SenseNova U1.5 Image Edit

A text-to-image node proves a model can render. An image-edit node proves it can think about an image - and SenseNova U1.5 is a multimodal model, which means it genuinely takes your reference image into its token stream rather than just compositing pixels on top of it. That's why this node exists and why it's worth trying, even though it's the least-traveled corner of the pack.

The workflow is: Loader → (optionally LoRA Loader) → Image Edit → SaveImage, with a LoadImage feeding the image socket. Feed it a reference, type something like the default Preserve the subject and composition; apply the requested edit, and it regenerates with your edit baked in. The inputs are the same sampler family as the Text-to-Image node - steps, cfg, timestep_shift, cfg_norm, width/height - plus one edit-specific dial:

• image_cfg - the image-conditioning strength, default 1.0, range 1–10. This is the knob that decides how hard the model holds onto your reference versus obeying the edit prompt. The README's no-LoRA edit config is 50 steps, CFG 4.0, image_cfg 1.0. If edits drift too far from your input, nudge image_cfg up; if the model is too glued to the reference to apply the change, ease it down.

The genuinely cool part is multiple references. The node accepts dynamic image inputs - wire a second, third, fourth LoadImage into it and they're collected in socket order after the primary image. The use case is obvious: show the model a subject from one angle and a style from another, or hand it several views of an object and ask for a consistent new composition. That's the kind of thing that needs a real multimodal model, not an SDXL inpainter.

Output is a single image, straight to SaveImage. Width and height follow the same rules as T2I: multiples of 32, 256–4096, default 2048×2048.

Installing and wiring

Same pack, same drill: ComfyUI Manager → search "SenseNova", or:

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

Restart, put the INT8 ConvRot checkpoint in ComfyUI/models/diffusion_models/ (the node never downloads weights), and a LoadImage into the image input. First run downloads the one-time runtime bootstrap, not the model.

Gotchas

• The 8-step LoRA is experimental here. The official distilled LoRA was released for Text-to-Image, and the README explicitly calls Image Edit with it experimental. If your edits look bad with the LoRA on, that's expected scope, not your fault. For edit work, run the base 50-step / CFG 4.0 / image_cfg 1.0 config.
• It needs at least one image. Empty image socket → error. That's by design; this node has no text-only fallback.
• Edits ignoring your prompt - image_cfg too high or cfg too low relative to your reference. It's a two-knob balance; change one, re-run, and you'll feel it.
• Non-32-multiple dimensions throw an explicit error rather than silently cropping, so that one at least fails loudly.
• On a 16 GB card, keep to a single reference image plus modest dimensions for the first run; multi-reference edits at 2048×2048 are the heaviest load in the pack.