Skip to main content

krea2-prompting

Prompting rules for the open-weight **Krea 2 Turbo** text-to-image model on the comfyui `krea2` / `krea2-2gpu` workflows. Use when the user asks to generate / draw / render an image on a Krea 2 workflow. Krea 2 is an aesthetic-first natural-language model (Qwen3-VL text encoder) - write descriptive prose or real descriptor phrases, not booru tags; pass no negative prompt; and leave the baked 8-step / cfg-1 recipe alone.

Ir a la instalación

Datos de origen

Repositorio
weikinhuang/dotfiles
Última actividad en el origen
31 de julio de 2026 a las 21:44
Idioma detectado de SKILL.md
inglés
Estrellas
21
Forks
3

Opciones de instalación

De forma predeterminada está seleccionado el prompt que primero revisa el origen. Puedes cambiar a un comando directo o descargar una copia local.

Revisa los archivos de origen

Lee SKILL.md y los archivos complementarios que muestra SkillsMP antes de decidir si quieres instalarlo.

Mostrando SKILL.md

SKILL.md
Instrucciones de origen · Vista previa de solo lectura
name
krea2-prompting
disable-model-invocation
true
description
Prompting rules for the open-weight **Krea 2 Turbo** text-to-image model on the comfyui `krea2` / `krea2-2gpu` workflows. Use when the user asks to generate / draw / render an image on a Krea 2 workflow. Krea 2 is an aesthetic-first natural-language model (Qwen3-VL text encoder) - write descriptive prose or real descriptor phrases, not booru tags; pass no negative prompt; and leave the baked 8-step / cfg-1 recipe alone.
# Krea 2 Turbo Prompting [Krea 2](https://huggingface.co/Comfy-Org/Krea-2) is Krea's open-weight, from-scratch 12B diffusion transformer (Qwen Image VAE + a **Qwen3-VL language-model text encoder**). The encoder is what gives it strong prompt adherence: it reads long, structured, natural-language descriptions rather than tag lists. The `krea2` workflow runs the **Turbo** checkpoint, an 8-step distilled model tuned for fast, high-quality photoreal output. Write the `prompt` you pass to `generate_image` per the rules below to drive it well on the first call. Two rules matter most: - **Natural language, not tags.** Describe the image as you would to a photographer - complete clauses or real descriptor phrases (`golden-hour backlight, 85mm, shallow depth of field`), never booru/quality tags (`1girl, masterpiece, 8k, score_9`). There is no quality prefix, no `@artist` syntax, no score tags. - **Aesthetic-first: trust it, prompt less.** Krea makes flattering art-direction choices on its own (rim light, colour grading, balanced framing). A short prompt yields variety; add structure only when you need control. **Safe default when unsure:** use the `krea2` workflow, write `prompt` as one or two descriptive sentences (subject, setting, lighting, style), set `width`/`height` (or `aspect`) for the framing, and pass nothing else - no `negative`, and do not touch `steps`/`cfg`/`seed` unless reproducing a prior image. ## The `generate_image` call | Arg | For Krea 2 Turbo | | ------------------ | ---------------------------------------------------------------------------------------------------------- | | `prompt` | Required. A natural-language description (flowing prose or comma-joined real descriptor phrases). No tags. | | `workflow` | `krea2` or `krea2-2gpu` (see below). Omit if `krea2` is the configured `defaultWorkflow`. | | `width` / `height` | ~1-4 MP: `1024x1024` square, `1216x832` landscape, `832x1216` portrait, up to `2048` on the long side. | | `aspect` | Preset (`16:9`, `portrait`, `square`, …) as an alternative to explicit `width`/`height`. | | `seed` | Omit for a fresh image; pass a prior seed to reproduce or to vary by tweaking the prompt. | | `count` | Batch size. | | `enhance` | Opt-in prompt rewrite before rendering (see [enhancement](#prompt-enhancement-enhance)). | Krea 2 maps **no `negative`** (it uses none) and **no `steps` / `cfg`** - the 8-step / cfg-1 distilled recipe is baked into the graph and is deliberately not exposed. Do not try to raise steps or cfg; it breaks the distilled model. Shape the result through the `prompt` and pick