Native ComfyUI nodes for SenseNova-U1.5-8B-MoT (any-to-any: text-to-image, single-image editing, multi-reference editing, batch generation), forked from T8mars' wrapper and extended with ConvRot quantization support so the 50 GB bf16 model runs on a 12 GB GPU.
Quantized weights live here: Milor123/ComfyUI-ConvRot-SenseNova-U1.5-8B-MoT-T8
| Download | Size | Place it in |
|---|---|---|
| SenseNova-U1.5-8B-MoT-T8-hybw4a8-L18-41.safetensors | 13.80 GiB | ComfyUI/models/diffusion_models/SenseNovaU1.5/ |
| SenseNova-U1.5-8B-MoT-T8-int8-convrot-tagged.safetensors — recommended | 17.58 GiB | ComfyUI/models/diffusion_models/SenseNovaU1.5/ |
| SenseNova-U1.5-8B-MoT-LoRA-8step-ComfyUI.safetensors | 0.76 GiB | ComfyUI/models/loras/ |
- ConvRot-aware quantized inference (
sensenova_u15/quant_bridge.py): explicit activation rotation + manual dequantization for INT8 ConvRot, and comfy-kitchen kernel routing for W4A4 / W4A8 — ComfyUI's generic dispatch silently skips the rotation these checkpoints require. - QuantizedTensor runtime guards (
qt_guards.py): neutralizes the dtype-cast paths in ComfyUI weight streaming (cast_to_device/Tensor.to(dtype)) that corrupt packed quantized tensors when LoRAs are applied. - Static checkpoint contract (
sensenova_u15/checkpoint_contract.py): deterministic key/shape/dtype validation immune to the partialstate_dictthat meta-model construction returns inside full ComfyUI sessions. - Hybrid checkpoint tooling (
tools/make_hybrid_ladder.py): byte-level merging of validated INT8 and W4A8 checkpoints into hybrid files — no re-quantization involved. - INT4 converter (
tools/convert_sensenova_int4_convrot.py): self-contained W4A4 / W4A8 / mixed-mode converter built directly on comfy-kitchen layouts, with sequential (mmap-free) reads for large files. - Headless test harness (
tests/headless/): capture, compare, first-divergence and validation scripts used to debug quantization numerically without launching the ComfyUI UI.
All upstream features are retained: text-to-image, single-image edit, 1-10 reference-image editing, 1-16 outputs per prompt, standard KSampler, U1.5 Final and SFT weights, the official 8-step speed LoRA, three-path img_cfg guidance, CFG Norm, and structured edit prompts. Workflows live in examples/: result-t2i-8step-2048.json is the maintained example for this fork; original upstream workflows are archived in examples/old examples (without my nodes)/ for reference.
Clone into ComfyUI/custom_nodes/:
git clone https://github.com/Milor123/ComfyUI-SenseNova-U1.5-ConvRot.gitRequires ComfyUI with comfy-kitchen >= 0.2.31. Download quantized weights from the model repo into ComfyUI/models/diffusion_models/SenseNovaU1.5/.
This model tolerates W4A8 activation quantization in its later transformer layers but not in its earliest ones: the hybrid release anchors layers 0-17 in INT8 (bf16 activations) and runs layers 18-41 in W4A8. The boundary was located empirically with a bisect ladder of hybrid checkpoints. See the model card for measured numbers (INT8: 0.43% pixel diff vs bf16; hybrid: visually indistinguishable in same-seed A/B).
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.
- T8mars - upstream wrapper (Apache-2.0)
- sensenova - original model (Apache-2.0)
- ConvRot - Hanzuliang et al., technical notes in Comfy-Org/ComfyUI#14735
- silveroxides/convert_to_quant - INT8 conversion lineage
- Comfy-Org/comfy-kitchen - quantized tensor runtimes