Pixal3D: Multi Views to 3D
Upload your own separate view images, or use one mixed multi-view turnaround sheet. The workflow prepares front, back, left and right views, then reconstructs a textured 3D model with Pixal3D Multi-View.
關於
Pixal3D: Multi Views to 3D turns a 4-view character or product turnaround into a textured 3D mesh. The Preprocess Multiview Images group takes either a single multi-view sheet or separate front, back, left, and right images. It can optionally remove the background (LoadBackgroundRemovalModel + RemoveBackground) to isolate the subject and improve consistency. CLIPVisionLoader provides vision conditioning from the views that drives the multi-view reconstruction.
您提供的內容
- 正在讀取工作流程
在 ComfyUI 中執行。 此工作流程無法在此頁面上執行。請在已載入所有內容的 Comfy Cloud 中開啟,或下載到您自己的 ComfyUI。
在 Cloud 中試用此工作流程尚無範例輸出。
問答
The workflow accepts a single mixed turnaround sheet or separate views. It works with fewer views, but geometry becomes more ambiguous. For best results, provide front, back, left, and right views with consistent scale and pose.
Ensure all views are aligned and similarly scaled, remove complex backgrounds (LoadBackgroundRemovalModel + RemoveBackground), increase KSampler steps, try a different seed, and keep CFGOverride/RescaleCFG at the provided defaults. Consistent, clean inputs yield more stable geometry.
Use DecimateMesh to reduce polygon count after VoxelToMesh. For textures, enable Trellis2UpsampleStage for higher fidelity and set your desired texture size in BakeTextureFromVoxel. Inspect results with PreviewImage before exporting.
VAELoader should load trellis_2_texture_vae_bf16.safetensors so the Trellis2 decoders (VaeDecodeStructureTrellis2/VaeDecodeTextureTrellis) can correctly decode shape and texture latents. CLIPVisionLoader provides vision conditioning from your views, and ModelSamplingSD3/KSampler control the diffusion sampling steps.