Reve AI is the editor-first image family: Reve 2.1 (mid-Jul 2026, ~3x parameters) plans layout before rendering, outputs native 4K with no upscaler (vendor claim), and edits for $0.01 per operation — its edit endpoints out-run its create endpoint on public registries. Five variants (2.1, Reve, Remix, Remix Fast, Edit Fast) run inside the invideo agent.
Updated August 2026
Reve AI is the image family built around editing rather than one-shot generation. Its current flagship, Reve 2.1 (mid-July 2026), plans an image — layout, hierarchy, text placement, spatial relationships — before rendering a single pixel, outputs native 4K without an upscaling pass, and edits for as little as $0.01 per operation. That last number is not a typo, and it explains the family's most telling statistic: on Replicate's public model registry, Reve's editing endpoints have logged more runs than its create endpoint — people use Reve as an editor at least as much as a generator (replicate.com/reve, August 2026). In invideo's picker the family appears as Reve 2.1, Reve, Reve Remix, Reve Remix Fast, and Reve Edit Fast. A caution runs through this page: several headline claims — native 4K, "fastest 4K model" — are the vendor's own, and they are labeled as such below.
How did Reve get to 2.1?
Reve is a Palo Alto lab that surfaced in March 2025 with a preview model codenamed "Halfmoon," which established the family's cinematic, photojournalistic look and topped a major blind image arena at debut. Reve 1.0 followed in 2025, trained on structured data describing composition, relationships, and style. Reve 2.1 arrived in mid-July 2026 with roughly 3× the parameters of its predecessor and a new architecture (Reve's model page; the 2.1 date is anchored by partner launch posts rather than a dated official announcement).
| Release | Date | Notes |
|---|---|---|
| "Halfmoon" preview | Mar 2025 | Established the aesthetic; strong arena debut |
| Reve 1.0 | 2025 | Structured-data training: composition, relationships, style |
| Reve 2.1 | mid-Jul 2026 | ~3× parameters, new architecture, Create/Edit/Remix modes, native 4K |
Note the sequence: the structured-data training arrived with 1.0, before the 2.1 scale-up — the editing emphasis is baked into how the family was trained, not bolted onto it.
What does "plan before rendering" actually mean?
Reve's core pitch is that its model represents images as inspectable code rather than going straight from text to pixels. Because the image exists as a structured representation, edits apply losslessly — change one element and the rest doesn't degrade through repeated regeneration — and every addressable element can be manipulated directly in Reve's editor. Reve also positions this as agent-native: a language model can read the image's code and reason about it (app.reve.com/model). This is the vendor's own framing — no independent teardown confirms the internals — but the observable behavior matches it: region edits that preserve typography and brand elements, without the artifact accumulation that plagues pixel-space models. The same page carries the vendor claims of native 4K×4K (~16MP) output with no upscaling step and "fastest 4K image model in the world."
What are Reve 2.1's specs?
| Spec | Detail |
|---|---|
| Resolution | Native 4K×4K, ~16MP, no upscaler (vendor claim) |
| Modes | Create (text-to-image), Edit (instruction-based), Remix (multi-reference) |
| Reference images | Up to 8, addressed in prompts as <frame>1</frame>, <frame>2</frame>… — some hosts cap this at 4 |
| Aspect ratios | 1:4 through 4:1, plus auto |
| Text rendering | Multilingual, legible in-image text; edits preserve existing text and brand elements |
| Weights | Closed; API and first-party app only |
Sources: Reve's model page and the Reve 2.1 listing on Replicate, as of August 2026. The distinctive row is reference handling — <frame> tags make multi-image prompts addressable rather than guessable — but check your host's cap before writing prompts around eight references.
What does each Reve variant do?
invideo's Image tab carries five entries spanning the family's generate/edit/remix split:
| Picker entry | What it does | Notes |
|---|---|---|
| Reve 2.1 | The unified current model — create, edit, and remix in one | Newest; native 4K |
| Reve | The core create endpoint | The established text-to-image workhorse |
| Reve Remix | Multi-reference combination | Blends subjects, styles, lighting across reference images |
| Reve Remix Fast | invideo's fast remix variant | Speed-optimized remixing; see note below |
| Reve Edit Fast | Instruction-based edits at $0.01 per edit | The repricing headline |
One transparency note: "Remix Fast" does not appear as a documented standalone endpoint in Reve's public API listings as of early August 2026 — treat it as invideo's fast remix option rather than an externally documented Reve release. (Flagged for editorial review.)
