New Open-Source 3D AI Texturing Workflow is Here (Runs on 6GB VRAM)

Source Download this video
en-origen
May 10, 2026 Jul 13, 2026
Video preview
Share:

This video demonstrates a local 3D AI texturing workflow that delivers sharp details, true albedo, runs on just 6GB VRAM, and is entirely free. The workflow uses Flux to Klein to project high-quality views onto a mesh from multiple camera angles.

Installation and Setup ⏱ 0:00

  • Download the one-click install script from pixelminusartistry.com/textureprojectionworkflows
  • First set up Trellis 2 using the Trellis 2 installation guide
  • Download both the Trellis 2 GGUF script and Trellis 2 script from the Trellis 2 GGUF folder
  • Paste both into the ComfyUI Easy Install add-ons folder and execute them
  • Run the Trellis 2 GGUF model downloader.bat to download all needed models
  • Download the texture projection installer.bat from the same website and save it in the add-ons folder
  • Double-click the installer .bat file; it will handle custom nodes and dependencies
  • Launch ComfyUI (use ComfyUI.bat or ComfyUI flash attention.bat)
  • Drag and drop the workflow JSON files into ComfyUI; there should be no red nodes
  • Download required models using the markdown node links within the workflow
  • Single-Image Workflow Usage ⏱ 4:35

  • Upload a single 2D image (e.g., a dwarf)
  • Set target face number (recommended 100k), texture resolution (1024 for 6GB VRAM), and character name
  • In the texture projection render conditions node, set resolution (e.g., 1024) and adjust flux subnets for side, back, front views
  • Adjust camera elevations and copy values to the texture projection bake textures node
  • Run the workflow; first launch may take time to download models and initialize
  • Result: a textured 3D model with normal map, with notably sharper details and fewer shadows compared to default Trellis 2 output
  • Multi-View Workflow for Better Accuracy ⏱ 10:10

  • Open the texture projection multi-view workflow file
  • Upload multiple images (e.g., 4 views of a robot) to texture from all sides
  • Flux to Klein reads small details like rust and metalness from each view
  • Result is more accurate and detailed than single-image projection
  • Side-by-side in Blender: multi-view projection shows way more detail than default Trellis 2 texture
  • Retexturing Existing Models and Fixing Issues in Blender ⏱ 11:43

  • For existing GLB models, upload the model and paste the mesh path into the mesh input node
  • The workflow is faster (e.g., 119 seconds) and projects from each side
  • Note limitations: top and bottom may lack texture detail; adjust resolution to 4K for better quality
  • To fix projection issues in Blender:
  • - Go to Texture Paint tab

    - Use clone brush (select clone tool, Ctrl+click to copy area, paint over flaws)

    - Save the image after painting (Image > Save)

  • Workflow currently outputs only diffuse map; for roughness/metallic:
  • - In shader editor, use a hue/saturation value node to desaturate base color → roughness

    - Use a color ramp to create black/white metallic mask from base color

  • Roughness and metallic Lora are being worked on by Error X and the community
  • Key Takeaways

  • The workflow runs locally in ComfyUI on as little as 6GB VRAM and is entirely free.
  • It uses Flux to Klein to project high-quality views onto the mesh, generating true albedo without baked-in lighting.
  • Two versions exist: single-image and multi-view; multi-view provides more accurate texturing from all sides.
  • Existing GLB models can be retextured by providing the mesh path; the process is faster (e.g., 119 seconds).
  • Projection issues can be fixed in Blender using the clone brush in Texture Paint mode.
  • Conclusion

    This open-source workflow offers a major upgrade for local 3D texturing, with future improvements like roughness and metallic maps expected from the community.

    Ask AI about this video

    Visual Highlightsbeta