Reve Image Edit

"Make the jacket red, remove the sign in the background, and change the lighting to dusk." That's the entire input format for this node. Reve's edit model takes an image plus a natural-language instruction and does a real, reasoned edit - not a blind img2img reinterpretation. You hand it a photo, describe the change, and get back a modified image with the parts you didn't mention left alone.

It's the middle sibling of the Reve family: Create makes from scratch, Remix combines several reference images, and Edit is the one for surgical changes to a single image. If you've used Qwen-Edit or GPT Image editing, the interaction model will feel familiar - the difference is it's wired straight into your ComfyUI graph with tensor in, tensor out.

Inputs that matter

• image - what you're editing
• edit_instruction - multiline, up to 2560 characters. Be specific and list changes in order; this model rewards clear instructions over vibes.
• model - dynamic combo with two versions: reve-edit@20250915 (quality) and reve-edit-fast@20251030 (the cheap and quick one). Under the hood there's an aspect_ratio option including auto (keep the source ratio) and a test_time_scaling 1–5 dial.
• upscale - disabled or 2x/3x/4x, adds cost
• remove_background - boolean, adds cost
• seed - re-run control, results non-deterministic

Output is an IMAGE tensor.

How it works

The image is base64-encoded and sent with your instruction to /proxy/reve/v1/image/edit; the edited result comes back as a WebP binary and is decoded into a tensor. Hosted, billed per call from your Comfy account credits. Ships with ComfyUI core - no install, needs a Comfy account, no model files.

Common issues

• The fast model is a different product. reve-edit-fast is substantially cheaper and substantially more approximate. Use it for iterating, switch to the full model for the version you actually ship.
• It's a re-render, not a clone. Edits will re-render the image, so fine details - exact text, a specific face - can drift even in unchanged regions. Verify text and logos after editing.
• Deprecated. The whole Reve family got the deprecation flag in August 2026 (PR #15331). Functional today, but flagged for eventual removal - don't wire a business process to it.
• Server-side content policy. A flagged result raises an error (via the x-reve-content-violation header) instead of returning. No local override exists.

Reach for it when a client hands you a photo and says "change X, keep everything else" - that's the exact job it's good at, and it's a lot less fiddly than masking and inpainting the same change locally. Just keep the deprecation timeline in mind.