ComfyUI-Sensenova-U1.5-AEGIS

ComfyUI SenseNova U1.5 AEGIS

High-performance local SenseNova U1.5 nodes for ComfyUI.

Features:

• local single-file INT8 ConvRot checkpoint loading;
• AEGIS hybrid VRAM residency and async layer prefetch;
• official SenseNova U1.5 LoRA support;
• Text-to-Image and multi-reference Image Edit;
• automatic FlashAttention → PyTorch SDPA fallback;
• automatic runtime/dependency preparation after ComfyUI restart;
• no automatic download of model weights.

This project targets the final SenseNova-U1.5-8B-MoT, not the older Preview checkpoint.

Example workflow

A cleaned example workflow is included in:

examples/SenseNova-U1.5-AEGIS.json

Load your own reference images after importing the workflow. Model and LoRA filenames are placeholders matching the recommended files from the Models section.

Models

Recommended ComfyUI checkpoint

SenseNova-U1.5-8B-MoT-pruned-int8_convrot.safetensors

ComfyUI-ready INT8 ConvRot checkpoint:

https://huggingface.co/joyfox/SenseNova-U1.5-8B-MoT-FP8

Place it in:

ComfyUI/models/diffusion_models/

The same repository also contains BF16, scaled FP8 and a ComfyUI-converted 8-step LoRA.

Official base model

https://huggingface.co/sensenova/SenseNova-U1.5-8B-MoT

Official 8-step LoRA

https://huggingface.co/sensenova/SenseNova-U1.5-8B-MoT-LoRAs

Official file:

SenseNova-U1.5-8B-MoT-LoRA-8step.safetensors

For the ComfyUI-converted LoRA used with this project, place the file in:

ComfyUI/models/loras/

Example filename:

SenseNova-U1.5-8B-MoT-LoRA-8step-ComfyUI.safetensors

Installation

1. Copy the complete node folder to:

ComfyUI/custom_nodes/ComfyUI-SenseNova-U1.5-AEGIS/

2. Put the INT8 ConvRot checkpoint into:

ComfyUI/models/diffusion_models/

3. Optional: put the 8-step LoRA into:

ComfyUI/models/loras/

4. Restart ComfyUI.

On startup the node checks/prepares its runtime and auxiliary Python dependencies automatically. It does not download SenseNova model weights.

For ComfyUI Manager installation: install the node, restart ComfyUI, then place the model files in the folders above.

Nodes

• SenseNova U1.5 AEGIS Loader
• SenseNova U1.5 LoRA Loader
• SenseNova U1.5 Text to Image
• SenseNova U1.5 Image Edit

Image Edit supports multiple reference images through dynamic image inputs.

Recommended settings

Base model — maximum quality

steps:          50
cfg:            4.0
cfg_norm:       none
timestep_shift: 3.0
LoRA:           disabled

Official 8-step LoRA — fast Text-to-Image

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

0.90–1.00 LoRA strength can be useful for tuning, but 1.0 matches the intended distilled configuration.

The official 8-step LoRA is released for Text-to-Image. Image Edit usage should currently be considered experimental.

Image Edit without LoRA

steps:          50
cfg:            4.0
image_cfg:      1.0
cfg_norm:       none
timestep_shift: 3.0

VRAM modes

aegis — recommended

Hybrid GPU residency with asynchronous layer prefetch.

Current default profile:

prefetch:           4
VRAM fraction:      0.94
headroom:           2 GiB
activation reserve: 3 GiB

balanced

Lower residency and lower VRAM pressure.

low

Maximum offload for GPUs with limited VRAM.

full

Moves the full model to GPU and may OOM on consumer GPUs.

Console progress

Since previous development builds, progress is measured from the model's actual fm_modules.timestep_embedder. The bar advances when the denoise timestep changes, so repeated CFG/image-CFG forwards do not create fake extra steps.

Example:

SenseNova U1.5 Edit: 5/8 [00:11<00:06, 2.1s/step, t=0.428571]

Notes

• The loader uses a local checkpoint only.
• attention=auto safely falls back to PyTorch SDPA if FlashAttention is unavailable or broken.
• INT8 ConvRot is loaded directly into a meta model to avoid a redundant full-precision model copy.
• AEGIS keeps generation weights resident where possible to reduce CPU ↔ GPU transfers.

Upstream

SenseNova-U1 / U1.5:

https://github.com/OpenSenseNova/SenseNova-U1

Official ComfyUI integration:

https://github.com/OpenSenseNova/ComfyUI-SenseNova-U1

SenseNova U1.5 cookbook / best practices:

https://github.com/OpenSenseNova/SenseNova-U1/blob/main/docs/u1.5_best_practices.md

License

Please review the licenses of the upstream SenseNova project and the model checkpoints you download.

Shared-core LoRA lifecycle

SenseNova U1.5 AEGIS uses a shallow-copy shared-core design for LoRA:

base runtime ─────┐
                  ├── shared _core -> model / tokenizer
LoRA copy.copy ───┘
        │
        └──────────── same shared patcher

model and tokenizer live only inside _core. clone_with_lora() uses copy.copy(), so LoRA clones share both the same heavyweight core and the same ComfyUI patcher.

This fixes RAM retention during unload: when ComfyUI unloads the shared patcher, the shared model/tokenizer are released for the base runtime and all LoRA clones at once. Lazy reload uses register=False and reuses the existing patcher, so no orphan model registrations are created.