Which GPT Image 1.5 settings actually change detail, what high quality costs per image, where you can try it for free, and tested prompt examples for portraits, products, landscapes and food photography.
Most people who say GPT Image 1.5 looks "soft" or "plasticky" never touched the quality setting. The default is low, and low is built for speed, not detail. Flip it to high, write a prompt that names real materials and real light, and the same model suddenly produces skin pores, stitched seams and believable reflections. This article shows exactly which settings matter, what high quality costs, where you can try the model without paying, and prompts you can copy today.
💡 Quick answer: Set quality to
high, pick the aspect ratio that matches your final use, and describe light, lens and texture in your prompt. Those three moves account for most of the visible improvement.
"High fidelity" gets thrown around loosely. With GPT Image 1.5 it points to two separate things, and mixing them up is the first reason people get disappointing results.
The first meaning is output detail: how much fine texture, edge sharpness and tonal depth the image has. That is controlled by the quality setting. The second meaning is input faithfulness: how closely the result follows a reference photo you upload, especially faces. That is controlled by a different setting called input_fidelity. They are independent. You can run high quality with low input fidelity, and the other way around.
GPT Image 1.5 on PicassoIA offers low, medium, high and auto. Here is how they behave in practice:
| Tier | Best for | Typical trade-off |
|---|---|---|
| Low | Fast drafts, composition tests, prompt iteration | Softer micro-detail, simpler textures |
| Medium | Social posts, thumbnails, blog illustrations | Good balance of speed and detail |
| High | Final renders, print, portraits, product shots | Slower and noticeably more expensive |
| Auto | When you do not want to decide | The model picks, so results vary run to run |
The macro eye above shows what "high" aims for: individual lashes, skin texture and a believable catchlight, with nothing smeared into a smooth blob.
Here is every control worth touching, in the order I would adjust them.
Work in low while you build the prompt. Change one sentence at a time, check the composition, and only switch to high when the framing is right. This workflow keeps iteration cheap and still ends with a polished file. Because the platform default is low, anyone who never opens the settings panel is judging the model at its weakest tier.
The model supports 1:1, 3:2 and 2:3 on PicassoIA, which map to the native square, landscape and portrait sizes of 1024×1024, 1536×1024 and 1024×1536. Pick the one closest to your end use before generating. Cropping a square into a wide banner afterward throws away exactly the pixels you paid for.
💡 Rule of thumb: 2:3 for portraits and posters, 3:2 for landscapes and website heroes, 1:1 for products and avatars.
If you upload one or more images through input_images, set input_fidelity to high when the face or a specific product shape must survive. At low, the model treats your photo as loose inspiration. At high, it works harder to match style and facial features. Use high for likeness and brand-accurate products, and low when you only want the mood of a reference.
auto, transparent or opaque. Transparent is excellent for stickers, logos and product cutouts, and it saves a trip through a background remover.You can request up to 10 images per run. Generate four at low quality first, pick the strongest composition, then re-run that single prompt at high. It is the cheapest way to find a good frame.
Pricing comes from OpenAI's API, and PicassoIA runs the model for you, so you do not manage keys. The numbers below come from third-party price trackers that agree with each other, so treat them as a close estimate and confirm on OpenAI's pricing page before budgeting a large project.
| Quality | Square (1024×1024) | Portrait or landscape |
|---|---|---|
| Low | about $0.009 | slightly higher |
| Medium | about $0.034 | slightly higher |
| High | about $0.133 | about $0.20 |
The jump from medium to high is roughly four times in price, and from low to high it is more than ten times. That is why the draft-then-finish workflow matters. Ten high-quality experiments cost more than a hundred low ones.
Blog illustrations, social posts and slide decks rarely benefit from the extra cost. Medium gives most of the sharpness at a fraction of the price.
I searched for current, official information on free access to GPT Image 1.5 and found sources that disagree. One guide says free ChatGPT users get the same image model as paid users, while another says free tiers cannot reach HD quality. OpenAI's own pages did not give a clear, current answer in what I could find. So here is the honest version: free limits change often, and you should check the app itself for today's allowance.
