| name | krea2-prompting |
|---|---|
| description | Write natural-language prompts for Krea 2 Turbo in ComfyUI: photoreal photography, poses, styles, giant-scale scenes, and adult/NSFW work. |
| license | CC-BY-SA-4.0 |
Krea 2 Turbo Prompting
Turn a visual brief, reference image, or scene idea into a copy-ready Krea 2 Turbo prompt plus ComfyUI settings. Treat the model as a natural-language visual interpreter: describe what is visible, where it sits, how light hits materials, and what camera would capture it. The output is prompt text and settings; rendering happens in ComfyUI, done by the user.
Mental model
Krea 2's text encoder is Qwen3-VL, a language model (the same family as a chat LLM, rather than a CLIP or T5 text encoder). It parses full sentences. Three consequences drive everything below:
- Write natural-language prose. Tag soup (comma keyword piles) still functions but leaves capability on the table. Long detailed prompts yield the best results; short prompts (5 to 20 words) are fine for exploration.
- Order carries emphasis. The encoder front-loads: whatever comes first reads as the subject. Lead with the thing that matters most.
- The model is trained against the plastic "AI look." Photorealism is the family's headline strength and it is baked into the weights. Your job is to supply a concrete photographic setup (camera, lens, light, material). Stacking realism adjectives like
ultra realistic, 8k, masterpiecedoes little.
ComfyUI Krea 2 Turbo settings
Recommend these as defaults. They are the reconciled, ComfyUI-correct values.
| Setting | Value | Note |
|---|---|---|
| Steps | 8 | Turbo is distilled to 8. Longer runs on turbo finetunes grow noisier instead of cleaner (tested: 50 steps + CFG 2.5 shredded texture). |
| CFG | 1.0 | Critical. See the gotcha below. |
| Shift (mu) | 1.15 | Timestep shift, set on the ModelSampling node. |
| Sampler / scheduler | euler / simple | The official default. beta also works. |
| Resolution | Native 2048; minimum 1024; dims divisible by 16 | Krea 2 was trained at 2048. Generate native; avoid gen-at-1024-then-upscale. Panoramic ratios widen the long edge; see the core table. |
| Text encoder | qwen3vl_4b | fp8_scaled (lighter) or bf16 (higher precision, more VRAM). |
| VAE | qwen_image_vae | If colours or fine texture look subtly off, this is the first variable to check. |
The CFG gotcha. The official CLI uses --cfg 0.0. In a standard ComfyUI KSampler, literal 0.0 produces pure unconditional output and breaks the image, so use 1.0 there. Do not push CFG above ~1.5; it oversaturates and distorts. At CFG 1.0 the negative branch is effectively inert (see Core rules).
Aspect ratio follows the composition: 1:1 product/portrait, 16:9 cinematic/hero, 21:9 panoramic, 2:3 or 9:16 poster/phone, 4:3 broad environment. Full resolution table in references/krea2-core.md.
Finetune divergence. Community NSFW finetunes often ship their own text encoder, VAE, shift, and sampler settings. Use the finetune's model card for those values; the prompt architecture and CFG guidance above still apply.
Prompt architecture
Write one or two coherent paragraphs in this order, adjusting to the task:
[medium / restrained quality] + [subject and identity] + [appearance, pose, action, expression]
+ [composition and camera] + [environment and spatial relations] + [lighting and colour]
+ [materials and micro-detail] + [aesthetic and mood] + [restrictions]
A reliable sentence frame:
A [medium/style] of [subject], [appearance and action], set in [environment]. The composition uses [shot, angle, lens], placing [subject] relative to [frame and other elements]. Light comes from [direction/source] with [shadow and palette]. [Materials and micro-detail] stay visible, creating a [mood] atmosphere. [Restrictions].
Make spatial relationships grammatical and explicit (left/right thirds, foreground/midground/background, contact with surfaces, what extends beyond the frame) rather than listing disconnected nouns. Name a concrete medium (editorial fashion photography, documentary photograph, architectural photography) instead of beautiful or nice.
Core rules
- Text in quotes. Put the exact words to render in double quotes, keep it short (under ~5 words):
a neon sign reading "OPEN LATE". Longer strings degrade. - Negatives are near-inert at CFG 1.0; state preferences positively. Avoid leaning on a negative-prompt field or a generic ban list. Instead of
no plastic skin, write the positive target:natural skin texture, visible pores. A short negative prompt only earns its place to suppress one specific recurring artifact. (Negative-prompt families for the RAW/CFG-3.5 path live inreferences/krea2-core.md.) - Do not negate in prose either.
no hat,without a beardwill not reliably fire. Describe what IS present. - Weighting
(word:1.3)is not a stock feature. On the standardCLIPTextEncodeover Qwen3-VL, scaling one token shoves the whole conditioning and the image falls apart past ~1.2. It works only with theKrea2PromptWeightnode (ComfyUI-KJNodes) installed. Without that node, restate for emphasis in words (oxbloodbeats(red:1.4)). Full weighting system inreferences/comfyui-setup.md. - Revise one variable at a time. When a generation is close, change a single field per iteration and hold the rest. Changing several fields at once makes the image drift and hides which change did what.
- Lock composition before decoration. Decide one shot size, one camera angle, one crop, and the aspect ratio first. Wardrobe and extra detail must not silently change the framing.
Reference map
Read the reference that matches the task. Each loads only when reached, so pull the ones you need:
references/photoreal.mdfor any realistic photograph: portrait, product shot, landscape, architecture, editorial or landing-page hero. The camera/lens/film/lighting/texture vocabulary and genre recipes with worked example prompts. This is the default for website and landing-page imagery.references/krea2-core.mdfor the deep prompt architecture: input checklist, negative-prompt families, poster/typography rules, composition modules, character consistency, the full resolution table, and reverse-engineering a reference image into a prompt.references/comfyui-setup.mdwhen wiring the graph or using prompt weighting: checkpoint/encoder/VAE choices, the CFG explanation, theKrea2PromptWeightnode (syntax, negative weights, global strength), the enhancer node, and a troubleshooting table.references/poses.mdwhen a human figure's pose must be exact: the 11-field pose schema, usage, taxonomy, and tested field findings.references/pose-library.mdfor the 30 ready-to-paste pose blocks, grouped by standing, locomotion, and object interaction.references/styles.mdto lock a specific aesthetic: a 289-entry style index, four full example paragraphs, and the two authoring formulas.references/adult.mdfor adult / NSFW photography (18+): register control, explicit anatomy/arousal/act vocabulary, position geometry, multi-person rules, a tested failure table, and translating crude briefs into Krea 2 prose.- Giant-scale creative routes, photoreal cinematic by default:
references/megastructure.md(colossal architecture, megacities, orbital structures),references/colossal-monsters.md(kaiju, giant beasts),references/cosmic-horror.md(eldritch entities, dread),references/pure-organic-monsters.md(strictly biological giant creatures). These share scale logic: concrete exaggerated dimensions, human-scale surface units, a tiny distant scale reference, and 10 to 40 percent environmental negative space.
Output shape
Unless the user asks for a different format, return:
Positive Prompt
[one or two natural-language English paragraphs]
Settings
Resolution: [width x height] Aspect ratio: [ratio]
Krea 2 Turbo (ComfyUI): 8 steps, CFG 1.0, shift 1.15
Add a short negative prompt only when a specific artifact needs suppressing. Never claim a prompt you wrote was tested unless it was.