Pixify feature

Character Consistency + Batch Render

Same character, multiple poses, scenes, and shots — kept consistent

Character Registry: save reference + style
Drag character cards from sidebar onto canvas
Batch Render: N prompt variants in one shot
Consistency Eval: auto-score similarity (aHash)

Use characters in the editor Explore all features

What is it

When you need many images of the same character (different poses, angles, scenes), rewriting prompts each time causes drift. Character Registry solves this: save the character once (reference image + style prompt), and downstream nodes (t2i / i2i / i2v) auto-reuse them — keeping the character stable. Pair with Batch Render for N variants and Consistency Eval for 0-100 similarity scoring.

How to use it

Get started in 6 steps

1
Prepare a character reference image
Generate a satisfying character via AI Image, or upload your own design. This is the "seed" — clarity matters.
2
Write a style prompt
1-3 sentences describing the character's key features: "long black hair, blue eyes, white shirt, silver necklace". Auto-injected into downstream prompts.
3
Save to the library
In the character panel, click "Save to Library". The character card appears in the sidebar Characters tab.
4
Drag a card to canvas
From Characters tab → drag onto canvas → auto-creates an imageInput node with the reference image pre-filled.
5
Connect to downstream
Wire the imageInput to Image to Image / Image to Video image ports. The reference image injects into image_urls.
6
Batch render + score
Add Batch Render with prompt variants ("at a cafe", "at the beach", "on a train"). Add Consistency Eval to auto-score each output against the reference.

Use cases

What other users build with it

Comics / picture books

Main character stable across 20 panels — no per-frame prompt babysitting.

Virtual host series

Same persona across 100 episodes — no face drift.

Product spokesmodel

One virtual model across all product asset variations.

Game character portraits

Same character — idle, attack, hit, victory expressions.

Why Pixify

No training needed

Unlike LoRA which needs 20+ training images, just one reference + style description.

Sidebar-native

Saved characters appear in sidebar Characters tab — drag to use.

Quantified evaluation

Consistency Eval gives 0-100 score, surfaces the most stable variant.

Workflow-first

Not a standalone tool — first-class workflow citizen, chains naturally with everything else.

Frequently asked questions

How consistent can it get?

Depends on reference quality and style prompt precision. Typical aHash scores 70-90 / 100. Pixel-perfect identity is impossible (each generation samples fresh), but "clearly the same person" is achievable.

What algorithm does Consistency Eval use?

v2 uses sharp aHash (average hash) — fully local, zero API calls. Sensitive to overall appearance similarity. CLIP embedding may be added later (pending new provider business).

Max concurrent Batch Render jobs?

Currently serial (one prompt finishes before the next), cap 8. Production-grade Cloud Tasks queue under evaluation.

Can I train my own character LoRA?

Not yet. LoRA fine-tuning needs a new provider and business approval (pending). Current approach (reference + style injection) works even with insufficient training samples.

Ready to start?

Use characters in the editor