GPT Image 2.5 for D&D: Flare vs Sunburst in CharGen

GPT Image 2.5 for D&D is useful when I need a character portrait, a campaign scene, a handout or a careful edit without leaving the CharGen workflow. The important choice is not the newest model name. It is whether I need a fast working image or a slower, more exact revision.

CharGen now offers GPT Image 2.5 Flare and GPT Image 2.5 Sunburst in Generate and Quick Edit. Flare is the everyday option. Sunburst is the precision option. Both have Low, Medium, High, XHigh and Max quality settings, and both let me choose the image size before I spend Gold.

The CharGen GPT Image 2.5 Flare page is the quickest place to start. I can select a model, add a prompt or reference, check the quote, and make a usable image for my next session. I do not need a perfect prompt on the first try. I need a prompt I can judge.

GPT Image 2.5 for D&D at a D&D table

I judge an image model by the job it completes at the table. A portrait needs a clear face and a recognisable silhouette. A scene needs room for the people, props and action that the session depends on. A handout needs a readable visual structure, but I still add important rules text myself rather than trusting generated lettering.

OpenAI describes Flare as the faster high-quality everyday model and Sunburst as the model for detailed edits and tighter control. Its GPT Image 2.5 announcement also calls out better reference fidelity and precision editing. I treat those claims as useful direction, then judge the result against my own campaign brief.

My usual D&D image jobs fall into four groups:

A model choice cannot rescue a vague brief. I still name the subject, camera view, action, setting, lighting, style and elements that must stay out. The model gives me a different balance of speed and control, not a replacement for decisions.

Flare and Sunburst do different jobs

Flare is where I begin when I am still exploring. I can ask for three related views of the same ranger, test a tavern scene, or try a new outfit without making every draft expensive or slow. The result is often enough for a session note, an NPC introduction, a token or a visual reference for the next prompt.

Sunburst earns its place when the image already works and the next request is narrow. I might need to keep a scar, a silver clasp, a specific relic or the shape of a familiar face while changing the background. I might also need a dense prop arrangement where small errors matter. In those cases, I want more control over what moves.

OpenAI's image generation guide lists the quality and size controls that sit behind the product. CharGen turns those controls into a Gold quote before generation. I use the quote as a practical boundary, not as a reason to avoid testing.

Choose quality and size before the prompt

Quality and size change the price, the wait and the kind of checking I need to do. I set them before I polish the prompt because they define the output I am asking the model to make.

For a 1024 x 1024 Flare image, the current CharGen prices are:

The figure is a 1024 x 1024 example, not a promise that every request has the same price. Size, quality and image count all affect the quote. Quick Edit and Sunburst use Gold too. I read the displayed estimate immediately before I generate.

I use Low for a quick composition check when the image will only guide a later pass. Medium is a useful everyday test. High is my normal starting point for a portrait I might keep. XHigh and Max make more sense when detail is the point, not when I am still deciding whether the character should hold a sword or a lantern.

Size needs the same discipline. CharGen supports square, portrait and wide formats, from 1024 x 1024 to options with a 3840-pixel long edge. Supported examples include 1536 x 1024, 1024 x 1536, 2048 x 2048, 2048 x 1152, 3840 x 2160 and 2160 x 3840. The picker shows the actual pixel dimensions.

I match the size to the destination:

My Flare workflow for character art

I use the same short loop for most D&D portraits. The model can produce a dramatic first image, but the loop gives me a way to decide whether it belongs in the campaign.

Name the table job

I start with the job, not the style. “A portrait for the party's new half-elf ranger” is useful because it tells me what the image must communicate. “Fantasy art” is too broad.

My opening brief might be:

A half-elf ranger for a level 5 D&D party, shown from the chest up, alert but tired after a night in the marsh. Dark green cloak, braided black hair, a small silver compass at the collar, longbow strap visible, wet reeds behind her, cool morning light.

The brief gives the model a subject, crop, mood, props and setting. I can add a visual treatment after the visible facts are stable.

