A lit candle is a deceptively hard product to shoot, and not for the reason most guides assume. The glass is not the problem anymore. The flame is not even the problem anymore. The problem is the flame's light. A real candle glows through its frosted jar, melts a small pool of wax around the wick, and throws a warm glow onto whatever sits near it. Get the flame shape right but miss the light, and the eye reads the whole image as fake in a fraction of a second.
So I tested it. I ran one brand-free brief, a lit scented candle in a frosted glass jar with a soft flame, a melted wax pool, and warm evening light, through four of the strongest image models with the same prompt: Nano Banana 2, GPT Image 2, Seedream 4.5, and FLUX.2 Pro. Every one drew a believable flame. They split sharply on whether that flame's light was real. This is the candle entry in our product-photography series, alongside the skincare, jewelry, supplements, makeup, food and beverage, footwear, clothing, and furniture tests and the broader best AI image model for product photography roundup.
Quick answer
- Frosted jar that should glow from within: GPT Image 2. It rendered the most convincing internal glow, the whole jar lit warm from inside.
- Premium macro hero with a real wax pool: Seedream 4.5. The best material, a genuine melted pool around the wick, warm light cast on the wood and wool, and the cheapest of the photoreal picks.
- Coziest full lifestyle scene: Nano Banana 2. The flame's glow integrated into a styled scene with lavender and cinnamon.
- Prettiest still jar, but watch the flame: FLUX.2 Pro. The most refined jar, but its flame has no visible wick, no melt pool, and casts almost no light. A decorative flame, not a lit one.
If you only remember one thing: drawing a flame is solved. Lighting the scene with that flame is not. Judge the glow, not the flame.
The test, model by model
One brief, four models, same prompt. I judged the flame's light first, then the glass, the wax pool, and the overall cozy read.
GPT Image 2 nailed the part that is actually hard. The frosted glass glows warm from within, the way a real candle backlights its own jar, and the flame sits in a soft halo that falls onto the surface around it. This is the headline test passed: the flame is not a sticker on top of the image, it is a light source the rest of the scene responds to. It is the most expensive and slowest model per image, but for a candle hero where the glow is the whole point, it earned it here.
Seedream 4.5 made the most premium image, the same way it did on skincare, jewelry, and supplements. The standout detail is the one most models skip: a genuine melted wax pool around the wick, glossy and slightly sunken, which is what sells a candle as actually lit rather than freshly poured. The frosted glass reads right, and the flame's warm glow falls believably on the wood and a wool throw. At roughly a fifth of GPT Image 2's cost, this is the value pick for a candle hero.
Nano Banana 2 built the most complete lifestyle scene. Rather than a tight macro, it staged the candle on a wooden table with a sprig of lavender, cinnamon sticks, and soft window light, and crucially the flame's glow ties that whole vignette together. The wax pool is believable. The one miss against the brief: it rendered a clear glass jar rather than a frosted one, so if your product is specifically a frosted vessel, that matters. But as a cozy, ready-to-style scene with real lighting, it is the most usable out of the box.
FLUX.2 Pro is the clean illustration of the trap this whole post is about. Taken as a still object, the jar is the most refined of the four: smooth frosted glass, a tasteful warm wax color, a calm composition. But the flame is the giveaway. There is no visible wick connecting it to the wax, no melted pool, and the flame throws almost no light onto the jar or the table. It is a pretty flame pasted onto a pretty jar, not a candle that is lit. This matches what FLUX does elsewhere in the series: a strong main object, but the physics-dependent details are where it slips. Cheapest per image, and fine if you will composite a real flame yourself, but do not trust it to light a scene.
The comparison
| Model | Flame's light (the test) | Glass | Wax pool | Best for | Rough cost/image |
|---|---|---|---|---|---|
| GPT Image 2 | Best, jar glows from within | Frosted, accurate | Present | Frosted-jar hero, glow | ~26.4 credits |
| Seedream 4.5 | Strong, warm cast on surroundings | Frosted, premium | Best, real pool | Premium macro hero, value | ~4.8 credits |
| Nano Banana 2 | Strong, integrated into full scene | Clear, not frosted | Present | Cozy lifestyle scene | ~9.3 credits |
| FLUX.2 Pro | Weak, decorative flame, no cast glow | Frosted, refined | Missing | Still jar, label-free | ~3.6 credits |
Credit costs are first-hand from this test on Masonry; per-image rates move, so check current pricing before you budget.
Why a candle is a lighting shot, not a product shot
Most product photography is about rendering an object faithfully. A candle is different because the product is also a light source, and that changes what "faithful" means.
The flame is solved; the flame's light is not. Every model in this test drew a convincing flame shape, because flames are common in training data and modern models are good at them. What separates a real candle image from a fake one is whether that flame behaves like light: a glow through the frosted glass, a warm halo on the surface, a melted pool where the heat has been working. GPT Image 2 and Seedream 4.5 rendered that physics. FLUX.2 Pro drew the flame and forgot the light. The eye catches the difference instantly even when it cannot name it.
Glass was not the weak point. This is worth flagging because candle-specific AI tools have historically warned that reflective glass jars are hard for background replacement, the vessel blends poorly into a swapped background. In from-scratch generation, all four models handled the frosted or clear glass fine. The vessel is not where these models struggle anymore. The lighting is.
The melt pool is the realism tell. A freshly poured candle has a flat, even wax surface. A lit one has a small molten pool around the wick. Seedream 4.5 rendered that pool, which is a big part of why its image reads as a candle in use rather than a candle on a shelf. If you want a "lit and cozy" mood, look for the pool.
How to shoot your candle line without a studio
The workflow is the one the product photography roundup lays out, applied to a lighting-driven product. Lead your judgment with the glow, not the flame. Run your jar through two models, because the one that lights your frosted vessel beautifully may flatten your clear or colored one. Look specifically for the wax pool and the cast light on the surface, the two details that separate a lit candle from a decorative one. And if a model gives you the perfect jar with a lifeless flame, that is a candidate to composite a real flame onto, not to discard.
With the Masonry CLI you can fire the same candle prompt at every model from one command and compare the glow side by side, which is exactly how the four images above were made:
masonry image "lit scented candle in a frosted glass jar, soft flame, melted wax pool, cozy wooden surface, warm evening light, macro" --model gpt-image-2 masonry image "lit scented candle in a frosted glass jar, soft flame, melted wax pool, cozy wooden surface, warm evening light, macro" --model seedream-4-5
The bottom line
A candle is a lighting shot wearing a product shot's clothes. Every modern model can draw the flame, so that is not where you choose. You choose on the light: GPT Image 2 for a frosted jar that glows from within, Seedream 4.5 for a premium macro with a real melted pool at a fifth of the cost, Nano Banana 2 for the coziest full scene. FLUX.2 Pro makes the prettiest still jar but the most decorative flame, so save it for when you will light the scene yourself. Run your own candle across two models from one place with the Masonry CLI, or see how the same fidelity-first logic plays out across every product type in our best AI image model for product photography roundup.
