MicroDecoder previews an AI render in a hundredth of a second

Generating an image locally means staring at a progress bar and finding out at the end whether the seed was any good. MicroDecoder is an attempt to make that wait visible, and it does it with seven files totalling less than 70 MB.

The set went up on Hugging Face on September 24, 2026 under the Apache 2.0 licence. Its author, Vlad Mandic, describes it as a micro-VAE: a small network whose only job is turning the data a diffusion model is working with into something a person can look at. He puts the reconstruction time at roughly a hundredth of a second per frame, which is fast enough to draw on every step of a render that takes minutes.

What MicroDecoder Actually Does

A diffusion model does not hold a picture while it works. It holds data that can be decoded into a picture. The decoder that turns that data into a finished image is expensive. The Fervor Creative AI briefing that flagged the release puts it at almost the cost of the generation step itself, which is why so many local setups show a progress bar instead of a preview.

MicroDecoder replaces that decoder with a much smaller one for viewing purposes only. It applies two adjustments that matter. The first is a correction for how far through the render the model is, because the working data at step four is far noisier than the data at step forty, and early previews would otherwise come out as noise. The second is deliberate blurring followed by an upscale. The card says both serve one aim: a reconstruction as fast as possible that stays usable and consistent at any noise level.

Why This Is Worth Fifteen Minutes

The card describes training a decoder for a new model in about fifteen minutes, and publishes the script that does it. The example run in the card, which trained the Qwen-Image-2.1 decoder, ran 350 epochs at 512 pixels over 400 samples, 360 for training and 40 held back for validation, and finished with a best validation PSNR of 35.83 dB. The briefing's view is that for a preview this is far better than it needs to be.

The practical gain is being able to abandon a bad generation at step four instead of waiting for step forty and then throwing it away. On a machine that takes two minutes a render, that changes how many ideas fit in an afternoon.

Where It Runs and What It Covers

MicroDecoder is wired into SDNext, the local generation interface Mandic maintains, where turning on the micro preview type routes the loaded model to a decoder and downloads it into a models/Preview folder. The seven files on the Hub are microdecoder-f1, -f2, -qwen, -qwen21, -sd, -sdxl and -wan21, and a file can serve a group of architectures rather than one model. In SDNext's current routing table, -f1 serves FLUX.1 and a long tail of other models, -f2 serves FLUX.2 and a few more, -qwen21 serves Qwen-Image-2.1, -sd serves Stable Diffusion, -sdxl serves SDXL and several others, and -wan21 serves Wan 2.1 and, among others, the original Qwen-Image. No entry in the table currently points to the -qwen file.

The example inference code is published as well, under eighty lines of SDNext code that take the model's working data plus how far through the render it is. Shapes and channels are read from the base model's VAE, so switching between covered architectures needs no configuration.

The Limits

The card describes the output as low-quality reconstruction intended for live preview, so the finished render still comes from the model's own decoder. There is no file for an architecture outside those groups (the table also lists SD3 and a MiniMax model, but the Hub has no file for either), though the training script is the answer to that if you are willing to spend the fifteen minutes and a few hundred images. As the briefing notes, the 35.83 dB figure is one run on one dataset, not a benchmark.

What It Says About Where the Work Is Going

The Fervor briefing's reading of the weekend is that no lab released a model and the useful releases were plumbing around the models people already have, with preview decoding the clearest case. The fix for a blind wait looked like it needed a better model; in the briefing's words, "It required a 10 MB file and fifteen minutes of training."

MicroDecoder is on Hugging Face under Apache 2.0, and SDNext downloads it automatically once the micro preview is switched on.