The Ultimate Guide to Fooocus Image Prompts (2026 Update)

Fooocus's Image Prompt feature is not a text trick — it's a dedicated input mode (Input Image → Image Prompt) that feeds reference images into generation alongside your text. Fooocus augments IP-Adapter with a pre-computed negative embedding, attention modifications, and adaptive weighting of its own — which is why reference images and text prompts work together here instead of the image drowning out the text.

This guide covers the four Image Prompt modes, the Stop At and Weight controls, the text-prompt syntax that actually changes results, and copyable prompt patterns. Everything factual below is sourced from the official Fooocus repository and its official Image Prompt documentation (v2.5, checked July 2026) — where a claim comes from community documentation instead, it's labeled. One thing to know upfront, honestly: Fooocus is in limited long-term support — bug fixes only, with no current plans to adopt newer model architectures (the project notes this may change). It remains a simple way to run SDXL-class generation locally, but it is not where new features land.

Pick your goal

How to use Fooocus Image Prompt: the four modes

Enable the Input Image checkbox, open the Image Prompt tab, and drop reference images. Each image slot can use one of four modes (under Advanced):

ImagePrompt

The default mode. Per the official docs, it combines IP-Adapter with a Fooocus-built negative embedding, attention hacking, and adaptive balancing — the practical result being that your text prompt still matters while the reference image steers style and appearance. Use it when you want "like this image, but…" behavior: style transfer, a consistent aesthetic, or appearance guidance.

PyraCanny

Pyramid-based Canny edge detection. The docs note that standard Canny "tends to miss some image details" at Fooocus's high working resolution, so PyraCanny detects edges at multiple resolutions and merges them. Use it to guide structure: a pose, a layout, a silhouette. Your text prompt then dresses that structure. Downloads a ~396MB control model on first use.

CPDS

A structure-extraction algorithm based on Contrast Preserving Decolorization — it takes only the structural part of the reference, and uses a fast preprocessor that requires no separate model download. Choose CPDS over PyraCanny when you want looser structural guidance or faster iteration. Its control model is also a ~396MB first-use download.

FaceSwap

Input Image → Image Prompt → Advanced → FaceSwap. Puts a reference face onto generated characters. Combine with a text prompt describing everything except the face. The usual caution applies more strongly here than anywhere else on this page: only use faces you have the right to use.

Combining modes: the docs are explicit that Fooocus supports multiple image inputs without quality loss — a deliberate design difference from implementations where each extra reference degrades output. A common combination is one PyraCanny image for pose plus one ImagePrompt image for style.

Stop At and Weight

Each image prompt slot has two sliders (community-documented behavior — the official announcement doesn't spell them out):

• Weight — how much this reference contributes, relative to the text prompt and other references. Raise it when the reference is being ignored; lower it when outputs look like clones of the reference.
• Stop At — how far through the generation the reference stays active (0–1). At lower values Fooocus stops consulting the reference partway and finishes on its own, which loosens the grip of the reference. If PyraCanny results feel rigid, lowering Stop At is the usual first move.

There are no universally correct values — they interact with your model, reference, and prompt. Change one slider at a time so you can tell what did what.

Prompt syntax that actually changes results

All of the following is official documented behavior:

• Weights: (happy:1.5) raises a token's influence — Fooocus uses A1111's reweighting algorithm, so weights behave the way A1111 users expect. Embeddings load as (embedding:file_name:1.1).
• Prompt expansion ("Fooocus V2"): a GPT-2-based expander that enriches short prompts automatically — it's why two-word prompts still produce detailed images. It's a style entry: unticking Fooocus V2 in Advanced → Style stops the GPT-2 expansion, but any other selected styles still add their own prompt text — untick those too for full literal control.
• Styles: Advanced → Style applies curated prompt/negative-prompt bundles. Styles stack — a few well-chosen ones beat ten.
• Wildcards: __color__ flower pulls a random entry from wildcards/color.txt on each generation — the built-in way to batch variations.
• Negative prompts: Advanced → Negative Prompt. With Fooocus's defaults doing heavy lifting, negatives are for specific exclusions, not paragraph-length incantations.

