Home Creators Posts Import Register Favorites Logout

Downloads

Content

Microsoft TRELLIS 2 premium app generates top quality 3D models with mesh and texture from images, with libraries precompiled for RTX 1000 to 5000 series GPUs.

Ultimate premium app with 1-click Windows and Linux installers, including RunPod and Massed Compute, and automatic character rigging with UniRig. Exports GLB, glTF, OBJ, PLY and STL files and runs each stage in its own subprocess for zero VRAM leftover.

Trellis 2 from Microsoft is the very best newest Image-to-3D generator with both mesh and amazing texture quality.

Patreon exclusive posts index to find our scripts easily, Patreon scripts updates history to see which updates arrived to which scripts and amazing Patreon special generative scripts list that you can use in any of your task.

Join discord to get help, chat, discuss and also tell me your discord username to get your special rank : SECourses Discord

Please also Star, Watch and Fork our Stable Diffusion & Generative AI  GitHub repository and join our Reddit subreddit and follow me on LinkedIn (my real profile)

=======

  • Latest installers zip file : Trellis2_v6_3.zip

  • This is an ultra advanced 1-click to install Trellis 2 app with so many features. Check below to see all features. I spent literally days to make this app and pre-compiled all the libraries to make it work otherwise it is impossible to install either or Windows or Linux. Also all models are downloaded into \models sub folder and loaded from there

    • I had to compile for both Windows and Linux : cumesh, flash-attn, flex-gemm, nvdiffrast, o-voxel (upgraded with more features by me), sageattention, xformers - all supporting all GPUs starting from RTX 1000 to 5000 series with all possible CUDA archs setup

  • This is early release and hopefully I will test on all platforms more and fix remaining errors if there are any

  • This model is VRAM hungry therefore you can reduce resolution and play with parameters

    • We are running everything as sub-process and minimizing the VRAM usage but still model is heavy since it is amazing quality

  • Hopefully I will make a full tutorial asap

    • First try some of the samples provided at the very bottom of the gradio app

    • Generated models looks better in Blender compared to Gradio

  • First click Cenerate and once done click Extract GLB to generate glb , gltf , obj , ply , stl files

    • Even if you get OOM at preview part of the Generate process, still click Extract GLB and it should work

17 January 2026 V4

  • Cloud installers fixed & improved

Windows Requirements - Super Important

Massed Compute (Recommend Cloud) :

RunPod (Cloud):

All the new features I did as below compared to official repo

TRELLIS.2 Premium App - New Features & Enhancements

🏗️ Core Architecture & Memory Management

Subprocess Stage Processing

  • Zero VRAM leftover mode: Each pipeline stage runs in a fresh Python subprocess, releasing all GPU memory between stages

  • Eliminates CUDA OOM errors: Perfect for high-resolution generation (1536+, 2048) and batch processing

  • Toggle-able: Can switch between subprocess mode and in-process mode

  • Graceful cancellation: Active subprocess stages can be immediately terminated

  • Cross-platform compatibility: Works on Windows, Linux, macOS with proper process isolation

Advanced Output Management

  • Auto-indexed run folders: Every generation saved to outputs/0001/, outputs/0002/, etc. (never overwrites)

  • Complete artifact persistence: Saves all intermediate files for debugging and reproducibility

    • 00_input.png

      - Original uploaded image

    • 01_preprocessed.png

      - After background removal and cropping

    • 02_cond_512.pt

      - 512px image embeddings

    • 03_cond_1024.pt

      - 1024px image embeddings (cascade only)

    • 04_coords.pt

      - Sparse structure coordinates

    • 05_shape_slat.npz

      - Shape latent

    • 05_shape_res.json

      - Resolution metadata

    • 06_tex_slat.npz

      - Texture latent (if texture generation enabled)

    • 07_preview/

      - Multi-view preview renders (JPEGs)