resolution via `width`/`height`/`aspect`. ### Which workflow | Workflow | Use it for | GPUs | | ------------ | ---------------------------------------------------------------------------------- | ------------------------- | | `krea2` | The default - whole pipeline on one GPU. | 1 | | `krea2-2gpu` | Same output, but text encoder + VAE run on a second GPU to free VRAM on the first. | 2 (+ multi-GPU node pack) | Pick `krea2` unless the host has two cards and you specifically want the split; the images are identical. ## Positive prompt recipe Write a coherent description, moving through these beats in roughly this order (as prose or descriptor phrases, not labelled fields): 1. **Subject + action** - who/what and what they are doing. `A weathered fisherman mends a net on a wooden dock.` 2. **Appearance + material** - concrete, depictable attributes: wardrobe, materials, age, expression, colour. 3. **Setting + composition** - where it is and how it is framed. Krea reads cinematic framing: shot size (`extreme close-up`, `portrait`, `full body`, `wide establishing shot`), angle (`eye level`, `low angle`, `three-quarter`, `top-down`), placement (`rule of thirds`, `off-centre`, `negative space`). 4. **Lighting** - do the heavy lifting here. Quality (`soft diffused`, `hard directional`, `golden hour`, `overcast`, `studio`, `volumetric`), direction (`backlit`, `rim light`, `side light`, `top light`), mood (`moody low-key`, `bright high-key`, `warm tungsten`, `cool blue hour`). 5. **Mood + atmosphere** - one phrase on the feeling. 6. **Medium + style** - name it explicitly (`editorial photograph`, `cinematic film still`, `oil painting`, `watercolour`, `anime illustration`) to avoid hybrid results. Name an era or film stock rather than an artist. 7. **Camera + texture detail** - lens (`35mm`, `85mm`), aperture (`f/1.4`), `shallow depth of field`, `soft bokeh`, `film grain`, and fine-texture cues (`natural skin texture`, `fabric weave`, `wet black stone`, `glossy reflections`, `micro-reflections`). You do not need every beat. Krea's aesthetic sense fills gaps; spend words on the specifics that matter for this image. ## How much to write Match length to intent: - **Explore (5-20 words):** a short prompt for variety - let the model's aesthetic drive. - **Controlled (30-80 words):** subject + composition + lighting + style, when you know what you want. - **Complex scene (80-140 words):** layered detail for a specific, busy composition. Prefer clear clauses over one long run-on. Flowing prose and comma-joined descriptor phrases both work - what matters is that the tokens are real descriptors, not tags or quality filler. ## Rendering text in the image Krea 2 renders short, legible text. Put the exact string in double quotes and keep it short - a word or a short sign renders cleanly, long paragraphs degrade: ```text A vintage enamel shop sign reading "FRESH COFFEE" in cream lettering on a dark green background, mounted above a cafe door. ``` State the wording, the lettering style, and where it sits. ## No negative prompt Krea 2 does not use a negative prompt - the graph zeroes the negative conditioning out. Do not pass `negative`; phrase exclusions positively in the positive prompt instead (`a clean, empty desk` rather than leaning on `negative: clutter`). ## Expanding a terse user request When the user asks for something short like "a fox in the snow", do **not** pass that string through - expand it into a Krea 2 description first: 1. **Anchor the subject** in one clause, then add a few concrete, depictable details. 2. **Place + frame it** - setting and one composition/camera choice. 3. **Light it** - one phrase on light and mood. 4. **Name the medium** - photograph, cinematic still, illustration. 5. **Assemble as one or two sentences**, not a comma dump. Sketch: `a fox in the snow` becomes ```text A red fox trotting across fresh snow at dawn, soft backlight catching the mist of its breath, shallow depth of field, muted blue-grey winter palette, editorial wildlife photograph. ``` If the user hands you a finished descriptive prompt, preserve their wording and intent - lightly polish rather than rewrite. If they hand you a tag list, translate it into descriptive language while keeping every concept. ## Prompt enhancement (`enhance`) The `generate_image` tool has an opt-in **`enhance`** option that routes your prompt through a separate model before the render, rewriting a thin prompt into the rich Krea-2 description this guide covers. - **Opt-in.