Keep the fixed facts together

I place the details that must survive future edits in one sentence. Face shape, hair, scar, eye colour, outfit, signature prop and age all deserve a clear place. I do not scatter them across a long paragraph where they compete with atmosphere.

For a recurring NPC, I keep a short reference note beside the prompt. I use the same spelling and the same core details each time. Consistency comes from a repeatable brief and references, not from hoping a model remembers yesterday's image.

Add references with a job

One reference can show the face. Another can show a cloak clasp or a sword. I say what each reference contributes. “Use reference one for the face and hair. Use reference two for the bronze clasp. Keep the pose and marsh setting from the written brief.”

The character consistency workflow helps when one NPC needs to appear across portraits, scenes and campaign notes. I still review each output. A reference supports the prompt, but it does not remove the need for a visual check.

Choose the smallest useful test

I start at 1024 x 1024 and Medium for a new character unless the destination says otherwise. A small square test tells me whether the face, silhouette, clothing and prop are working. If the brief is wrong, a larger image only gives me a more expensive wrong answer.

Flare is especially useful at this stage because I want to compare directions. I might test a quiet portrait, an action pose and a group scene. I keep the subject facts fixed and change one visual decision at a time.

Change one thing per edit

When the image is close, I stop rewriting the whole prompt. I ask for one controlled change:

Keep the face, hair, cloak, compass and lighting. Replace the bow with a short spear. Keep the same chest-up crop and marsh background.

If the identity shifts, I know which part of the request caused trouble. If I change the weapon, background, pose and art treatment at once, I cannot tell what worked.

Save the winning handoff

I keep the final prompt and the selected reference with the campaign note. A portrait becomes more useful when I can reuse its fixed facts for a token, an encounter scene or a session recap. The Token Maker is a natural next step when the portrait needs to reach a VTT.

When Sunburst earns the extra wait

I switch to Sunburst when the cost of a changed detail is higher than the cost of a slower render. A final villain portrait may need the same face, crown, eye glow and ceremonial blade after I revise the room. A magic item may need its gem setting, engraving shape and leather wrap to stay in place. A group scene may need the characters to remain distinct while I adjust the weather.

My Sunburst requests are more conservative. I state what must stay in the opening sentence, then name the one change. I also describe the reference images by role. If I want a darker sky, I do not add a new costume, a new camera view and a new painting style in the same instruction.

The model is not a licence to ask for microscopic control over every pixel. I still accept a changed background texture or a slightly different fold in the cloak. I reserve it for changes that affect the identity or the campaign meaning of the image.

I also use Sunburst for a final edit when the first Flare image has the right character but the wrong practical detail. The face may be perfect, but the hand has an extra ring. The outfit may be right, but the relic is hidden by a fold. I name the defect, state what remains fixed, and make one edit.

Prompt examples I would actually use

The following prompts keep the subject facts separate from the visual choice. I can run the first with Flare, then try a controlled edit with Sunburst.

A player character portrait

A chest-up portrait of a human paladin for a D&D campaign. Mid-thirties, close-cropped dark hair, a thin scar over the left eyebrow, blue enamel shield badge, worn cream surcoat over chainmail, calm expression after a difficult victory. Soft firelight from the left, dark stone chapel behind, realistic painterly fantasy art. Keep the face and shield badge clear. No readable text, no logo, no extra limbs.

A location scene

A wide D&D campaign scene inside a flooded archive beneath a ruined monastery. Three adventurers stand on a stone shelf while floating books circle a sealed iron door. Blue lantern light reflects on shallow water. Leave open space near the top for session notes. The door, waterline and lanterns must be visible. No readable writing on the books.

A handout illustration

A square fantasy illustration for a D&D puzzle handout: four brass bells around a cracked moon dial, silver fish mosaics in the floor, damp stone, one candle and a clear central apparatus. Use simple shapes and strong contrast so the components remain easy to point at during play. No words, no letters, no numbers.

A controlled Sunburst edit

Keep the character's face, braided hair, dark green cloak, silver compass and marsh lighting exactly as shown. Change only the weapon to a short spear with a wrapped leather grip. Keep the chest-up crop, the same expression and the same background. Do not add text.

