Turn One Photo Into PBR Texture Maps for Free: A New AI Add-On Makes Albedo, Roughness and Metallic, and One Map Needs Your Eyes
A free Hugging Face tool splits a product photo into the three maps a renderer needs. Here is a ten-minute routine to get a material into Blender, and the honest score behind the map you should not trust
Every photo of an object is a lie about its surface. The shine on a ceramic mug is the window behind the camera. The dark side of a leather bag is the absence of a lamp, not the colour of the leather. A renderer does not want any of that. It wants to know what the surface is, so it can light it fresh.
That is why 3D artists build materials from separate maps instead of photos. An albedo map holds the pure colour, with every shadow and reflection removed. A roughness map says, pixel by pixel, how glossy or matte the surface is. A metallic map says which parts are bare metal. Together they are called a PBR material, short for physically based rendering, and they are how Blender, Unreal, Unity and most product renderers decide how a surface reacts to light.
Making those maps by hand is slow, fussy work. A new add-on that appeared on Hugging Face on October 27 does a first pass from a single photo, for free, in a browser tab. Two of its three maps are a real head start. The third is barely better than a coin flip with loaded dice, and the author's own numbers show it.
What appeared, and why it is unusual
The tool is Mike0021/qwen-image-edit-2511-pbr-lora, an add-on for Qwen-Image-Edit-2511, an open image editing model you instruct in plain sentences. The add-on teaches that editor three new instructions. You give it a photo and one of these, word for word from the card:
- "Turn this photo into its albedo map: the base colour of every surface, with all lighting, shadows and reflections removed."
- "Turn this photo into its roughness map: grayscale, black where the surface is glossy, white where it is rough."
- "Turn this photo into its metallic map: white where the surface is bare metal, black everywhere else."
You do not need to install anything to try it. The author put it in a free Hugging Face Space, projtex-pbr, created the same night. Upload a photo, press Decompose, and you get all three maps back.
The unusual part is how it learned. An editing model normally learns from pairs of real photos. You cannot photograph a roughness map, so the author rendered his own training set in Blender: about 5,600 views of simple shapes, assemblies and everyday objects, wearing materials from the free MatSynth library (CC0 and CC-BY) and lit with free Poly Haven environment lighting (CC0). Because every view was rendered, the true colour, roughness and metal value of every pixel was known exactly. The add-on learned to look at a lit picture and subtract the light.
The licence is clean all the way down. The add-on is Apache 2.0 and so is the Qwen base model, which means you can use the maps in paid work.
How good is it, really
The author tested the add-on on 42 views of 14 standard sample assets from the Khronos group, the body behind the glTF 3D format, and compared it with two published research methods and a deliberately dumb baseline. That baseline is the key to reading the results.
The baseline assumes the whole object has medium roughness and contains no metal at all. It is what you would get if you skipped the AI and filled two maps with flat grey and black.
On the metallic map, the version the card recommends scores 0.74 on the author's accuracy measure. The dumb "no metal anywhere" guess scores 0.72. The add-on wins, but by two points, on a test the card does not fully define. Its metal error is lower than the baseline's and lower than both research methods', so it is not useless. It is also not something to ship without looking.
The card tells a second story about metal. The first version kept reading glossy paint as metal, which makes sense: shiny red enamel and polished steel both throw sharp reflections. The second version added training views where a quarter of the materials were glossy but not metal, and the card says it stops making that mistake. The benchmark table does not show that fix clearly, and the second version's roughness error is about 10 to 14 percent worse than the first's. So you are trading a little roughness accuracy for, the author says, fewer painted surfaces turned into chrome.
The albedo map is where the add-on looks strongest. It ties for the best colour score in the author's table and beats both research methods on it.
My position: treat the colour and roughness maps as a real head start, and treat the metallic map as a suggestion that you confirm with your own knowledge of the object. You already know whether the mug has a steel rim. The model is guessing.
The ten-minute routine
1. Shoot for the tool, not for Instagram
Even, soft light from several sides. No hard window reflections. A plain, flat background in one colour.
The background matters for a reason you would not guess. The Space pads every photo out to a square using the colour of the top-left pixel, per its code. If that corner is a busy shelf, the padding is a smear of shelf. A plain backdrop keeps the edges clean.
Fill the frame with the object. The model works at 768 by 768 pixels and then resizes back to your photo's size, so a small object in a big frame gets very few pixels of attention.
2. Run it in the Space
Upload the photo, leave the Steps slider at its default of 4, and press Decompose. The Space uses a fast companion add-on (Lightning, which cuts the model's work to four passes) and the recommended second-version file. You get albedo, roughness and metallic as three images. Save all three.
If a map looks muddy, the slider goes up to 8; try a higher setting and compare.
3. Check the albedo first
Hold it next to the photo. The highlights should be gone. The shadowed side of the object should be the same colour as the lit side. If you still see a bright window reflection baked in, reshoot with softer light rather than fighting it in software. The tool removes lighting; it cannot recover detail a hot reflection erased.
4. Check the metallic map against reality
Look at every white area and ask: is that bare metal? Painted metal, chrome-look plastic and glossy ceramic are not metal for rendering purposes. Paint over any wrong white areas with black in any image editor. This takes a minute and fixes the map's biggest weakness.
5. Build the material in Blender
In a Principled BSDF material, plug the albedo into Base Color. Plug roughness into Roughness and metallic into Metallic, and set both of those image nodes to Non-Color, so Blender reads them as data rather than as a picture. Leaving them on the default colour setting is the common mistake, because Blender then shifts the values as if they were a photograph.
6. Light it somewhere else
Swap the environment to an HDRI that looks nothing like where you took the photo. A sunny exterior if you shot indoors. This is the real test. A good material looks right in any light. If the object still carries a ghost of the original window, the albedo kept some of it; go back to step 3.
7. Going further: whole objects
The card says the add-on was built for a projection-texturing tool by the same author, where each generated view of a 3D model gets split into maps that are projected and baked onto the mesh. The single-photo routine above is the honest starting point. Projection onto a full model adds its own seams and alignment work.
Where it breaks
It learned from renders, not photos. Every training image came out of Blender. The card reports no test on real photographs. Your phone's lens, noise and colour processing are all new to it.
The test set is small. 42 views of 14 objects is a sanity check, not a verdict. Expect surprises on materials it never saw: fabric weaves, translucent plastics, wet surfaces.
The metallic map is close to a guess. Two points above "no metal anywhere." Check it every time.
Resolution tops out at 768 pixels. Fine for a prop in a scene, too soft for a hero close-up. Upscale with care, since upscaling a roughness map can invent texture that is not there.
It splits one view, not an object. The back of the mug you did not photograph does not exist in these maps.
No seed control in the Space. The seed is fixed, so pressing the button again on the same photo gives the same result. To get a different take, change the photo or the step count.
What this changes
For years the honest answer to "can AI make my materials" was "it can make a texture that looks nice in a thumbnail." This add-on is more useful than that, and more honest than most, because its author published the score that shows its weakest map barely beats a blank one.
That kind of candour is how you should work with it. Use it to skip the dull first hour of a material. Keep your own eyes on the part where the model is guessing. Shoot one object you know well this week, a kitchen knife or a painted mug, run the routine, and light it in a scene it has never seen. You will know within ten minutes whether it belongs in your texturing kit.