    • 07_preview_manifest.json

      - Preview metadata

    • 08_extract/

      - Final GLB and optional export formats

    • logs/

      - Per-stage subprocess logs

    • running_logs.txt

      - Consolidated real-time generation log

    • run.json

      - All generation parameters for reproducibility

Real-Time Logging System

  • Dual output: Logs stream to both UI overlay and disk simultaneously

  • Per-stage logs: Each subprocess stage writes detailed logs to outputs/<run_id>/logs/<stage>.log

  • Consolidated view: running_logs.txt combines all stages for easy debugging

  • Smart truncation: UI logs auto-trim to prevent browser memory issues

  • Timestamped entries: Every log line includes [HH:MM:SS] timestamp

🎨 User Interface Enhancements

Preview Panel Improvements

  • Larger preview: 820px height vs 722px (better visibility)

  • Progress overlay: Status and logs appear on top of preview (no sidebar clutter)

  • Semi-transparent log box: Dark overlay with syntax-highlighted monospace font

  • Auto-hide on completion: Overlay disappears when preview is ready

  • Fullscreen support: 3D Model viewer can go fullscreen via CSS

Control Buttons & Safety

  • Two-step cancel button: First click arms, second click confirms (prevents accidental cancellation)

  • Open outputs folder: One-click OS-native file explorer (Windows Explorer, macOS Finder, Linux file manager)

  • View Extracted GLB: Quickly switch to 3D viewer after extraction

  • Visual feedback: Buttons update state based on active operations

Configuration Presets

  • Save current settings: Name and save all UI parameters as JSON presets

  • Load presets: Instantly restore all sliders, checkboxes, dropdowns

  • Preset dropdown: Quick access to all saved presets

  • Delete presets: Clean up unused configurations

  • Reset to defaults: One-click restore factory settings

  • Auto-load last used: Remembers your last preset across sessions

  • Storage location: presets/<name>.json

Help System

  • Comprehensive guide tab: Full documentation of every parameter

  • Real-world examples: Practical use cases for each setting

  • VRAM guidance: Clear explanations of memory-intensive options

  • Workflow tips: Step-by-step quick start instructions

  • Settings explanations: What each slider/checkbox actually does

🖼️ Image → 3D Pipeline Features

Resolution & Quality Options

  • Custom resolution support: Enter any resolution ≥ 512 and divisible by 128 (768, 1280, 1920, etc.)

  • Preset resolutions: 512 (fast), 1024 (balanced), 1536 (high quality), 2048 (maximum detail)

  • Smart pipeline selection: Automatically chooses direct or cascade based on resolution

  • Force high-res conditional: Use 1024px embeddings even for 512px generation (quality boost)

Texture & Material Controls

  • Skip texture generation: Generate shape-only models (faster, lower VRAM)

  • Texture size: 1024, 2048, 4096 (baked texture resolution)

  • Per-stage guidance intervals: Fine-tune when CFG is applied (start/end sliders for SS, Shape, Texture)

  • Advanced sampling params: Full control over guidance strength, rescale, steps, rescale_t for all 3 stages

Memory Optimization Options

  • Low VRAM mode: Keeps models on CPU when not in use

  • Chunked triangle processing: Reduces memory spikes during generation

  • Tiled mesh extraction: Processes mesh in tiles to avoid OOM

  • Max tokens slider: Control cascade memory usage (adjust for high-res OOM)

  • Subprocess stage isolation: Complete VRAM release between stages

Export & Mesh Processing

  • Multiple export formats: Auto-save GLB, GLTF, OBJ, PLY, STL (checkbox selection)

  • Decimation target: 100k - 9M faces (default 1M for high quality)

  • Remesh method: dual_contouring (fast) or faithful_contouring (better topology, requires deps)

  • Simplify method: cumesh (GPU-accelerated) or meshlib (CPU-based alternative)

  • Prune invisible faces: Remove hidden geometry for cleaner exports

Preview Rendering

  • Multi-view snapshots: 8 angles with smooth slider navigation

  • 6 render modes: Normal, Clay, Base Color, HDRI Forest, HDRI Sunset, HDRI Courtyard