Why the $0.01 edit matters
Reve Edit Fast lists at $0.01 per edit on Replicate — an order of magnitude below what most families charge for instruction-based editing (replicate.com/reve/edit-fast, August 2026). At that price, "fix the typo on the label" and "relight the product shot" stop being regeneration decisions and become retouching keystrokes. The market noticed: Reve's edit endpoints show 154K+ combined runs against roughly 109K for create on the same registry. Most image families are generators with editing bolted on; Reve's numbers suggest the reverse.
What Reve gets right — and what it doesn't
Strengths: typography that survives editing (change the scene, keep the label readable), preservation of brand elements through region edits, relighting and product-photo updates, strong prompt adherence in multi-character scenes, and the multi-reference Remix mode — official ComfyUI partner documentation highlights transferring "shadows, color grading, or aesthetic styles" between images (ComfyUI docs).
Weaknesses: vague prompts produce conservative edits — the model under-changes rather than over-changes, so "make it better" gets you almost nothing; cluttered reference images defeat precise region targeting; edit quality is heavily dependent on input quality; weights are closed with no self-host path; and the family does images only — no video or audio.
How to prompt Reve
- Be surgical in Edit mode. Name the element and the change: "replace the headline text with 'OPEN LATE', keep font and color" beats "update the sign."
- Use frame tags in Remix. "Put the jacket from
<frame>2</frame>on the person in<frame>1</frame>" — explicit addressing is the whole point of the syntax. - Feed clean references. One clear subject per reference image; clutter degrades targeting.
- Iterate in small steps. Lossless editing is the model's edge — five precise edits outperform one paragraph-long instruction.
What is Reve pricing like?
Only the edit-fast figure is rock solid: $0.01 per edit. Reve publishes no official per-image price list; third-party hosts quote roughly $0.20–$0.25 per full-resolution generation for Reve 2.1 — reseller listings, not official rates, so treat them as directional. The first-party app includes a free tier.
What should you make with Reve?
Native 4K output makes it a finals-quality engine inside an AI image generator workflow — no upscaling pass between generation and delivery. The penny-edit economics matter most in a photo editor context: object swaps, relights, and text corrections priced like operations instead of regenerations. And the combination of brand-element preservation and relighting is effectively purpose-built for product photography — same product, new season, new lighting, label intact.
Five questions about Reve
What is Reve AI?
An image model family from a Palo Alto lab, currently at version 2.1 (mid-July 2026), distinguished by a plan-then-render architecture, native 4K output, and unusually cheap instruction-based editing ($0.01 per edit).
Is Reve better for generating or editing images?
It does both, but editing is the standout — public registry run counts show its edit endpoints out-running its create endpoint, and its lossless iterative edits preserve typography and brand elements.
How many reference images does Reve Remix support?
Up to 8 on the reference implementation, addressed with <frame>N</frame> tags in the prompt — though some API hosts cap it at 4, so the limit depends on where you run it.
Does Reve really output native 4K?
Reve states 2.1 generates 4K×4K (~16MP) natively with no upscaling step, and calls it the fastest 4K image model available. Both are vendor claims from Reve's own model page, not independently benchmarked as of August 2026.
How much does Reve cost per image?
Officially unpublished. The verified figure is $0.01 per edit on Edit Fast; third-party hosts list full generations around $0.20–$0.25, which should be treated as unofficial. Reve's own app has a free tier.
Where is Reve available?
Every flavor of the family — Reve 2.1, Reve, Reve Remix, Reve Remix Fast, and Reve Edit Fast — sits in the Image tab of the invideo agent's model picker, a natural home for a model that behaves like an editor: generate a frame with whichever of the platform's 200+ models fits, then hand it to Reve for penny-priced corrections. Family details and examples are collected at invideo.io/ai-models/reve-ai.
Version history: "Halfmoon" preview Mar 2025 → Reve 1.0 (2025) → Reve 2.1 with ~3× parameters, Create/Edit/Remix, native 4K, mid-Jul 2026 (date per partner launch posts).