Milor123/ComfyUI-ConvRot-SenseNova-U1.5-8B-MoT-T8

SenseNova-U1.5-8B-MoT-T8-convrot

Community quantizations of sensenova/SenseNova-U1.5-8B-MoT for ComfyUI, using ConvRot rotation-fused quantization. These files bring the 50 GB bf16 any-to-any model down to consumer GPUs (tested on RTX 4070 12 GB at 2048x2048).

Downloads & file placement

Download	Size	Format	Place it in
SenseNova-U1.5-8B-MoT-T8-hybw4a8-L18-41.safetensors	13.80 GiB	Hybrid INT8 + W4A8 ConvRot	ComfyUI/models/diffusion_models/SenseNovaU1.5/
SenseNova-U1.5-8B-MoT-T8-int8-convrot-tagged.safetensors	17.58 GiB	INT8 ConvRot — recommended, maximum fidelity	ComfyUI/models/diffusion_models/SenseNovaU1.5/
Loras/SenseNova-U1.5-8B-MoT-LoRA-8step-ComfyUI.safetensors	0.76 GiB	Official 8-step speed LoRA	ComfyUI/models/loras/

Requirements

• ComfyUI (tested with commit 82f839f5) with comfy-kitchen >= 0.2.31
• The ComfyUI-SenseNova-U1.5-ConvRot custom node (ConvRot-aware fork of the T8 wrapper)
• Download the files below into the folders listed in the placement column

Quality

Measured against the bf16 original:

• INT8 ConvRot: 0.43% pixel difference in a full-pipeline A/B (same seed); per-layer weight reconstruction error well under 2%.
• Hybrid W4A8: per-layer 4-bit weight reconstruction error ~7% rel-L2 (Lloyd-Max codebook, group size 16, fp8 group scales); output visually indistinguishable from bf16 in same-seed A/B tests.

bf16 (reference)	INT8 ConvRot (same seed)	Hybrid L18-41 (same seed)

The INT8 variant tracks the bf16 composition almost exactly; the hybrid follows its own (equally coherent) trajectory — same-seed trajectories are chaotic, so treat these as quality samples rather than pixel comparisons.

Note: the base model has a characteristic glossy/"plastic" skin rendering at photorealism. It is present in bf16 as well, so it is a trait of the model, not of the quantization.

Performance

On an RTX 4070 12 GB, both the INT8 and the hybrid W4A8 variants run surprisingly fast even though the model exceeds 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 turns into a slowdown. bf16 is a different story: it streams ~47 GB per step and feels drastically slower. On this GPU class, quantization is not just about fitting the model, it is what makes it enjoyable to use.

Why the hybrid? (for quantization practitioners)

We found that this model tolerates true W4A8 activation quantization in its later layers but not in its earliest ones: quantizing the first transformer blocks breaks prompt coherence, while quantizing layers 18+ is visually transparent. This matches the intuition that the earliest layers process the freshly embedded token stream, where precision matters most.

The hybrid files were produced by byte-level merging of two independently validated checkpoints (an all-INT8 and an all-W4A8 conversion), so no layer was ever re-quantized during merging. A bisect ladder over layer ranges was used to locate the fragility boundary empirically.

ConvRot applies a grouped regular-Hadamard rotation (group size 256) fused offline into the weights; at runtime the activations are rotated with the same matrix inside the comfy-kitchen kernels.

Credits & attribution

• sensenova/SenseNova-U1.5-8B-MoT - original model (Apache-2.0)
• T8mars/Comfyui-SenseNova-U1.5-Wrapper-T8 - upstream ComfyUI wrapper (Apache-2.0)
• ConvRot - Hanzuliang et al.; see the author's technical notes in Comfy-Org/ComfyUI#14735
• silveroxides/convert_to_quant - converter lineage
• Comfy-Org/comfy-kitchen - quantized tensor runtimes

License

Apache-2.0, inherited from the base model and the upstream wrapper.