Before you start
Q4 is the smallest published trellis.cpp GGUF set. It can create more working headroom than Q8, but its roughly 6 GB footprint does not by itself put a particular GPU-memory class in range. This guide is about measuring whether the complete workflow fits and whether the stated texture-quality cost is acceptable for your output.
Prerequisites
- A runtime that explicitly supports the Q4 weight set
- A known GPU, backend, resolution, and workflow whose complete peak can be recorded
- A fixed input and seed for comparison
- A tolerance for texture softness that you can state before testing
Step by step
Understand the stated trade-off
The project describes Q4 as suitable for lower-VRAM systems with slight texture graininess. That places the cost in the material channels, not primarily in the silhouette — which matters, because it means the defect is easiest to see in close-up renders and hardest to see in thumbnails.
Budget beyond the 6 GB
Roughly 6 GB of stored model files does not mean a 6 GB card works. The texture stage, activations, geometry, rasterization, and working buffers sit on top, and voxel resolution changes that peak. Prove one shape-only run and one complete textured run on the exact stack before recording compatibility.
Test on your worst case, not your best
Compare Q4 against Q8 or f16 using the input with the most fine texture detail you actually work with — fabric, printed surfaces, weathering. If Q4 holds up there, it will hold up everywhere else.
Consider a two-pass workflow
If Q4 is the only format that fits, use it for iteration and selection, then re-run the chosen seed and input through a higher-precision route — locally on a bigger machine or hosted — for the final asset.
Known trade-offs
- Q4 is the format with an acknowledged quality cost; do not present its output as equivalent to the reference pipeline
- There is no 6 GB guarantee — 6 GB is the weight footprint, not the peak allocation
- Texture graininess can survive into baked maps and downstream renders
- If Q4 is required and still fails, reduce voxel resolution before adding CPU offload