FLUX 3 Image's New, Anchor and Move elements as Sume edit prompts

FLUX 3 Image labels every element in a layout as New, Anchor or Move, and Sume's Image API has no field for any of them. You get the same three jobs done differently: New becomes an instruction in the prompt, Anchor becomes a preserve list plus, on GPT Image 2.5, a mask, and Move becomes two edits because no Sume model has a move operation.

That is the honest answer to "can I do FLUX 3 style box edits on Sume": the effects are reachable for most edits, but the control is looser. BFL's docs describe a structured request. Sume's current docs describe a prompt, up to 10 reference images on most edit models (16 on ChatGPT Image 2.5) and an optional mask_url for ChatGPT Image 2.5 edits.

What BFL's three tags mean

From BFL's FLUX 3 Image documentation: New is generated content added to the image, Anchor is content preserved unchanged from a reference, and Move is content repositioned from a source location to a target location. Boxes use [y0, x0, y1, x1] coordinates on a 0 to 1000 grid. The docs also show recolor and replace edits, including several replacements in one request.

Writing the three as a prompt

A small builder keeps the wording consistent across edits. Anchors become one closing sentence listing what to keep, which is the form the habit our GPT Image 2.5 edit-drift post recommends repeating on every turn, because edits drift over several rounds. It deliberately refuses move rather than pretending the prompt can do it.

def edit_prompt(elements):
    """elements: list of dicts with kind new|anchor|move|replace|recolor, name, and optional where/to."""
    keep = [e["name"] for e in elements if e["kind"] == "anchor"]
    steps = []
    for e in elements:
        k = e["kind"]
        if k == "new":
            steps.append(f"Add {e['name']} {e.get('where', '')}".strip())
        elif k == "replace":
            steps.append(f"Replace {e['name']} with {e['to']}")
        elif k == "recolor":
            steps.append(f"Recolor {e['name']} to {e['to']}")
        elif k == "move":
            raise ValueError("no move operation: run an erase pass, then a place pass")
    prompt = ". ".join(steps) + "."
    if keep:
        prompt += " Keep exactly as in the input: " + ", ".join(keep) + ". Change nothing else."
    return prompt

print(edit_prompt([
    {"kind": "recolor", "name": "the sofa", "to": "forest green"},
    {"kind": "new", "name": "a floor lamp", "where": "in the left third"},
    {"kind": "anchor", "name": "the window and curtains"},
    {"kind": "anchor", "name": "the rug"},
]))

Where the mapping breaks

A mask on Sume is guidance. The docs for GPT Image 2.5 list mask_url as an optional public HTTPS URL and no other model accepts it; a request that sets an unsupported parameter is rejected with 400 unsupported_parameter. Reports of edits leaking outside a mask are the reason to verify unchanged areas yourself instead of trusting the label "Anchor".

Move is the weakest mapping. FLUX 3 Image takes it as an element type. On Sume you erase the object at the old location with a masked edit, then place it at the new one with a second masked edit that takes the object as a reference. Each pass is a paid generation that can shift pixels, so check the result against the original before you ship.

Which one to reach for

If your edits are recolor, replace and add on a stable photo, GPT Image 2.5 with a mask and a preserve list gets you most of the way. If your core need is a layout of many labelled boxes with exact pixel preservation, that is what BFL built FLUX 3 Image for, and it is not something Sume lists today. Both facts can be true in one pipeline: keep Sume for the generation and edit steps it serves and decide on BFL separately.

For the box-to-mask conversion itself, see FLUX 3 bounding box to mask URL in Python. For the Sume side of mask edits, the Image API docs list the fields.