yanaimngvov

If you want to test Alibaba's new Qwen-Image-2512 model but don't have the GPU memory (or patience) to host it locally, we put up a free instance here: https://zimage.run

It currently supports Qwen-Image-2512 and Z-Image Turbo. No account needed.

It might be a bit slow if traffic spikes, but it works for quick prompt testing.

You can open another tab in your browser to generate a new image. Since there's no logged-in status, previously generated images cannot be retrieved after refreshing the page.

We will be implementing an update soon that will allow you to view your image generation history once logged in. Thank you for your feedback!

Hello HN,

I built zimage.run to provide a frictionless way to try out Alibaba's newly open-sourced Z-Image model (specifically the Z-Image-Turbo 6B variant).

While there are many FLUX/SDXL wrappers out there, I found Z-Image interesting because of its efficiency (6B parameters) and its ability to render legible text in both English and Chinese.

The site is dead simple: - Completely Free: No credits, no hidden costs. - No Login Required: Just type a prompt and hit enter. - Uncensored/Unlimited: I haven't put artificial limits on it (for now), so you can stress test the model's capabilities.

Tech Stack: Model: Alibaba Tongyi-MAI Z-Image-Turbo (6B) Frontend: Next.js Backend: ComfyUI

I'd love to hear your feedback on the generation speed and how you think this model compares to Flux or SDXL.

Try it here: https://zimage.run

You can try the website: https://sora2watermarkremover.net

It allows you to select the watermark area online, and the AI automatically processes it.

This is great, thanks for sharing. The open-source path is always appreciated. For anyone looking for a no-code, web-based alternative, I've also had good results with https://sora2watermarkremover.net

This is a technically impressive release, and its architecture reveals a lot about why it achieves such high quality in dialogue generation. It's not just another TTS model; it's a well-thought-out system.

Diving into its design, VibeVoice employs a sophisticated two-stage, cascaded generation process, which is key to its performance:

Semantic Generation (The "What"): It first uses a powerful LLM backbone (a Qwen2 variant) to convert the input text into semantic tokens. This is where the model's deep contextual understanding comes from. It's not just reading words; it's interpreting the structure of a conversation, which is why it handles complex, multi-speaker scripts (Speaker 0: ..., Speaker 1: ...) so effectively and maintains long-form coherence. Acoustic Generation (The "How"): The semantic tokens are then passed to a diffusion-based acoustic model. This is the core of its audio quality. Unlike older GAN-based vocoders, the diffusion process allows it to synthesize incredibly rich and natural-sounding audio with realistic prosody, intonation, and emotional cadence. This is a computationally intensive but sonically superior approach, and likely the reason for its quality.

The impact of this architecture is significant. It moves open-source TTS closer to the quality of proprietary leaders, especially for use cases that require more than single-sentence narration, like character-driven AI video or podcast prototyping.

For those interested in the technicals, here are the direct resources:

GitHub (Source Code): https://github.com/microsoft/VibeVoice

Hugging Face (Model & Config): https://huggingface.co/microsoft/VibeVoice-1.5B

Live Demo (for Dialogue Engine Testing): https://vibevoice.info/

One of the most fascinating (and challenging) parts of building this was seeing just how wildly different the "best" model can be depending on the document type.

For example, during testing, I found that Marker is an absolute champion for clean, single-column layouts like blog posts. But throw a dense, multi-column academic paper at it, and MinerU often produces a far superior, structured output with proper LaTeX. Then, for a complex invoice table, PP-StructureV3 frequently beats both of them.

This really solidified my belief that a "one-size-fits-all" parser is a myth. The future seems to be less about finding a single perfect model and more about building a quick, effective workflow for selecting the right specialist for the job. It's a classic "routing" problem, and this tool is my attempt at solving the first step of that puzzle.