DEV Community

AI-KSK
AI-KSK

Posted on

Pixal3D: Turning One Image Into a GLB Asset Draft (Ten Subject Tests)

#comfyui #ai #tutorial

Image-to-3D has been around long enough that the interesting question is no longer "can it" but "what is it good for". This episode's answer, after ten subjects: Pixal3D is a high-fidelity draft machine for props, fantasy assets and scene objects, not a modeler replacement. Its distinguishing feature is pixel alignment - pixel features from the image are back-projected into 3D so the 2D-to-3D correspondence is tighter. This episode shipped no workflow JSON and no mirror, so the notes below are the author's test results, reported in the video.

Evidence note: no workflow JSON and no workflow/mirror link were included with this episode - only the YouTube and Bilibili videos. Everything below is the author's reported testing; no graph parameters exist to quote.

When to use it

  • You need a fast blockout of a prop, a game item or a fantasy object.
  • You want a GLB you can open in local 3D software and iterate on.
  • You are building a composite AI + 3D + video pipeline and need asset drafts, not final topology.

What is different in this workflow

  • Pixel alignment is the pitch: rather than guessing geometry from semantics, the model back-projects image pixels into 3D space, which tightens the 2D/3D correspondence.
  • Output is GLB, usable outside the tool.
  • It is a draft generator: the author's repeated framing is "3D blank/rough", good for blockout, not for rigging.

Step by step

  1. Pick a subject with a normal, roughly symmetric viewing angle - the author used stock assets and notes that odd angles hurt results.
  2. Run the image through Pixal3D (the author points to a one-click package linked from his Bilibili description; the RunningHub side also has a ComfyUI plugin for this kind of conversion).
  3. Open the GLB in your 3D software and inspect silhouette, material separation and back side.
  4. Keep it as a blockout; retopologize before rigging.

Parameter basis

  • No workflow JSON was shipped and no step/CFG/sampler values were published, so none are quoted here.
  • The only quantitative content is the ten-subject test set: character, metal coffee pot, bicycle, leather chair, robot, glass perfume bottle, backpack, treasure chest, plant, guitar.

Limits

  • Character: overall silhouette, clothing layers, cape, boots and arm proportions survive; face, hands and clothing detail get simplified; topology is very unlikely to be riggable.
  • Bicycle: thin structures (rims, spokes, chain, frame crossings) come out as a rough blank - not a usable mechanical model.
  • Robot / guitar: appearance and colour read correctly, fine mechanical detail merges.
  • Glass bottle: looks convincing, but transparency in the GLB is likely texture and surface effect, not physically accurate glass.
  • Input angle matters: the author recommends normal, roughly symmetric views.
  • It is a draft, not a replacement for a modeler.

Episode resources

Top comments (0)