A 10 MB File Lets You Watch an AI Image Form, So You Can Kill a Bad One at Step Four
Local image generation has been a blind wait for three years. The fix turned out to be small, and somebody just published how to make your own.
Here is a thing nobody defends and everybody tolerates. You type a prompt, you press generate, and then you stare at a progress bar for forty seconds with no idea whether the composition is right, whether the hands came out, whether the model heard the word "profile." At the end you find out. If it is wrong you change one word and do it again.
The reason is dull and mechanical. An image model does not work in pixels. It works in a compressed form of the picture, something like a heavily abbreviated sketch, and turning that sketch into something a human eye can read is a separate step at the end done by a separate piece of software. That final step is expensive. Running it at every stage of the render would cost you most of the speed you were generating locally to get.
So the picture stays hidden while it forms. Everybody accepted this, because the obvious fix looked like it needed a faster, better decoder, and nobody was going to build one of those for free.
A file published on Wednesday, updated Friday, suggests the fix was much smaller than that.
What actually shipped
MicroDecoder is seven files totalling about 68 MB, one per group of model architectures, released under Apache 2.0. Each one takes the model's half-finished internal sketch and turns it into a rough but recognizable picture in roughly a hundredth of a second. Not the final picture. A preview.
It is wired into SDNext, a local image generation app with around 7,300 stars on GitHub, where it downloads itself on demand and draws to your screen at every step of the render. The author, Vlad Mandic, maintains that app, which tells you what this was built for: he was annoyed by the same blind wait you are.
The honest framing is that fast previews are not new. A tool called TAESD has done the same job since July 2023, it is MIT licensed, it is tiny, and it appears in something like a hundred public demos. If you have run Stable Diffusion locally with previews turned on, you have used it. So the interesting question is not whether a small preview decoder can work. It obviously can.
The interesting question is why you are usually still waiting.
The part that is actually new
Two things.
First, the new decoders know how far through the render they are. The model's sketch at step three is mostly noise, and at step thirty it is nearly a finished picture. Feed both through a decoder that treats them identically and the early previews come out as garbage that tells you nothing. MicroDecoder takes the render's progress as an input and corrects for it, which is why the preview at step four is worth looking at rather than worth ignoring.
It also deliberately blurs its output and then scales it up. That sounds backwards until you think about what a preview is for. A fast, honest blur reads correctly to a human eye. A fast, confident sharpness that is wrong reads as a bad image, and you kill a good generation on the strength of it.
Second, and this is the part I think matters more, the training script is published along with the weights. The card says fifteen minutes per model. The example run in the card trained on a 400-image set at 512 pixels, 360 of those images used for training and 40 held back to check the result, over 350 passes, and finished at 35.83 dB reconstruction quality. For a preview that is far better than it needs to be.
That single fact changes who is waiting for whom.
Why the recipe matters more than the files
The problem with every preview decoder before this one is that somebody has to make one for each new model, and then publish it, and then your app has to know about it. New image models arrive weekly now. Each one uses a different internal format. Each one therefore needs its own preview decoder, and until one exists, you are back to the progress bar.
TAESD solved that for the model families of 2023 and 2024. It could not solve it for models nobody had built yet.
Publishing the recipe is the difference between "here are seven decoders" and "here is how you stop needing me." If you are using a model outside the seven, you gather a few hundred images, point the script at your model, and come back after a coffee. That is a real answer. It is also, to be blunt, a shape of release I wish more people copied: the artifact plus the means of production, not just the artifact.
Put this into practice
If you use SDNext, this is a settings toggle. Turn the preview type to Micro. It works out which file matches the model you have loaded, pulls it into models/Preview, and starts drawing. There is nothing to configure, because the file itself carries the shape information the app needs. Total cost: one setting and a 10 MB download.
If you use something else, the code that runs it is under eighty lines and takes two things: the model's current internal sketch, and a number between zero and one saying how far through the render you are. That is a small enough surface to port into another app or a custom script in an evening. The seven published files cover FLUX.1 and FLUX.2 and a long tail of architectures that route to the same two files, plus Qwen-Image-2.1, Stable Diffusion, SDXL, and Wan 2.1. Check the app's routing table rather than guessing from the filenames, because the mapping is not one file per one model name.
If your model is not covered, the training run is the shape of a lunch break. You need a few hundred images, the name of your model's decoder class, and its repository path. The script encodes your images once, then trains the small network against them, saving whichever version scores best on the held-back set. Defaults are sensible: 500 images maximum, 512 pixels, 200 passes.
What to actually do with it. The payoff is not speed. Your renders take exactly as long as before. The payoff is that you can abandon one. Watch the first four or five steps. If the composition is wrong, if the subject is facing the wrong way, if the model has given you three people when you asked for one, stop it there and change the prompt. On a forty-second render that is thirty-five seconds back, every time you are wrong, and when you are exploring you are wrong a lot.
That habit is worth more than it sounds. Most of the frustration of local generation is not that renders are slow. It is that the feedback loop is long enough that you stop experimenting and start guessing.
Where it breaks
The author calls it low quality reconstruction and he means it. This is a preview and never an output. Do not judge fine detail, do not judge text, do not judge skin. Composition, colour, pose, count, layout: those you can judge. Anything finer, you wait.
The seven published files do not cover everything, and the routing is less tidy than the filenames suggest. Two of the files each serve a dozen or more architectures grouped together, and at least one common model type routes to a file whose name does not match it. If your previews look wrong, check the routing table before you assume the decoder is broken.
Every number in the card is one run on one dataset by one person. The 35.83 dB figure is his measurement of his own training run, not a benchmark anybody else has reproduced. The fifteen minute claim is his too, and it will depend entirely on your card and how many images you feed it. Treat both as an order of magnitude, not a promise.
And there is a quieter cost nobody mentions. A live preview invites you to judge an image before it exists. I have caught myself killing generations at step five that would have resolved fine, because the blur happened to read as a mistake. The preview is honest about being rough, but your eye is not. Give it a few days before you trust your own step-four instincts.
What I would do with it
If you generate locally, install it today. It is free, it is Apache 2.0, it is a single setting in SDNext, and the worst case is you turn it off again.
If you maintain a generation tool, the more interesting move is to copy the pattern rather than the file. Ship the artifact and ship the means of making more of them. The seven decoders will be stale in six months. The training script will not be.
And if you are running a model that nobody has made a preview for, you now have a choice you did not have on Tuesday. You can wait for someone else, or you can spend a lunch break. That is a small thing, but it is the kind of small thing that stops being a complaint without anybody announcing it.