I do not ask a model to create the final rules handout with critical instructions embedded in the artwork. I create the visual, then add the exact text in a design tool or document. Generated lettering can be useful for mood, but it is not a reliable place for an initiative order, a spell name or a clue the party must read.

References, edits and campaign consistency

The best workflow is not “generate once and move on”. I make a small set of images that share a character or place, then check whether the set still reads as one campaign.

For a character set, I check:

• face shape and hair;
• ancestry cues and age;
• signature clothing;
• weapon or magic item;
• palette and lighting;
• crop and camera distance;
• expression appropriate to the scene.

Quick Edit is useful when I have one strong image and a specific correction. Flare is a good place to explore a new direction. Sunburst is a better place to finish a precise revision. The AI image model comparison gives a wider view of model selection, but I use the Flare and Sunburst split when the request is already inside CharGen.

I also keep an eye on crop safety. A portrait can look excellent at full size and fail when the face is reduced to a token. I check the intended destination before I call an image finished. If the VTT needs a circular token, I leave breathing room around the head and shoulders. If the art will sit in a session card, I keep the important prop away from the edge.

Gold allowances and limits

Elite includes 1,000 GPT Image 2.5 Flare Low images at 1024 x 1024 each month. Ultimate includes 3,000. The allowance is separate from the existing GPT Image 2 Low allowance, which is 5,000 for Elite and 15,000 for Ultimate.

The Flare allowance renews on the subscription month's date and does not roll over. A batch must fit the remaining allowance in full. Other Flare sizes, higher qualities, Sunburst, Quick Edit and API requests use Gold. Failed or cancelled Flare requests return the allowance to the same month.

I read those limits as part of the creative brief. A 1024 x 1024 Low draft can be a sensible test for an everyday idea. High quality, a large size or a precise Sunburst edit should have a real destination. The Gold estimate is there to help me make that choice before I submit.

OpenAI's GPT Image 2.5 Flare model page lists its API input and output modes and its quality choices. API prices and CharGen Gold prices are different accounting systems, so I do not convert one into the other. I use the CharGen quote for a CharGen request.

My pre-session check

Before a campaign night, I check the images in the place where I will show them. My list is short:

1. The character is recognisable at the smallest intended size.
2. The image has the crop I need, not only a nice full-frame composition.
3. Important props are visible and not hidden by hands or shadows.
4. Any rules text or handout copy was added separately and checked by me.
5. The prompt, reference and final file are saved beside the campaign note.
6. The image cost was clear before generation.

FAQ

What is GPT Image 2.5 for D&D?

GPT Image 2.5 is a family of image models available in CharGen for D&D portraits, scenes, handouts and edits. Flare is the faster everyday option, while Sunburst is aimed at detailed revisions and tighter control.

Should I choose GPT Image 2.5 Flare or Sunburst?

Choose Flare for quick character art, campaign scenes and early drafts. Choose Sunburst when a final image needs detailed revisions, complex props or stronger reference control.

How much does GPT Image 2.5 cost in CharGen?

A 1024 x 1024 GPT Image 2.5 Flare image costs 2, 3, 9, 14 or 30 Gold at Low, Medium, High, XHigh or Max quality. The displayed price changes with size, quality and image count.

What sizes does GPT Image 2.5 support?

CharGen supports GPT Image 2.5 sizes from 1024 x 1024 up to a 3840-pixel long edge, including square, portrait and wide formats. Every side must use a multiple of 16, and the exact options appear in the picker.

Can I use character references with GPT Image 2.5?

Yes. Use reference images for a character, outfit or prop, then describe what must stay and what should change. Flare is useful for a fast pass, while Sunburst suits a precise final edit.

Start with one controlled render

I would start with Flare, a 1024 x 1024 Medium request and one character brief. Check the face, silhouette, prop and crop. If the direction works, make the next image with the same fixed facts. If one small detail needs to change, try a focused edit with Sunburst.

Open the GPT Image 2.5 Flare workflow, read the Gold estimate, and make one image for the next session. A useful campaign set grows from decisions I can check, not from a pile of unreviewed renders.

Try GPT Image 2.5 Flare→