Modelverse Step1X Edit

Here's the workflow this node is for: you've got an image you mostly like, and you want to change one thing about it - swap the background, change the lighting, remove the sunglasses, make the subject look in the other direction. That's not regeneration, that's editing, and Modelverse Step1X Edit does it with plain instructions. Feed it an image plus a sentence like "turn this into a rainy night scene," and you get an edited IMAGE back.

Step1X-Edit is StepFun's instruction-driven editing model (the same lab behind the Step series), and UCloud hosts it on their ModelVerse API. The README describes it as "precision local editing and inpainting" - which is shorthand for: give it a target and it changes that, not the whole image. It's the closest thing this pack has to a "photoshop with words" node, and it's the one I'd reach for when a Flux-style redraw would be overkill.

The inputs that matter

The full set is small and honest:

• image - the source image, wired in as an IMAGE tensor from any loader or upstream node. It's required; the node raises an error if it's missing.
• prompt - your edit instruction. The tooltip says it plainly: "The prompt to guide the image edit."
• negative_prompt - optional, for keeping unwanted things out ("blurry, distorted, extra limbs" style).
• num_inference_steps (default 30, 1–50) and guidance_scale (default 4.0, 0–20) - the familiar diffusion dials. More steps means slower but cleaner; guidance 4.0 is a reasonable starting point and you'll rarely need to push it.
• num_requests (1–10) - fires multiple concurrent edit calls, each with the seed bumped, so you get variations to pick from.
• seed - -1 for random, or a fixed number to reproduce an edit.

Output is a single batched image tensor (one per request, concatenated).

How it works

Under the hood the node base64-encodes your image, builds a request to Modelverse's Step1X-Edit endpoint, and fires requests concurrently via the shared async client. Because it's hosted, the "steps" happen on their servers - your GPU never wakes up. That also means each run costs credits, and num_requests multiplies the bill. The result URLs are downloaded and converted back to tensors, so the node behaves like any ComfyUI image node despite being a cloud round-trip.

Install and the practical loop

Standard pack install: ComfyUI Manager → search "ComfyUl-UCloud" → Install → restart, or git clone https://github.com/ucloud/comfyui-plugin into custom_nodes. No model downloads, light dependencies, and you need a Modelverse API key wired through the Modelverse Client / Secret Client.

The practical pattern for photo edits: load your image, drop it into this node, describe the change, and use num_requests of 2–3 to get options. Then take the result you like and feed it through the pack's Preview or Save nodes. One gotcha from the source: the image input is a plain tensor, so an oversized source can produce weird results - keep inputs at sensible working resolutions rather than feeding it a 4K render and wondering why the edit looks off.