This weekend I used a firefly photo of a stone wall in IT3D and it came out incredible after trying a couple of different images. No I’m trying to create cedar shake roof panels via the same process, I’m using the highest quality tool offered and consistently no matter what Image I use, singles are always offset by about half from one row to the next and every image shows that distinctly. After four tries it keeps giving me perfectly square tiles with no offset. looks literally nothin g like the images I’ve tried.
Welcome to AI! 
Even proponents of these kinds of tools admit they’re far from perfected. Here’s a great evaluation of one of the others: I tested PrintU from Bambu Lab - an AI tool for quickly creating 3D-printed bobbleheads from photos
For your specific need, I’d hit one of the 3D model meta-search engines like Yeggi or STL Finder, and also try synonyms like “shingles”.
my surface not flat, from image to 3d, anyone can helo me to fix that surface ?
I have pretty good luck with this tool I have created many 3d models from AI generated images. But the one area that I have never been able to get working is if there is any lettering in the image. Even if it is obvious raised letters. No matter what model I use the lettering is always jumbled. Letters missing, letters merged, spacing changed. Is this a known issue or is there a way to make sure that any letters maintain the original context.
I had that issue and still do, so I now leave an area where I want to put the text and use the Bambu Slicer to add the text later. That way you can select any font you want, size, and bold/italic, also whether or not you want the lettering raised or embedded. I am new to the 3D creation/AI/GenAI/Print world, so found a way that works for me 
Create a single row stretched top-to-bottom so about 25% can be overlaid just like real shingles. Then print a bunch after overlapping them in the edit box, or print and then glue them together as single sections staggered half a shingle
Почему я не могу скачать сгенерированные модели? При экспорте выводит предложение с выбором формата: stl, 3mf, glb, при выборе формата и нажатии кнопки Download, происходит бесконечная загрузка и файл модели не скачивается
The giant problem is that ai thinks depth is darker colors, not physical dimensions. basically, a hole in a flat surface you see it as a hole, but ai sees it like a flat wall with a black stain
Is an awesome tool. I use TripoAI and I love it.
I’ve had two separate instances of image to 3D making a cartoon dragon with two tails, specifically when the dragon is sitting. It’s frustrating you can’t give prompt or feedback in the tool so I can tell it “only one tail please.”
I’ve had 2 files - a 3d dimensional card, both very similar in style and color. and used tripoAI, First one on 6-4 turned out great printed all one color, all other file conversion attempts since; remove the background or it has added multiple colors. As colors are deleted, so goes the detail. Quite strange and I cannot get the quality back like in the first file. any ideas?
yes, you would need to feed it a different view or different version of photo… tripo ai is the best model generator…
I’d like to share some feedback about the Image to 3D feature in MakerLab.
I understand that generating a 3D model from a single image is a complex task and that AI has to reconstruct parts of the object that are not visible in the photo. However, I find it difficult to justify spending a significant number of credits before knowing whether the final result will actually be usable.
For example, Tripo 3.1 costs 36 credits and Hi3D costs 50 credits. On top of that, users are asked to choose between different mesh types, but it is not always clear which one will produce the best result for a specific image. As a result, we end up spending credits without any real confidence that the generated model will meet our needs.
Even when starting with a high-quality image, the outcome can vary dramatically depending on the selected model and mesh option. In practice, it often feels like a gamble: the result may be excellent, or it may be completely unusable, requiring additional credits for another attempt.
I believe the user experience could be improved by adding features such as:
-
A free low-resolution preview before spending credits.
-
Multiple mesh variations generated from a single credit purchase.
-
A confidence score indicating how reliable the expected result is.
-
Partial credit refunds for clearly unusable generations.
At the moment, the process feels too much like trial and error for the number of credits being charged.
Has anyone else had the same experience with Image to 3D?
Same here. The day I saw that you needed to spend credits for generating a model i was like “okay, maybe it’s not such hit or miss as before and that’s why its more expensive”, but I got very disappointed. The results were identical to before and you couldn’t even see a preview without spending credits.
I would love a solution for this @BambuLab
Bambulab Ticket response..
Thank you so much for this thoughtful feedback on Image to 3D in MakerLab — we understand how frustrating it can feel to spend a large number of credits on options like Tripo 3.1 or Hi3D, and on mesh choices that aren’t always clear, before knowing whether the result will actually be usable. Your ideas for a free low-resolution preview, multiple mesh variations from one credit spend, a confidence score, and partial refunds for clearly unusable generations are specific and very helpful. We’ve shared them with our product team for evaluation, and we’ll keep assessing how we can make this experience less of a trial-and-error process. We truly value feedback like yours and would love to hear more of your ideas anytime. Thanks again for helping us make MakerWorld better.
Make My Statue should be dropped back to version before the new one. It was better at getting a correct model than this new version is.
now we loose credits for just creating instead of downloading, but if the model is not good enough now you loose your credits for a model you can not use. Please go back to the old ways and credit us for exporting the model not to create a model
Worth being specific in that ticket, because “add a preview” is easy to file and easy to defer.
These pipelines build the geometry in stages. A coarse mesh exists well before the high resolution pass and before any texture work, and that coarse mesh already answers the two questions that decide whether you’ll use the result: did it understand the object as one object, and did it invent the unseen side in a way that makes sense. That stage costs a fraction of a full generation. So a low res preview isn’t a discount on the product, it’s showing you something the pipeline already produced on the way. That’s a much harder thing to defer than a general request for better UX, and I’d put it in those words if anyone follows up on the ticket.
On being charged at creation rather than at download, I doubt that one is coming back. Bambu pays the upstream provider the moment the generation runs, whether or not you export it, so the old behaviour meant they ate the cost of every failed attempt. The version of that ask which can actually win is domored’s partial refund, or a cheap rerun on the same image at reduced cost.
And since nobody answered the mesh option question: the spread between those options is much smaller than the spread coming from your photo. If a generation comes back unusable, rerunning it with a different mesh setting mostly gives you a differently unusable model. The gamble is real, but you’re not losing it at the dropdown.
Where you can actually shift the odds is the input, and it’s worth knowing what these models are weak at before you spend fifty credits. They have no depth information, they infer shape from shading, so a dark stain and a real recess look the same to them and a flat straight on photo gives them almost nothing to work with. A three quarter angle with even light and a plain background does far better than a perfect straight on product shot. Anything thin and protruding, straps, handles, thin limbs, is where they fail most often, and lettering below a certain size comes back as a smooth surface with the letters painted on, so leave the text off and add it in the slicer.
None of that makes it deterministic. It does move the hit rate more than anything in the options panel will.
(For context, I work on PHORM, which is in this category.)