  • PBR rendering: Full physically-based rendering when nvdiffrec_render available

  • Fallback shading: Simple lambert shading when PBR unavailable (no hard crash)

  • Incremental progress: Per-view progress updates during rendering

  • Preview recovery: If preview fails but latents exist, can still extract GLB

📦 Batch Processing System

Batch Operation Features

  • Folder-based input: Process entire directories of images

  • Supported formats: PNG, JPG, JPEG, WEBP, BMP, TIF, TIFF

  • Smart output structure: Each image saved to outputs/<sanitized_filename>/

  • Resume capability: Skips folders that already exist (safe to re-run)

  • Per-image seeding: Randomize seed applies per-image or use same seed for all

Batch Progress & Monitoring

  • Real-time statistics: Processed / Skipped / Failed counts

  • ETA calculation: Estimated time remaining based on average per-image time

  • Live status log: Scrolling log with per-image progress

  • Current item display: Shows which image is being processed

  • Graceful error handling: Failed images don't stop the batch

Batch Settings

  • Same settings as single mode: Uses all UI parameters (resolution, guidance, etc.)

  • Preview skipping: Batch mode skips preview generation for speed (direct to GLB)

  • Custom output folder: Optionally specify different output location

  • Path support: Handles relative paths, absolute paths, quoted paths with spaces

  • Cross-platform paths: Works on Windows (D:\folder), Linux (/home/user/folder)

🎨 Texturing Tab (Standalone Pipeline)

Mesh Texturing Features

  • Upload existing mesh: Add textures to untextured models (PLY, OBJ, GLB, GLTF)

  • Reference image guidance: Use an image to guide texture appearance

  • Resolution options: 512, 1024, 1536 (same as Image→3D)

  • Texture size control: 1024, 2048, 4096 baked texture resolution

  • Seed control: Reproducible texturing with seed input

Advanced Texturing Controls

  • Guidance strength: Control how strongly the reference image is followed

  • Guidance rescale: Reduce over-saturation artifacts

  • Guidance interval: Start/end sliders for when CFG is applied

  • Sampling steps: Quality vs speed tradeoff (1-50 steps)

  • Rescale T: Sampler stability parameter

📂 View 3D Files Tab

3D File Viewer

  • Standalone viewer: Preview 3D files without running ML pipeline

  • Supported formats: GLB, GLTF, OBJ, PLY, STL

  • Fullscreen mode: Dedicated fullscreen button for immersive viewing

  • Drag & drop: Easy file upload

  • No processing: Instant preview (no GPU computation)

🛡️ Error Handling & Resilience

Cancellation System

  • Session-based tracking: Per-user cancellation state (multi-user safe)

  • Graceful subprocess termination: SIGTERM → wait → SIGKILL escalation

  • Process group killing: On POSIX systems, kills entire process tree

  • Batch vs all: Cancel only batch or cancel everything including active subprocess

  • No sticky cancels: Flags auto-clear when no work is active

Recovery & Fallback

  • Preview failure recovery: If preview OOM but latents saved, can still extract GLB

  • Visual error reporting: Preview panel shows error with recovery instructions

  • Enable Extract button: Even if preview fails, extraction works if latents exist

  • Detailed error messages: OOM errors show context and suggest fixes

  • Graceful degradation: Missing optional deps (nvdiffrec_render, faithcontour) don't crash app

State Management

  • Running task tracking: Knows which operations are active per session

  • Active subprocess registry: Tracks which stage subprocess is running

  • Session cleanup: Auto-cleanup temp files on session end

  • Lock-protected state: Thread-safe cancellation flags and subprocess tracking

🔧 Developer & Reproducibility Features

Reproducibility Tools

  • RNG state preservation: Subprocess mode preserves torch random state across stages

