Before you start
TRELLIS.2 memory use cannot be reduced to the size of one model file. This planner separates the ComfyUI BF16 and INT8 ConvRot SafeTensors routes from GGUF Q8 and Q4 routes, then asks you to validate GPU, resolution, loader, shape generation, texturing, mesh processing, and export as one recorded combination.
Prerequisites
- The exact GPU model, architecture target, VRAM, operating system, and driver
- A chosen runtime: official Python, ComfyUI, a community ROCm stack, or trellis.cpp
- System RAM and disk headroom for support assets, caches, and any offload
- One repeatable input, seed, resolution, and output workflow
Current snapshot
- Decision order
- Runtime → precision → resolution → stagesLoader support decides whether a smaller file is useful at all.
- ComfyUI routes
- BF16 or INT8 ConvRot SafeTensorsINT8 requires a compatible ConvRot loader and backend.
- Native routes
- GGUF Q8 or Q4trellis.cpp uses a multi-file model set and does not load the Comfy-Org SafeTensors directly.
- Compatibility unit
- GPU + OS + runtime + resolution + workflowA card name or checkpoint size alone is incomplete evidence.
Choose a route before you install
The file-size scopes differ: Comfy-Org numbers below refer to one diffusion SafeTensors file; trellis.cpp numbers describe an approximate complete GGUF model set.
| Route | Published disk scope | Primary runtime | Start here when… |
|---|---|---|---|
| BF16 | 10.3 GB diffusion file | ComfyUI / reference-style Python | You need the cleanest comparison baseline and your tested workflow has headroom |
| INT8 ConvRot | 5.25 GB diffusion file | ComfyUI with a ConvRot-aware loader | You want a smaller diffusion payload and your GPU has a working native or validated fallback backend |
| GGUF Q8 | ≈9.5–10 GB complete set | trellis.cpp; some custom ComfyUI nodes | You prefer a native cross-platform runtime and want the higher-quality GGUF route |
| GGUF Q4 | ≈6 GB complete set | trellis.cpp; some custom ComfyUI nodes | Q8 does not fit the tested workflow and you accept a larger quality risk |
Step by step
Choose the runtime before the quant
BF16 and INT8 ConvRot are SafeTensors routes intended for a compatible Python or ComfyUI loader. GGUF Q8 and Q4 are native to trellis.cpp and also appear in community ComfyUI integrations. Select the runtime your operating system and GPU can actually execute before comparing file sizes.
Build a 512 shape-only baseline
Disable texture generation and optional mesh operations, keep one input and seed fixed, and run the lowest supported shape workflow. Record peak VRAM, system RAM, cold and warm time, backend messages, and whether CPU offload or an emulated kernel was used.
Move through precision one step at a time
Start with the highest-precision route that your selected runtime can attempt, then move BF16 → INT8 ConvRot → GGUF Q8 → GGUF Q4 only when the previous route fails the measured memory budget. Do not change quantization and resolution in the same comparison.
Add texture and mesh stages separately
Enable VAE decode, texture flow, mesh cleanup, UV work, baking, and export in stages. A shape-only run can fit while the textured PBR route fails later. Keep the maximum allocation from the entire run, not just model load.
Test the working resolution ladder
After the complete 512 route is stable, test 1024 or cascade workflows using the same precision and input. Stop when the workflow falls back unexpectedly, pages system memory, produces invalid values, or exceeds the time budget — not only when it raises an out-of-memory error.
Publish the combination, not a badge
Record GPU, architecture target, OS, driver, PyTorch or native runtime version, loader commit, quant, resolution, workflow type, peak VRAM, peak RAM, wall time, and output result together. That row is reusable evidence; a standalone label such as '8 GB supported' is not.
Compatibility evidence by GPU, resolution, and workflow
This table records only the scope supported by the linked source. It deliberately leaves untested combinations unclaimed and treats a shape-only run differently from a textured PBR export.
Evidence reviewed 25 Aug 2026. 'Published route' is weaker than an end-to-end card-specific measurement.
| GPU / backend | Resolution + workflow | Precision route | Evidence level |
|---|---|---|---|
| NVIDIA A100 / H100 · Linux CUDA | Upstream reference examples; H100 timing published at 512³, 1024³, 1536³ | BF16 reference | Official upstream baseline; not a consumer-GPU claim |
| RX 7900 XTX · Linux ROCm gfx1100 | 512 shape-only; 1024 shape-only; 1024 PBR textured | INT8 ConvRot | Community end-to-end runs with pinned Python, PyTorch, ROCm, and Triton |
| CUDA / ROCm / Vulkan backends | Native full pipeline; resolution depends on selected build and card | GGUF Q8 / Q4 | Published trellis.cpp route; verify the exact GPU in project results or your own run |
| Any other GPU / stack | Any resolution or full textured workflow | Any quant | Unverified until a complete, versioned run is recorded |
Measure the whole pipeline, not the checkpoint
Checkpoint bytes are useful for storage planning and a rough lower bound. They do not include every live tensor or downstream 3D stage. The peak can move when a model is unloaded, when VAE decode begins, when sparse geometry expands, or when a high-resolution texture atlas is rasterized.
- Capture GPU peak and system-RAM peak for both shape-only and textured workflows.
- Record whether weights stayed resident, were unloaded between stages, or were offloaded to CPU.
- Note resolution, face count, texture atlas size, and optional remesh or hole-fill operations.
- Treat severe paging or silent CPU fallback as a failed performance target even if the run eventually completes.
Known trade-offs
- There is no universal 6 GB, 8 GB, 12 GB, or 16 GB guarantee
- The 5.25 GB INT8 number and ≈6 GB GGUF Q4 number describe different package scopes and are not directly comparable
- System-RAM offload can avoid one GPU allocation limit while causing severe paging and much longer runs
- Quantization can change geometry or material output and only helps when the selected loader and backend support it
- A successful 512 shape-only run does not establish compatibility with 1024, cascade, or textured PBR workflows