SenseNova-U1.5-8B-MoT-FP8

SenseNova U1.5 8B MoT — Pruned BF16 / FP8 / INT8 + 8-step LoRA

This repository contains three pruned checkpoints derived from sensenova/SenseNova-U1.5-8B-MoT, plus a ComfyUI-ready copy of the official 8-step distilled LoRA.

Install the ComfyUI nodes

All three checkpoints use the dedicated ComfyUI-SenseNova nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/starsFriday/ComfyUI-SenseNova.git
pip install -r ComfyUI-SenseNova/requirements.txt

Restart ComfyUI after installation. Project documentation and example workflows are available at starsFriday/ComfyUI-SenseNova.

Model files

For ComfyUI, download one of the three checkpoint files into:

ComfyUI/models/diffusion_models/

Place the optional 8-step LoRA in:

ComfyUI/models/loras/

The FP8 checkpoint uses float8_e4m3fn storage with per-tensor scales for 592 linear layers. Token embeddings, convolution weights, normalization parameters, and other non-linear tensors remain in their source precision. The INT8 ConvRot checkpoint contains 588 int8_tensorwise linear layers with group size 256. The BF16 checkpoint retains the pruned model at high precision and requires substantially more system memory and GPU memory than the quantized variants.

ComfyUI usage

Use these nodes:

1. SenseNova U1.5 Model Loader
2. Optional: SenseNova U1.5 LoRA Loader
3. SenseNova U1.5 Text to Image or SenseNova U1.5 Image Edit
4. A standard ComfyUI image output node such as Save Image

The generation nodes expose independent width and height controls from 256 to 4096 pixels in steps of 32. Image Edit provides dynamic reference-image sockets and supports up to 64 ordered references, including images expanded from ComfyUI batches. Sampling progress is shown with the native ComfyUI progress bar and with step timing in the terminal. Enable low_vram for layer streaming on 24 GB GPUs.

SenseNova uses a stateful pixel-space Flow Matching process with prefix KV caches. It therefore uses its dedicated sampling nodes instead of ComfyUI's standard KSampler.

8-step LoRA

Connect Model Loader → LoRA Loader → Text to Image and use these exact settings:

steps: 8
cfg: 1.0
cfg_norm: none
timestep_shift: 3.0
LoRA strength: 1.0

The node applies the adapter as a low-rank residual without expanding or permanently modifying the base weights. The same LoRA therefore works with the BF16, scaled-FP8, and INT8 ConvRot checkpoints in this repository. It requires the final SenseNova U1.5 base, is incompatible with the earlier Preview checkpoint, and is intended for text-to-image generation only. Use the base model without the distilled LoRA for image editing.

The 882 LoRA tensors are bit-identical to the official SenseNova-U1.5-8B-MoT-LoRA-8step.safetensors. Only ComfyUI base-model and sampler-setting metadata was added to the adapted file.

Pruning scope and limitations

Only language_model.lm_head.weight was removed. The understanding path, _mot_gen image-generation path, vision encoder, and Flow Matching modules are retained. This checkpoint supports:

• Text-to-image generation
• Single-image and multi-reference image editing

The pruned checkpoint does not support Think mode, VQA text output, or interleaved text/image output.

Recommended starting settings for this base model are 50 steps, CFG 4.0, and timestep shift 3.0.

License and attribution

The checkpoint follows the upstream Apache-2.0 license. See the SenseNova U1 project and the base model card for upstream details. The 8-step adapter comes from sensenova/SenseNova-U1.5-8B-MoT-LoRAs.