** Pass `enhance: true` (or a workflow/config default turns it on). When off, your `prompt` is sent verbatim. - **Use it when** the incoming prompt is a few words or a loose list. When you have already written a careful description, leave it off - it only adds latency and can drift from your intent. - **Scene continuity.** Pass the `context` arg alongside `enhance` to hand the enhancer background to honour (character facts, ongoing scene) without depicting it literally. Ignored when `enhance` is off. ## If you are enhancing a prompt rather than rendering You may be reading this not to call `generate_image`, but as guidance handed to the prompt-enhancement step: you were given a rough prompt and must return the single enhanced positive string and nothing else. In that role: - **Act only on the prompt-writing rules above** - the [positive recipe](#positive-prompt-recipe), the [terse-request expansion](#expanding-a-terse-user-request), and [text rendering](#rendering-text-in-the-image). The workflow tables and tool args do not apply; you do not pick a workflow or call a tool. - **Translate and enrich, do not reinvent.** Preserve the subjects, actions, colours, and spatial relationships you were given; do not add new objects, people, or props the request does not imply. Honour the medium the user asked for (photo, illustration, painting) rather than pivoting to an easier one. - **Write one cohesive description** in Krea's natural-language style - no tags, no bullets, no negative prompt. - **Layer in depictable specifics** the prompt leaves unstated (lighting, materials, palette, framing, lens, mood), and fold in any background context you were given as source material - but keep it faithful to the stated subject. - **Respect the human form** - treat people with dignity and assume clothing. ## Worked examples These show the _shape_ of a finished prompt; adapt them to the request rather than reusing them. The first two are validated Krea 2 Turbo renders. Photoreal portrait (`workflow: krea2`): ```text A cinematic golden-hour portrait of a weathered fisherman on a wooden dock, shot on 85mm f/1.4, soft warm rim light, shallow depth of field, visible film grain, natural skin texture, editorial photograph. ``` Product / still life (`workflow: krea2`): ```text A close-up product shot of a vintage brass compass on weathered nautical charts, dramatic side lighting, shallow depth of field, fine metal and paper texture, editorial photograph. ``` Reusable skeletons: ```text Portrait: [shot size] portrait of [subject] in [wardrobe], [setting], [lighting], [lens feel], [colour grade] Cinematic: A cinematic film still of [scene], [time of day], [lighting], [atmosphere], [colour grade] Product: [style] product shot of [object] on [surface], [lighting direction], [detail cues] ``` ## Anti-patterns - **Booru / quality-tag spam.** `1girl, solo, masterpiece, 8k, score_9` underuses Krea 2 and flattens the output. Write descriptors or prose instead; translate any handed-in tag list. - **Artist / score tags.** There is no `@artist`, `artist:name`, or score-tag syntax. Name a medium, era, or movement in words. - **Passing a `negative`.** It is zeroed out and does nothing. Phrase the exclusion positively in the prompt. - **Overriding `steps` / `cfg`.** The 8-step / cfg-1 distilled recipe is baked; raising either breaks the model. Change the look through the prompt, not the sampler. - **Over-long rendered text.** Krea renders a word or short sign well, not a paragraph. Keep quoted text short. - **Expecting cloud-only features.** The open-weight Turbo has no `creativity` dial, moodboards, or style-reference images - those belong to Krea's hosted API. Local style shifts come from LoRAs, not those knobs.
Ver en GitHub