  • Parameter logging: Every run saves complete settings to run.json

  • Artifact versioning: All intermediate tensors saved for inspection

  • CSV-ready structure: Batch runs create organized folder structure

Utility Infrastructure

  • subprocess_utils.py: Shared utilities for path handling, run allocation

    • allocate_run_dir()

      : Thread-safe run folder creation with auto-indexing

    • next_indexed_path()

      : Find next available numbered path

    • ensure_dir()

      : Safe directory creation

    • safe_relpath()

      : Cross-platform relative path display

  • subprocess_stage.py: Worker process implementation with 7 stage handlers

  • Environment setup: Auto-configures CUDA allocator, TF32, module loading

Capability Detection

  • o_voxel auto-install: Checks for CUDA extension and installs if missing

  • nvdiffrec_render detection: Enables PBR rendering only if available

  • Faithful contouring check: Shows remesh option only if deps installed

  • Platform detection: Adjusts subprocess handling for Windows vs POSIX

🎯 Quality of Life Improvements

Path & File Handling

  • Smart path resolution: Handles ./relative, C:\absolute, quoted paths

  • Safe folder names: Sanitizes filenames (removes <>:"/\|?*)

  • Reserved name protection: Avoids Windows reserved names (CON, PRN, NUL, etc.)

  • Cross-platform compatibility: Works on Windows, Linux, macOS

UI Polish

  • Tips overlay: Hoverable help icon with usage tips

  • Active state indicators: Buttons and mode icons show current state

  • Smooth transitions: CSS animations for mode switching

  • Monospace logs: Code-style font for better readability

  • Color-coded status: Success (green), warning (yellow), error (red)

Related posts

All posts in this topic: Image to 3D Generation Apps · New to SECourses? Start Here: Best and Most Popular SECourses Posts

Files

Comments

Nedo Braun

thanx Doc! this is superb!

BigVI

Great, thanks very much. One thing I wonder is will it support multiple angle image references? At the moment it will sometimes create holes in the back of a character.

SECourses

it doesnt support. you need to increase resolution it fixes that but needs more VRAM.

SECourses

i may make later. currently improving texture tab. also generate multiple times because different seeds fixes your problem i testd

BigVI

Thanks, though it seems this is for the original Trellis. I've also read that some people are working on this functionality for Trellis 2, though it's not something they provide "out of the box"

Firat Güven

Looks interesting but even with 24GB VRAM (4090) it is not possible to finish an execution, since the VRAM gets full within a minute and gets extremely slow. i tried to reduce the resolution and activated a couple of features which should help to reduce VRAM usage but no... may be it works with 32GB, for me it seems useless unfortunately.

SECourses

512 px should work for you but it is vram heavy

Taiga

this works very well!

Taiga

is there a way to select by color when imported glb into blender 5?

SECourses

it generates texture as well. but if you mean without texture probably you can change in blender

Taiga

how do I get more solid renders? Right now, it's very difficult to get theses 'manifold'

SECourses

sadly i am not expert . can you eloborate more so i will try to come up with some specific answer

Taiga

For a 0.4 nozzle, aim for at least 0.8–1.2 mm wall thickness (2–3 lines). In Blender: add Solidify modifier → set Thickness to something like 1.2 mm (or 0.0012 m if you’re in meters) → Apply.

SECourses

i am going to add massive new features to this app. working on it. hopefully output quality will be much better. i also added unirig so you can auto rig characters as well. you can message me on discord i can show you.

