SenseNova U1.5 Text to Image

This is the node that actually draws the picture. Feed it the model output from the SenseNova U1.5 AEGIS Loader (optionally through the LoRA Loader), type a prompt, and out comes an IMAGE. The interesting thing is that one node covers two completely different regimes - the full 50-step base model and the distilled 8-step LoRA - and it will switch between them for you.

SenseNova U1.5 is a flow-matching model, and that changes what the dials mean compared to SDXL habits. There's no sampler or scheduler selector here at all; instead you get timestep_shift (default 3.0), which is the flow-matching knob that decides how sampling effort is spread between composition and detail. Leave it at 3.0 - the README's quality config is built around it. cfg_norm (none / global / channel / cfg_zero_star) is a normalization scheme on top of classifier-free guidance; the recommended setting is none, so just leave it.

The inputs that actually matter:

• prompt - multiline. Default is a cinematic portrait; change it to whatever you're making.
• steps and cfg - the pair that defines your regime. Base model: 50 steps, CFG 4.0. With the 8-step LoRA attached: 8 steps, CFG 1.0. This mirrors the distillation pattern you see across 2026's models - CFG 1 means guidance is baked into the weights, and pushing it back up to 7 just burns the image and doubles render time.
• auto_8step_lora_settings - default true, and this is the neat bit. If a LoRA is attached, the node silently forces steps=8, cfg=1.0, cfg_norm=none, timestep_shift=3.0. You can't screw up the distilled settings by forgetting to change them. Turn it off only if you deliberately want to fight the LoRA's intended config.
• width / height - default 2048×2048, and they must be multiples of 32 or the node throws a clear error. Range is 256 to 4096.
• batch_size - 1 to 4. On a 16 GB card, don't go cranking this just to see what happens.
• seed - standard, with control_after_generate so you can walk variations.

The only other output is image, which you wire into a SaveImage (the pack's example workflow uses SaveImage). There's a console progress bar that reads the model's real timestep embedder, so repeated CFG passes don't fake extra steps - nice touch when you're counting seconds per image.

Installing and wiring

The pack: ComfyUI Manager → search "SenseNova", or clone it into custom_nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/vizart-vj/ComfyUI-SenseNova-U1.5-AEGIS

Restart, make sure the INT8 checkpoint is in ComfyUI/models/diffusion_models/, and the LoRA (if used) in ComfyUI/models/loras/. The first run does a one-time runtime bootstrap download; model weights are never auto-downloaded.

Gotchas

• Feels slow and blurry? You're running the base 50-step config without a LoRA - that's "maximum quality" mode, not broken. 50 steps on an 8B multimodal at 2048×2048 is not a sprint.
• Blown-out colors at high steps? You've probably left cfg at something SDXL-ish. Base model wants 4.0; LoRA wants 1.0.
• "Width and height must be multiples of 32" - this one bites everyone eventually; the 2048 default is safe, and any size stepping by 32 is fine.
• Black images with the LoRA at 50 steps - that's a settings mismatch, not a model failure. Let auto_8step_lora_settings do its job, or match 8/1.0 manually.