Discover Wan 2.2, FLUX.1 Krea, and Qwen-3-Image — Open Source Power
🌀 Meet the New Models
Wan 2.2
Developed by researchers behind the open-source video model series “Wan,” Wan 2.2 pushes the boundaries of text-to-video generation. It leverages a powerful MoE (Mixture of Experts) architecture to create cinematic-quality short clips from simple prompts—often at 720p, 24fps.
License & Access
Wan 2.2 is fully open-source—users can freely access model weights, code, and documentation on GitHub, with no license fees for most use cases. Commercial licensing is available for enterprise-level support
NSFW Friendliness
Some third-party platforms note that the model does not support NSFW content—for instance, a version on one interface specifies “This model does not support NSFW content” wan2.ai. The open-source GitHub source doesn’t explicitly mention NSFW policy; it’s safer to assume it’s not NSFW-friendly by default.
Performance
Wan 2.2 delivers high-quality text‑to‑video (T2V), image‑to‑video (I2V), and hybrid generation at up to 720p @ 24fps, optimized for consumer GPUs like RTX 4090 GitHubVideomakerVidful. Its Mixture‑of‑Experts (MoE) architecture and high‑compression VAE enable efficient, cinematic-quality outputs, outperforming leading commercial video models in realism, motion consistency, and prompt accuracy
Get started with Wan 2.2 by watching the full tutorial:
- WAN 2.2 Tutorial: Massive Upgrade to AI Video Generation
- Get the most out of Wan 2.2 – Image & Video Generation – FASTER!
FLUX.1 Krea [dev]
A collaborative project from Black Forest Labs and Krea, FLUX.1 Krea [dev] is a text-to-image model designed with an eye for photorealism and natural aesthetics. It’s known for producing more organic, less “AI-polished” results that are especially strong in faces and fashion.
License & Access
FLUX.1 Krea [dev] is an open‑weights text‑to‑image model—weights are freely available (e.g., on Hugging Face), while commercial licenses can be obtained through BFL’s licensing portal or via API partners like Replicate
NSFW Friendliness
The developers have incorporated pre‑ and post‑release mitigations to minimize generation of unlawful or unsafe content, implying NSFW content is likely restricted or filtered.
Performance
FLUX.1 Krea [dev] excels with photorealism and a unique, more natural aesthetic that consciously avoids the “oversaturated AI look.” It matches or even outperforms previous open models and is on par with closed‑source alternatives in human preference tests
Qwen‑3‑Image
Part of the broader Qwen-3 release by Alibaba’s team, Qwen‑3‑Image is a new multimodal capability integrated into their large language model suite. Though still evolving, it shows promise as an image-generation model embedded into conversational AI, capable of generating images alongside text reasoning.
License & Access
The Qwen‑3 model family—including both dense and MoE variants up to 235B parameters—is released under the Apache 2.0 license, allowing broad usage and redistribution QwenOllamaAge of LLMs. For commercial use of Qwen language models in high‑traffic scenarios, some tiered licensing may apply—but for lower‑scale scenarios, free usage is allowed
NSFW Friendliness
No explicit mention of NSFW filtering for image‑generation specifically. However, the Qwen ecosystem (e.g., in Qwen Chat) includes image support; absent concrete NSFW policy, it’s safest to assume neutral stance, possibly allowing NSFW unless explicitly restricted.
Performance
Although Qwen‑3 is a powerful LLM family, documentation indicates image generation support exists in Qwen Chat, especially for the MoE variants, though actual open‑source model weights may not directly support vision input yet QwenHugging Face. Performance appears promising, but concrete benchmarks for image output quality on KaijuGen are not yet clear—suggesting early-stage capability with potential room to mature.
ACCESS AND GENERATE ON KAIJUGEN
