SenseNova-U1.5 ConvRot for ComfyUI: 50GB Model on a 12GB GPU
ConvRot quantization runs SenseNova-U1.5-8B-MoT in ComfyUI on a 12 GB GPU: INT8 (17.6 GB) and hybrid W4A8 (13.8 GB) checkpoints plus an 8-step speed LoRA.
What is in the release
Two quantized checkpoints plus the official speed LoRA:
| File | Size | Format | Notes |
|---|---|---|---|
SenseNova-U1.5-8B-MoT-T8-int8-convrot-tagged.safetensors | 17.58 GiB | INT8 ConvRot | Recommended, maximum fidelity |
SenseNova-U1.5-8B-MoT-T8-hybw4a8-L18-41.safetensors | 13.80 GiB | Hybrid INT8 + W4A8 | Layers 0-17 in INT8, layers 18-41 in true W4A8 |
SenseNova-U1.5-8B-MoT-LoRA-8step-ComfyUI.safetensors | 0.76 GiB | 8-step speed LoRA | Official LoRA converted for ComfyUI |
Both checkpoints go into ComfyUI/models/diffusion_models/SenseNovaU1.5/, and the LoRA into ComfyUI/models/loras/. Inference runs through a ConvRot-aware fork of the T8 wrapper node, which adds explicit activation rotation and manual dequantization for INT8, kernel routing for W4A8 through Comfy-Org's comfy-kitchen, and runtime guards that stop ComfyUI's weight streaming from corrupting packed quantized tensors.
Quality
Measured against the bf16 original with same-seed full-pipeline A/B runs:
| bf16 (reference) | INT8 ConvRot (same seed) | Hybrid W4A8 (same seed) |
|---|---|---|
![]() | ![]() | ![]() |
| Reference | 0.43% pixel diff | Visually indistinguishable |
The INT8 variant reports a 0.43% pixel difference over a full-pipeline same-seed A/B, with per-layer weight reconstruction error well under 2%. The hybrid reports about 7% relative L2 error on its 4-bit layers (Lloyd-Max codebook, group size 16, FP8 group scales) yet is visually indistinguishable from bf16 in same-seed tests. Same-seed trajectories are chaotic, so the authors present the images as quality samples rather than pixel comparisons.
Why the hybrid
The interesting finding behind the release: SenseNova-U1.5 tolerates true W4A8 activation quantization in its later layers but not in its earliest ones. Quantizing the first transformer blocks destroys prompt coherence, while layers 18+ quantize to W4A8 transparently. The authors located the boundary empirically with a bisect ladder of hybrid checkpoints, then produced the hybrid files by byte-level merging of two independently validated checkpoints, so no layer was ever re-quantized during the merge.
Performance
On an RTX 4070 12 GB, both variants run fast even though the weights exceed VRAM: ComfyUI streams weights on demand, and the quantized formats move 3-4x fewer bytes per step while computing through fast integer tensor-core kernels, so the overflow never becomes a slowdown. bf16 on the same card streams about 47 GB per step and is drastically slower.
Getting started
- Install the custom node:
git clone https://github.com/Milor123/ComfyUI-SenseNova-U1.5-ConvRotintoComfyUI/custom_nodes/, withcomfy-kitchen >= 0.2.31installed (tested with ComfyUI commit82f839f5). - Download the recommended INT8 checkpoint (or the hybrid) into
ComfyUI/models/diffusion_models/SenseNovaU1.5/and the speed LoRA intoComfyUI/models/loras/. - Run one of the example workflows from the repo's
examples/folder:


Comments
Sign in with GitHub to join the discussion.