpo variants and Tencent | Flux | Srpo—into your apps with simple endpoints for Text to Image and Image to Image tasks.

Explore via the interactive Playground for instant testing with custom prompts, or leverage the robust SDK for production-scale deployments, supporting batch processing and pipeline chaining. Benefit from optimized inference, real-time monitoring, and easy scaling without managing infrastructure.

Sign up to explore the full flux-tencent model family on each::labs.