If your goal is photoreal quality without spending, a few other models on the platform are worth trying side by side:
Some websites advertise "free GPT Image 1.5 High" but quietly run a lighter model behind the button. A simple test: generate a close-up of a hand holding a pair of glasses. Weak models bend fingers and warp frames. If the result looks sloppy, you are not seeing the real high tier.
The model page has everything in one panel. Here is the sequence I follow.
number_of_images at 3 or 4 to compare compositions.auto for scenes, choose transparent for cutouts.input_fidelity to high.number_of_images to 1.💡 If you need more resolution afterward, a super resolution model can upscale the finished file 2x to 4x without regenerating it.
High quality rewards specific prompts. The formula I use is:
Subject + action + setting + light direction + lens + texture details + film or color feel
The sample photographs below show the target look for each prompt type. They illustrate the style you are aiming for, and the prompts underneath are written to push GPT Image 1.5 toward the same result.
Candid street portrait of a woman in a camel wool trench coat standing at a crosswalk in a European old town at golden hour, wet cobblestones reflecting warm light, pastel facades softly blurred behind her, low sun backlighting her hair, 85mm f/1.4 lens, natural skin with fine pores and peach fuzz, visible wool weave, subtle film grain, Kodak Portra 400 color
Why it works: a named lens, a light direction and two texture cues keep skin from turning waxy.
Close-up of a potter's hands holding a freshly glazed teapot on a rustic workbench, glaze drips and fingerprints visible, soft northern window light from the left, shelves of pottery blurred behind, 85mm f/1.8, glossy and matte surfaces side by side, warm earthy palette, fine film grain
Why it works: contrasting finishes (glossy and matte) force the model to render material differences, which is where high quality shows.
Wide photograph of a rocky coastline at dawn with a stone lighthouse, sea mist, wet dark rocks with barnacles and tide pools in the foreground, pale pink and blue sky, sunrise light from the right, 24mm f/8, deep focus, fine foam and stone texture, natural color grading
Why it works: foreground detail plus a distant subject gives the model depth to work with. Use the 3:2 ratio here.
Close-up of a chef's hands plating seared scallops with microgreens on a matte stoneware plate, light steam rising, stainless steel kitchen blurred behind, warm overhead tungsten light with soft side fill from the left, 100mm macro f/2.8, glistening sauce, crisp caramelized crust
Why it works: food depends on moisture and crust. Words like glistening and caramelized tell the model where the highlights belong.
Single ceramic coffee mug with a matte sage glaze, three-quarter view, soft studio lighting, crisp edge detail, realistic contact shadow removed, isolated subject
Set background to transparent and output_format to png. The result drops straight into a design file.
Because low is the default, many first impressions come from the weakest setting. Compare the same prompt at low and high before deciding.
"Ultra, hyper, 8K, masterpiece, stunning" adds noise, not detail. Replace that pile with concrete physical facts: fabric type, surface finish, time of day, lens.
If you edit the subject, lighting and ratio together, you will not know which change helped. Alter one variable per run.
Uploading a portrait and leaving input_fidelity on low often gives a lookalike instead of the actual person. Raise it whenever likeness matters.
Image models still slip on long strings of text. Keep any in-image text to one or two short words, and add longer copy in a layout tool.
| Goal | Good pick |
|---|---|
| Controlled settings, transparency, reference faces | GPT Image 1.5 |
| Cheaper everyday drafts | GPT Image 1 Mini |
| Original-generation GPT look | GPT Image 1 |
| Large, sharp 4K frames | Seedream 4.5 |
| Fast edits and variations | Nano Banana |
Newer entries such as GPT Image 2.5 Flare and GPT Image 2.5 Sunburst are also available, so it is worth running one of your best prompts through each and comparing.
The fastest way to see what high fidelity means is to run one prompt twice. Open GPT Image 1.5 on PicassoIA, paste the portrait prompt above, generate a batch at low quality, then re-run your favorite at high. Put the two results next to each other. Then change the light direction, swap the lens, and watch how the skin, fabric and reflections respond. Ten minutes of side-by-side testing will teach you more than any settings table, and every prompt you refine becomes a template you can reuse for the next project.