Copyable prompt patterns

These are templates built on the documented behavior above — fill the brackets, and adjust weights to your reference images. They are starting points, not tested recipes: what works depends on your model, references, and settings.

1. Style transfer (ImagePrompt + text)

Reference: [image with the style you want], mode ImagePrompt.
Text: [subject you want], [setting], (detailed:1.2) — describe the content; let the reference carry the style.

2. Pose copy (PyraCanny + text)

Reference: [image with the pose/composition], mode PyraCanny.
Text: [character description], [clothing], [environment], [lighting] — the text dresses the structure; if the output clings too hard to the reference, lower Stop At.

3. Reuse a reference face (FaceSwap + text)

Reference: [clear front-facing face image], mode FaceSwap.
Text: [everything except the face: body, outfit, scene, mood] — keep the face description out of the text so the two inputs don't fight.

4. Multi-character scene (text weighting)

Two characters: (a tall man in a red coat:1.2) standing on the left, (a short woman in a blue dress:1.2) standing on the right, [shared setting], [lighting]
Weighting both characters equally may balance their influence, but it does not prevent attribute mixing — SDXL-class models still mix characters sometimes; regenerate rather than over-engineering the prompt.

5. Literal control (expansion and styles off)

Untick Fooocus V2 — and every other style — in Advanced → Style, then: [exact scene description, every element you want, nothing you don't] — with expansion and styles off, no extra prompt text is added, so write more than you would with them on.

6. Batch variations (wildcards)

a __color__ [subject] in a __weather__ [setting] — with matching wildcards/*.txt files, each generation rolls new combinations.

Hardware requirements

Official minimums: 4GB Nvidia VRAM and 8GB system RAM for RTX 20/30/40-series cards. GTX 10-series is listed at 8GB VRAM, with 6GB explicitly marked uncertain — "some people report success, some report failure." Full setup: see our Fooocus installation guide.

Should you still use Fooocus in 2026?

Honest answer: it depends what you want. The project is in limited long-term support — bug fixes only — and states there are no current plans to incorporate newer model architectures (its wording leaves room for that to change). That means: simple, free local SDXL generation with the image-prompting approach described above — and, for now, no Flux-class or newer models. If you need newer architectures, that's ComfyUI's territory, at the cost of Fooocus's simplicity.

Troubleshooting

Community-reported patterns (not our test results):

• Reference image seems ignored → confirm the Input Image checkbox is on and the Image Prompt tab (not Upscale/Inpaint) is active; then raise Weight.
• Output is a near-clone of the reference → lower Weight, then lower Stop At.
• PyraCanny output feels stiff → lower Stop At before touching Weight.
• Faces mangled in wide shots → generate closer crops or use inpainting to repair.
• Style won't apply → check whether a stacked Style is fighting your reference; strip styles to isolate.

Related guides

• Fooocus installation and first run
• Using LoRAs in Fooocus
• Fooocus inpainting guide
• SDXL models in ComfyUI
• The PromptZone AI prompt library

FAQ

Can Fooocus use multiple image prompts at once?
Yes — and per the official docs, quality does not degrade with multiple image inputs. Combining a structural reference (PyraCanny) with a style reference (ImagePrompt) is a standard workflow.

What's the difference between PyraCanny and CPDS?
Both control structure. PyraCanny extracts multi-resolution edges — precise, best for exact poses. CPDS extracts looser structure and uses a fast preprocessor requiring no separate model download. Start with CPDS for speed, switch to PyraCanny when you need exactness.

Is Fooocus still maintained in 2026?
It's in limited long-term support: bug fixes only, with no new model architectures currently planned. Maintained, but not evolving.

---

Last reviewed and updated: July 2026. This revision is a source-corrected documentation guide: factual claims are sourced from the official Fooocus repository and its Image Prompt documentation (v2.5, checked July 2026); Stop At/Weight behavior and troubleshooting patterns are community-documented, as labeled; the prompt patterns are templates, and no generated-image tests were run for this revision — a tested-outputs update is planned. Change log: July 2026 — rewritten around the four documented Image Prompt modes, added controls guide, syntax reference, prompt patterns, hardware and maintenance status, troubleshooting, and FAQ; removed unrelated citations from the previous version.