Empty SenseNova Pixel Latent

The name is a lie, and it's the most useful thing about this node. Empty SenseNova Pixel Latent doesn't create a latent at all. SenseNova U1.5 is a pixel-space model - no VAE compression, it generates raw RGB pixels directly - so the "latent" here is really a blank canvas of zeros that the sampler fills in with color. That's the whole reason it exists: stock EmptyLatentImage hands the sampler a compressed latent in the wrong space, and this node hands it the right-shaped tensor.

Where it sits: every SenseNova text-to-image workflow is SenseNova U1.5 Loader → SenseNova Sampling Options → KSampler → VAE Decode → Save Image, and this node is the thing feeding the sampler's latent input. It's your EmptyLatentImage, just built for this one model.

What you actually set

Three inputs, and honestly only two matter for day one:

• width / height - default 2048×2048, range 64–4096 in steps of 32. The default is the model's showcase resolution, not a great starting point on a modest card (more on that below).
• batch_size - 1 to 16, and this is the interesting one. The tooltip says it all: "Generate 1-16 variants with the same prompt and reference images." Crank it to 4 and you get four different results from one prompt and one reference set, each with its own noise. It's the pack's built-in "roll the dice N times" knob, and it works for edits too - every result in the batch shares the same references.

Output is a single LATENT that goes straight into the sampler. Done. That's the whole node.

The trap

That 2048×2048 default is a trap on anything short of a 24 GB card. The author's own README says start at 512 or 768 with batch_size 2 on 24 GB, and explicitly warns against 2048×2048, batch_size 16. VRAM scales with batch size, and this model is a brute: a 2048×2048 50-step dual-reference edit took about 506 seconds on an RTX 5090 Laptop 24 GB. Batch of 16 at full res is how you OOM before the first step finishes.

Also worth knowing: this is the edit workflow too. You don't paste your source image here - the reference image goes into SenseNova Reference Image, and this node just provides the output canvas size.

Install

It ships with the Comfyui-SenseNova-U1.5-Wrapper-T8 pack, so install is the pack install: search "SenseNova U1.5 (T8)" in ComfyUI Manager, or

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-SenseNova-U1.5-Wrapper-T8.git

then restart ComfyUI. No extra Python dependencies, and the pack never downloads models at runtime - you grab the checkpoint from the README's Hugging Face link and drop it in models/diffusion_models/ yourself. If you already have that model loaded, this node costs you nothing else.