How to Generate Photorealistic Images with Krea 2
Generate photoreal portraits, product shots, and concept art with Krea 2 using light, texture, crop checks, and a one-variable refinement loop.

A photoreal image fails in the details: plastic skin, dead materials, missing shadow falloff, or a shallow focal stack that never feels photographed. Clean composition alone does not save it. This page is the photoreal workflow spoke. Product definition and speed tiers live on what Krea 2 is. Prompt wording order lives in the prompt guide. Here the unique job is subject-class tactics, material review, and a refinement loop that survives full-size QA.
A photoreal frame is an image that still reads as shot or art-directed when zoomed and when checked for material logic. A material review is a pass that checks light, texture, edges, and reflections before you accept a keeper.
That one-purpose split follows Google's helpful content guidance.
Key Takeaways
- Choose one subject class before writing the brief: portrait, product, environment, or lifestyle.
- Light and material language beat long "ultra realistic" adjective stacks.
- Review at full size and reduced size; photoreal can fail in either view.
- Iterate with one variable at a time, usually light or angle first.
- Package art still needs poster, cover, or quote checks after the photoreal plate is locked.
- Use the hub for tiers and the prompt guide for stack order.
What "photoreal with Krea 2" means in practice
| Review signal | Pass condition | Common fail |
|---|---|---|
| Light logic | Source, direction, and falloff agree | Flat beauty fill with no motive |
| Material | Surface roughness and reflection match object | Wax skin, chrome matte plastics |
| Edge detail | Contours stay clean under zoom | Soft mush or aliased edges |
| Depth | Foreground to background separation feels optical | Everything equally sharp |
| Crop honesty | Frame still works in the real publish ratio | Pretty 16:9 forced into 1:1 |
Text-to-image systems improved surface fidelity quickly, while control and evaluation of fine structure remain hard (Zhang et al., arXiv 2025; Croitoru et al., diffusion survey). Photoreal work is therefore a review process, not a single lucky generation.
Step 1: Choose the subject class before any style words
Different photoreal jobs need different success tests:
- Portrait / character: skin, eyes, hair edges, catchlights
- Product / still life: material accuracy, reflection control, clean silhouette
- Environment / concept: scale cues, atmospheric layers, light paths
- Editorial / lifestyle: people plus place, readable gesture and context
Pick one class first. Mixing "hero product + cinematic crowd scene + magazine cover" into one prompt usually produces average everything.
Step 2: Build a photoreal brief from light and material
Use prompt-stack order from the prompt guide, then specialize the fields for realism:
- subject noun
- material or wardrobe truth
- place and time of day
- light source and quality
- lens and depth cues
- crop for the first channel
Portrait recipe
Close-up portrait of a woman by a window, soft directional daylight, 85mm lens look, shallow depth of field, natural skin texture, calm expression, 3:4
Product recipe
Studio product shot of a matte ceramic mug on dark stone, single softbox key, controlled reflection on rim only, crisp focus, minimal background, 1:1
Environment recipe
Misty mountain pass at dawn, volumetric light through layered haze, wet rock texture in foreground, wide establishing camera, painterly realism, 16:9
Avoid padding with "8k, ultra detailed, masterpiece." Prefer measurable optical and material cues.
Step 3: Generate directions, not infinite noise
Photoreal quality comes from a short branch, not from re-rolling forever.
Practical batch plan:
- two light directions
- two angles
- one crop family only
- keep two candidates max before refinement
Supported crops still matter: 1:1 for product tiles and avatars, 3:4 for portraits, 16:9 for cinematic environments, 9:16 for vertical stories. Ratio families used in production and print workflows are covered in practical references (poster size chart; print aspect ratios guide).
For explore-stabilize-finalize economics, use the shared map on the product hub. Do not relearn tier lectures here.
Step 4: Run a full-size material review
Use this testing method before you call a frame photoreal:
- Open at 100% and inspect skin or surface microstructure.
- Check whether the light source implied by highlights exists in the scene.
- Trace reflection behavior on metal, ceramic, glass, or wet stone.
