Reve Image: how to write prompts the model actually understands

What Reve Image does well

Reve's core strength is exceptional prompt fidelity: the model builds exactly what's described, without «creative reinterpretation». That's a rare property — most top models add their own composition, lighting, and incidental details. Reve follows the prompt literally.

Second feature: a proprietary Typography Engine trained on 50M font samples, delivering 98% text accuracy — higher than any competitor. Third: hyperrealistic portraits, especially strong with diverse ethnic features and celebrity likeness. Fourth: full commercial rights on all output, no restrictions.

Available on two platforms: preview.reve.art (official) and Higgsfield (unlimited generations on subscription). Architecture: Hybrid Diffusion with relational attention.

• 12B parameters, native 2048×2048, 4K upscaling (4096×4096)
• #1 on Artificial Analysis Image Arena (ELO 1167)
• 98% text accuracy via Typography Engine
• Exceptional prompt fidelity — builds exactly what you describe
• Full commercial rights on output

Prompt structure

Photographic-style prompting works best. Base formula: [Photographic framing] + [Subject with details] + [Lighting/atmosphere] + [Texture/materials] + [Camera/lens hints].

Lead with clear photographic framing: «A high-resolution, photorealistic studio portrait, captured with shallow depth of field...». That locks composition and quality immediately. Then subject with concrete detail, lighting and atmosphere, textures and materials, photographic hints about the camera.

Reve understands photography language well: «shift focus to the man», «render as if shot at 1.8 f-stop», «captured at 85mm». These aren't decoration — the model actually factors them into composition and focus.

Photography language

Camera: studio portrait, shallow DoF, 1.8 f-stop, 85mm, wide-angle, macro. Lighting: warm side light, foggy dusk, golden hour, studio softbox, rim light, hard flash. Textures: skin pores, fabric weave, weathered wood, brushed metal, frosted glass. Mood: cinematic, editorial, intimate, dramatic, serene, melancholic.

Reve processes these terms like a real photographer — every parameter influences the final frame. Specific «warm side light, foggy dusk atmosphere» yields a very different result than «nice lighting». The more precise the language, the more precise the result — that's a Reve strength.

Text and the Typography Engine

The Typography Engine with 50M font samples delivers 98% accuracy — the best in the industry. To render text, specify it explicitly in the prompt: «A storefront sign reading "Bloom & Co" in serif gold lettering». The more specific the font type, color, and placement, the more stable the result.

For branding and packaging this is a deciding feature: you can generate posters, signage, and labels with production-grade text in a single pass, no separate typography step. Latin, digits, and main European diacritics are supported. For CJK and Arabic, Qwen Image or GPT Image 2 are stronger.

Common mistakes

1. 1. Minimal prompts for commercial work

   Reve handles short prompts well thanks to strong inference, but for commercial results specificity always wins. «A portrait of a woman» yields something decent; «A photorealistic studio portrait of a woman with auburn hair, captured at 85mm f/1.8, warm side light» yields production grade in one iteration.

2. 2. Complex transparent objects

   A full wine glass, crystals, glass structures with many refractions — Reve's weak spot. The model can render the optics of liquid and surfaces incorrectly. If the task demands complex glass, either simplify the scene (empty glass) or combine Reve with another model at the final stage.

3. 3. Specific fictional characters

   «Mario from Mario Bros», «Pikachu», «character from Fortnite» — Reve usually yields generic results because it doesn't know every franchise. Celebrity likeness works well (Reve was trained on real faces), but recognizable IP characters are weaker. For franchise content, describe the character physically rather than by name.

4. 4. Multi-object compositions

   Scenes with 5+ objects and complex placement can come out with wrong spatial relations on Reve. The model's strength is simple centered scenes with one main subject. For complex compositions, either simplify or use ControlNet via another model and then transfer the style to Reve via I2I.

5. 5. Evaluative adjectives instead of descriptions

   «Beautiful», «stunning», «amazing» are noise for Reve. The model values photographic precision over emotional adjectives. Replace «beautiful portrait» with «studio portrait with warm side light and shallow DoF, captured at 85mm». That gives a concrete frame instead of a random «beautiful» result.

Before / after examples

beautiful portrait of a girl in a studio

A high-resolution, photorealistic studio portrait of a young woman with auburn hair and soft freckles, captured with shallow depth of field. Warm side light from camera left, soft falloff into deep shadow on the right cheek. Visible skin texture, fine pores, natural makeup. Cream linen background, gentle bokeh. Shot at 85mm, f/1.8, intimate framing from collarbone up.

coffee shop sign with the words Bloom & Co

A photorealistic storefront facade at golden hour. Vintage cafe with a glass door and a wooden window frame. Brass storefront sign mounted above the door (EXACT text): "Bloom & Co" in elegant serif gold lettering with subtle aging patina. Soft warm light spilling from inside, slight reflections on the glass. Captured at 35mm, slight low angle, cinematic warm color palette.

product shot of a perfume bottle

Macro product photography of a frosted glass perfume bottle on a polished marble surface. Side light from a softbox at camera left, subtle rim light highlighting the bottle's curve. Visible frosted glass texture, brushed gold cap, single drop of condensation on the bottle's shoulder. Shallow depth of field, focus on the cap engraving. Cream background gradient, editorial composition.