SE Talk Track

This is one Workers AI model doing two jobs: upload a photo and describe an edit in plain English ("add a sunset background", "remove the background", "add falling snow") and it generatively re-renders the photo with that change — or skip the upload entirely and generate an image from scratch with the same endpoint. This isn't a documented feature of this model on the public model page — it was confirmed empirically: uploading a receipt photo with an edit instruction returned a result that visibly preserved the original photographed-on-a-table composition with the requested change applied, proving the image is genuinely used as an editing base, not ignored. Be upfront in the room: this is a generative, diffusion-based edit, not a pixel-precise crop or overlay — fine detail, especially small text, can shift. And it takes 10-15 seconds per call, which is a real UX constraint worth designing around, not something to gloss over.

Demo Steps

1 Start in Edit mode — upload or drag in a product/room photo
2 Click one of the example prompt pills (e.g. "Add a sunset background") or type your own edit instruction
3 Click "Edit photo" and narrate the wait — point out the explicit ~10-15s ETA in the loading state, not a spinner that looks stuck
4 When the result lands, use the Before/After/Side-by-side toggle to compare — call out anywhere fine detail or text shifted, that's the diffusion model being honest, not the demo failing
5 Switch to Generate mode — the upload area disappears, it's now pure text-to-image on the exact same endpoint
6 Type a prompt like "a minimalist product shot of a red sneaker on a white background" and generate — same model, same latency, no input photo
7 Show the Worker code panel — one multipart POST, one model, both modes decided by whether an `image` field is present

Cloudflare Products

Workers AI@cf/black-forest-labs/flux-2-klein-4b

"Why not just use Photoshop's generative fill or a dedicated image-editing API?" — Those work great but need either a licensed desktop tool or a separate vendor integration with its own auth, pricing, and latency profile. This is one Workers AI binding, billed per-tile like every other Workers AI model, editable in the same codebase as your other AI features — the tradeoff is it's generative (re-renders pixels) rather than a precise masked edit.

🎨 E-commerce / Real Estate / Marketing

AI Photo Editor

Upload a photo, describe an edit in plain English, get an edited version back — or skip the upload and generate an image from scratch. One Workers AI model, two modes, no image-editing pipeline to build.

The Problem

""We want to let users (or our own team) restyle product photos, stage empty rooms, or spin up creative variations — but a real image-editing pipeline means a diffusion model to host, GPUs, and an inpainting/masking UI just to get started.""

The Outcome

Workers AI model (@cf/black-forest-labs/flux-2-klein-4b) handles BOTH photo editing and text-to-image generation — no separate model, no masking UI, no GPU to provision.

Live demo below

Productionising this

What changes when you ship this for real

Content moderation

Users can submit inappropriate images or prompts to a public-facing endpoint. Before shipping this beyond an internal demo, add a moderation pass — Cloudflare's Workers AI text/image classification models can screen prompts and uploads, or route flagged content to a manual review queue.

Rate limiting

At $0.000059/input-tile + $0.000287/output-tile per 512x512 call, this is far from free at demo-day volume, let alone public traffic. Add a KV or Durable Object-backed rate limiter per session/IP before this goes anywhere near production.

Storage strategy

This demo returns the result as an inline base64 string — fine for a single ephemeral view, wasteful for anything users expect to revisit. If results need to persist, write the decoded image to R2 and return a signed URL instead of re-sending base64 on every load.

Latency is a real UX constraint

10-15 seconds is not "basically instant" — design the surrounding UI (loading states, disabled buttons, an explicit ETA) around that reality rather than treating it as a background operation users won't notice.

Editing photos that contain real people raises consent and likeness questions independent of Cloudflare's tooling — a production deployment should have a policy on whose photos can be uploaded and what edits are permitted (e.g. no photos of people without consent, no impersonation-style edits).

Undocumented model behaviour

The image field for edit-mode input on this model is not documented on the public model page as of this writing — it was confirmed via live testing. Undocumented behaviour can change without notice; pin a specific model version if this becomes a production dependency, and re-verify after any Workers AI catalog update.