- Confirm hair, fabric, or object edges do not smear into halos.
- Reduce to thumbnail size and verify silhouette and focal subject still hold.
- Compare against the job class: portrait eyes, product logo plane, environment scale cue.
If the image only looks real as a small card, it is a social thumbnail, not a photoreal master.
Step 5: Refine the winner with one variable
Send the best candidate back as a reference. Change only one of these at a time:
- light quality or direction
- camera height or lens feel
- material roughness
- palette temperature
- background calm vs activity
Common repairs:
| Symptom | Likely cause | One-variable repair |
|---|---|---|
| Plastic face | Beauty light and over-smooth skin | Softer contrast + pore texture language |
| Dead product | No key motivation | Add single softbox or window key |
| Fake environment | Missing atmosphere layers | Add volumetric depth and scale cue |
| Correct mood, wrong crop | Style locked first | Rebuild in the real publish ratio |
Artifact patterns such as over-smooth skin, broken reflections, or smeared microdetail still appear in current model outputs; practical artifact writeups such as the ZSky artifacts guide help name what you are seeing during review.
Subject-class playbooks
Portrait playbook
- prioritize catchlights and eyelid structure over background drama
- prefer 85mm or three-quarter feel for faces
- keep wardrobe materials concrete: linen, leather, wool
- reject wax skin even when pose is perfect
Product playbook
- decide matte vs gloss before style words
- control reflections; do not ask for "shiny" without direction
- keep background quieter than the hero silhouette
- check the logo plane or label plane if branding will sit there later
Environment / concept playbook
- include one scale cue: figure, vehicle, doorway, rock ledge
- define weather and light path through volume
- hold a simple palette so atmosphere stays readable
- review mid-ground separation, not only the hero prop
Lifestyle / editorial playbook
- show gesture and context, not a floating mannequin
- choose time-of-day light that matches the story
- keep one focal action; secondary props should not compete
Campaign hierarchy thinking still helps after the photoreal plate is locked: dominant motif, readable contrast, intentional empty areas (movie poster anatomy overview). If the plate will become packaging or social type, hand off to the poster, cover, or quote guides.
Production checklist and rights note
Before publish or client handoff:
- subject class matches the job brief
- light logic passes at 100% zoom
- materials pass roughness and reflection checks
- crop matches the first channel
- winner refined with one-variable references
- commercial use reviewed against plan terms on pricing
- remaining rights or account questions can go through contact
Treat AI photoreal output as a strong production plate. Final retouch, legal review, and brand marks still live in the design toolchain.
Common failure modes and fixes
| Failure | Cause | Fix |
|---|---|---|
| "Looks AI" at full size | Texture and light never specified | Rewrite material and light fields |
| Great small, fake large | No full-size review | Always inspect at 100% |
| Endless re-rolls | Changing everything each pass | Reference lock; change one field |
| Pretty plate fails package | Photoreal only, no type mass | Apply packaging spoke rules next |
| Wrong tool expectations | Need novelty spam, not realism | Use the comparison guide |
FAQ
Is "photorealistic" enough as a prompt word?
No. The model needs subject, material, light, and lens cues. The word alone is a preference, not a setup.
Should I start with portraits or products?
Start with the subject you must ship first. Portrait work stresses skin and eyes. Product work stresses reflection and silhouette. They teach different reviews.
How many variations should I generate?
Enough to cover two light and two angle options inside one crop family. More than that usually burns budget without better decisions.
How is this different from the prompt guide?
This page owns photoreal subject workflows and material review. The prompt guide owns wording order and general control language.
Where do speed tiers fit?
In the explore-stabilize-finalize path on the product hub. This page stays on photoreal QA and refinement.
Bottom line
Generating photorealistic images with Krea 2 is a directed production loop:
- pick one subject class
- write light-and-material briefs
- generate a short branch of directions
- review materials at full size
- refine one variable at a time
- hand off package art to the matching use-case guide
Open the generator, start from a recipe above, and refuse frames that only look real as thumbnails. For product fit return to the hub; for prompt stack order use the prompt guide.