Step1X-Edit Generate

This is the node that actually does the thing the Step1X-Edit marketing promised: you drop in a photo, type "add a pendant with a ruby around her neck" or "make her smile", and out comes an edited image. No mask, no ControlNet, no inpainting prep. That's the whole appeal of an instruction-based editor, and Step1X-Edit is the closest open-weights model to the GPT-4o image-editing demo that most people can actually run locally.

Where it sits in the workflow: a Step1X-Edit Model Loader feeds a model wire into this node, you give it an image and a prompt, and it emits an edited IMAGE you pipe straight into a Preview or Save Image node. That's the entire graph - the loader article in this pack covers the heavy part; this node is where the words turn into pixels.

How it works

Step1X-Edit isn't a masked-inpaint model. Its text encoder is a whole multimodal LLM, Qwen2.5-VL-7B, and it reads both your sentence and the reference image before anything is generated. That's the secret sauce: the edit instruction gets conditioned on what the model actually sees, which is why "let her cry" works without you drawing a mask over her face. That encoding feeds a flow-matching DiT (think Flux-sized) that denoises over num_steps steps, and a VAE decodes the result back to pixels.

One detail worth knowing: unlike the guidance-distilled models that dominate 2026-era workflows, this model does real CFG. negative_prompt is genuinely wired in, not decoration - the sampler runs both your prompt and your negative and extrapolates. So a negative prompt like deformed, blurry, bad hands is worth filling in, not wasted effort.

The inputs that matter

• prompt - write full natural-language instructions, not comma-tag spam. This is an LLM-encoded model; the KB's "quality tags are dead on LLM-encoded models" rule applies. "A candid photo from a point-and-shoot" will move it more than "masterpiece, 8k".
• input_image - the reference photo to edit.
• num_steps - default 28. It's a slow model, so people trim this; quality degrades gracefully, but 20+ is the safe floor.
• cfg_guidance - default 6.0. Too high and edits get overcooked; 4.5–6 is the sensible band.
• size_level - the output canvas, max dimension in pixels (256–1024, default 512). Your image is resized to fit inside that square, aspect ratio preserved. This is your main VRAM lever.
• seed - reproducibility, and the thing to change when the edit is subtly wrong.

The only output is image. Wire it to Save/Preview.

Installing

Install the pack, drop the three model files (FP8 diffusion model, VAE, Qwen2.5-VL-7B-Instruct folder) into the right models/ subfolders, and pick the loader. Full steps live in the pack's Step1X-Edit Model Loader article - this node has zero extra dependencies of its own.

The honest part

Expectations first: the community reception when Step1X-Edit launched (April 2025) was "it's real, and it's nowhere near the demo quality." Early users reported the model changing things it shouldn't and missing precise edits, especially with vague prompts. It got better with clearer, more specific instructions - treat prompting it like instructing a capable-but-literal assistant. And it is heavy: at 512–768px without offload this pack wants ~31GB VRAM, so on a consumer card you live on the offload + quantized toggles and accept slower edits. CUDA out-of-memory on this node is almost always a size_level too high or offload switched off. When an edit misses, resist re-rolling seeds forever - rephrase the prompt instead.