Jared Newbold

I just started digging into the v6 release and it's pretty nice so far, I was already using Ultrashape in ComfyUI to kind of shore up the messy meshes that Trellis2 kicks out so it was really nice to see you had added it to v6. For me, however, it needed some extra love in order to make it work as it initially complained about not being able to find the model because ULTRASHAPE_ROOT was missing. I had to clone the Ultrashape git repo in the models subdirectory, move around the downloaded Ultrashape model (as the default installer placed it in 'Ultrashape' and the code seemed to want 'Ultrashape-1.0') and finally do a soft-link / junction (I'm on Windows but it's basically the same as Linux in this instance) at the root of 'Trellis_2_3D_Generator' to 'Ultrashape-1.0' in order to make it work but it's running now, with around a 40 minute runtime on my box (a 5060 Ti with 16GB of VRAM, but I'll admit my CPU is a bit low end). In my previous ComfyUI runs it was only maybe 10 minutes to get a 1.3 Million face mesh out of a Trellis2 rough mesh input, but I'll wait until it gets finished to see how it does. There's a little work to be done around the Ultrashape installer, though, I think. Would be happy to send you a PR if it helps, but I'm actually not sure where the code is. Anyway, thanks as always for the work, it's excellent.

SECourses

thanks for comment. if you have any issues let me know

Kieran Hawk

When I enable "Enable UltraShape Refinement" I'm getting the following error during GLB extraction: "[extract_ultra] refinement failed: FileNotFoundError: UltraShape source not found. Set ULTRASHAPE_ROOT to the UltraShape source checkout, or place it in one of: A:\1_APPS\TRELLIS2\Trellis_2_3D_Generator\UltraShape-1.0, A:\1_APPS\TRELLIS2\Trellis_2_3D_Generator\ComfyUI-UltraShape1\UltraShape-1.0, A:\1_APPS\TRELLIS2\Trellis_2_3D_Generator\UltraShape_v2, A:\1_APPS\TRELLIS2\Trellis_2_3D_Generator\models\UltraShape-1.0. Falling back to input mesh blob." No errors during installation. All models downloaded without error. I made two separate complete installations to make sure.

SECourses

please get v 6.2 zip file, overwrite and run installer again and let me know after that if works accurately

Kieran Hawk

Seems to be working fine. Thank you for the quick response! :)

Philipp Capetian

**Question: cumesh "no kernel image is available" (CUDA error 209) on RTX 4000 Ada (sm_89) — bug or something on my side?** Hi Furkan, thanks for the app. Before I assume it's a bug, I'd like to check whether I'm misreading this or missing a step. **System** - GPU: NVIDIA RTX 4000 Ada Generation (Compute Capability 8.9 / sm_89) - OS: Ubuntu 24.04.4, kernel 6.17.0-29 - NVIDIA driver 580.159.03, CUDA 13.0 - venv torch: 2.9.1+cu130 (from your installer) - Installer: Massed_Compute_Install.sh, already on v6.2, all 60 models downloaded, install completed clean, no errors **What works** The full diffusion pipeline runs perfectly: preprocess, encode_cond, sample_sparse_structure, sample_shape_slat (1024 cascade), sample_tex_slat. ~5 min, ~6 GB VRAM peak. So flash_attn, flex_gemm, xformers etc. all seem fine for sm_89. **Where it stops** Only cumesh calls fail, with: ``` RuntimeError: [CuMesh] CUDA error: File: /workspace/CuMesh/src/connectivity.cu Line: 125 Error code: 209 Error text: no kernel image is available for execution on the device ``` First in preview_decode_mesh (base.py fill_holes -> cumesh.get_edges), and in Extract GLB at o_voxel/postprocess.py to_glb -> cumesh.fill_holes (cumesh.py line 288). **My question** Error 209 usually means the wheel has no compiled kernel for this GPU's arch. Since everything else works on sm_89, my guess was the bundled cumesh wheel just doesn't include sm_89 — but I'm not sure that's the right read. I noticed another user here with a 4090 (also sm_89) hit VRAM limits rather than this kernel error, which makes me wonder: is cumesh supposed to work fine on sm_89 and I have a Linux-specific or env issue, or is the Linux cumesh wheel built with different archs than the Windows one? Just want to confirm whether this is expected to work on sm_89 before I chase the wrong thing. Thanks!

SECourses

working on this right now to recompile will let you know

SECourses

hello fixed with v 6.3 zip file. i tested on L40S same genre. compiled lbrary again. so please use new zip file and let me know