FLUX 3 Image Klein · Open Skill 公開スキル

FLUX 3 Image Klein

**Text to Image · Image Editing · Up to 10 References · Up to 2K** *Describe a scene, bring a photo, or combine references into a new image.* **FLUX 3 Image Klein** is an image creation Agent Skill for Digen. Turn a short description into an image, make a focused edit, or give each reference a role in a new composition. **Describe → compose → create → refine** Start with words or upload images. Choose photography, illustration, product imagery, or a graphic layout. Continue in the same conversation to adjust the result, with replies and guidance in your language. ## What you get | For creators | For the image | | --- | --- | | **Words to image** — start with a short description | **A clear composition** — subject, setting, light, and style | | **Edit a photo** — add, remove, replace, or restyle | **Focused changes** — specify what changes and what stays | | **Up to 10 references** — bring people, products, places, or styles | **Defined roles** — say what to take from each image | | **Keep refining** — make the next change in the same chat | **Your language** — localized replies and requested on-image wording | ## Pipeline | Stage | Output | | --- | --- | | **Describe** | Your idea, edit request, or reference images | | **Compose** | Subject placement, visual style, and reference roles | | **Create** | A generated or edited image | | **Refine** | A focused change to the selected result | ## Production stack | Parameter | Default | Notes | | --- | --- | --- | | **Model** | FLUX 3 Image Klein | Image generation and editing | | **Orientation** | Portrait, **9:16** | Other ratios or source framing on request | | **Resolution** | **1K target**, up to **2K** | Request 2K for a larger render; dimensions vary with aspect ratio | | **References** | Optional, up to **10 images** | One image can be the editing base | | **Look** | Photography | Illustration and other styles on request | | **Language** | Matches your request | Keep supplied on-image text in its original language | ## Modes | Mode | You bring | Agent delivers | | --- | --- | --- | | **Text to image** | A scene or idea | A new image shaped around your description | | **Photo edit** | An image and a change | An edit guided by your preservation instructions | | **Reference composition** | Up to 10 images and their roles | A new composition using those references | | **Layout control** | Placement notes or defined regions | A layout guided by your spatial instructions | | **Follow-up edit** | A result and your next change | A revision in the same conversation | | **2K render** | An image or brief and a request for 2K | A new render at the requested tier, keeping your framing | ## Example prompts **Text to image** > A forest-green perfume bottle on dark stone. Soft studio light, warm gray background, no lettering. Vertical. **Photo edit** > Change the blue mug to dark green. Keep its shape, position, and lighting, and leave the rest unchanged. **Reference composition** > Use the person from image one, the coat from image two, and the corridor from image three. Full-length portrait with natural lighting. **2K render** > Render this at 2K. Keep the same subject, style, and vertical composition. ## When to use - Create a scene, portrait, product image, illustration, or poster. - Remove, replace, recolor, or restyle part of an existing image. - Combine references with a clear role for each. - Specify where subjects and text should appear. - Refine an image without leaving the conversation. - Request a larger render at up to 2K. > Not the right tool when you need video, animation, or a layered, editable design file. ## Invoke Select **FLUX 3 Image Klein**, or ask: *"Create this image with FLUX 3 Image Klein."* Your idea · Your references · Your next edit.

@official

509 24 117

https://d7qlwi7tl2u27.cloudfront.net/apps/demo-agent-7d7fa70dd996.webp

# FLUX 3 Image Klein

Create a new image from text, edit an existing image, or compose a new scene from up to ten references. Official templates define the image prompt format only; they do not govern conversation replies. Preserve requested on-image wording verbatim.

## When to use

Use for text-to-image and image-to-image requests, including reference composition, local changes, whole-image transformations, and follow-up edits. A reference is optional for text-to-image.

## Steps

1. Identify whether the user wants an image or only a prompt. For prompt-only requests, prepare and return the requested prompt; do not call `generate_image` or run generation retries. Otherwise, extract subjects, counts, style, wording, and changes. Ask only about ambiguity that materially changes the result.
2. Choose the actual reference images. Follow-up edits use the image the user identifies, or the latest successful result when that is unambiguous, as the editing base; add original references only if needed. Inspect each selected reference with `recognize_image`, reusing reliable existing analysis where appropriate. Keep the selected order stable and assign each image a role. Accept at most ten images; never invent content, URLs, or dimensions.
3. Read `references/tools.md` and decide the target framing and output tier before expanding. The default target tier is 1K; honor explicit 1080P or 2K requests, with 2K as the maximum. Default to 9:16 portrait. An explicit ratio or a request to preserve source framing takes priority. Use confirmed metadata for source aspect, never a guessed pixel size. Supply `TARGET CANVAS: W:H` to the expansion step using the decided ratio, then remove that marker from the final image prompt. When a different target canvas requires extending or recomposing an edit, state that transformation explicitly while preserving the requested subject and content.
4. Read exactly one matching official template with `read_skill_file`: no images uses `references/t2i.md`; one to ten images uses `references/i2i.md`. Use `references/t2i_bbox.md` or `references/i2i_bbox.md` instead only for explicit boxes or separately controlled regions. These are internal expansion instructions, not text to paste into the image prompt.
5. Expand once into the selected template's final caption or instruction. Resolve one consistent composition, retain user choices, and specify concrete relationships, light, and material cues. For edits, describe the change and its local effects with a compact preservation clause; do not redescribe the entire source. For composition, bind each reference to its contribution and reconcile scale, perspective, contact, light, and shadows. Keep any newly authored visible copy in the user's requested language. Do not add a second planning template or rewrite pass after producing the final prompt.
6. Call `generate_image` with that final prompt: `model='t2i.flux'` with no references or `model='i2i.flux'` with the selected references. Pass the framing already decided in step 3. For the default 1K target, omit `resolution`; for a requested 1080P output, pass `resolution='1080P'`; for a requested 2K output, pass `resolution='2K'`. A request to upgrade an existing image changes its tier, not its aspect or content: preserve its established framing and composition unless the user asks to change them. Do not request a tier above 2K or change the selected tier during a retry. Do not change modality or reference order between expansion and generation.
7. Check tool success and inspect the image directly if visible, otherwise use `recognize_image` when inspection is necessary. Check the requested content, text, reference ownership, framing, and preservation constraints. Also inspect `param_adjustments` and reliable returned dimensions when available; a requested tier is not proof of actual output dimensions. Make at most one targeted visual correction for an observed failure within the request's scope. A transient-failure retry is

Digen をシェア

FLUX 3 Image Klein · Open Skill 公開スキル

FLUX 3 Image Klein