Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Train and manage custom FLUX LoRA fine-tunes on Replicate for consistent brand character imagery
tier
extended
applyTo
**/*lora*,**/*finetune*,**/*fine-tune*,**/*flux*
metadata
{"inheritance":"inheritable"}
FLUX Brand Fine-Tune
Train custom LoRA models on FLUX Dev for consistent Alex (or any character) imagery across all generated content.
Pattern: Train once with 10–20 curated images → get a trigger-word-activated LoRA → generate unlimited consistent images at ~$0.003–0.04/image.
Why Fine-Tune?
Approach
Consistency
Cost/Image
Setup Cost
Best For
Nano-Banana Pro (reference images)
Good
$0.025
Free
Quick one-off sets
Flux 2 Pro (input_images)
Good
~$0.05
Free
Multi-ref scenes
FLUX LoRA Fine-Tune
Excellent
$0.003–0.04
~$1.50
Brand-scale consistency
Fine-tuning wins when you need hundreds of images across marketing, docs, presentations, and social media — all with the same character looking exactly the same without passing reference images every time.
Trainer Models
Trainer
Replicate ID
Hardware
Cost
Training Time
Notes
Fast FLUX Trainer (recommended)
replicate/fast-flux-trainer
8× H100
$0.0122/sec ($1.50 total)
~2 min
Newer, faster, subject/style selector
FLUX Dev LoRA Trainer
ostris/flux-dev-lora-trainer
H100
$0.001525/sec ($0.98 total)
~10 min
Original trainer, more parameters
Recommendation: Use replicate/fast-flux-trainer for most use cases — faster, cheaper, simpler.
Training Data Requirements
Image Guidelines
Spec
Recommendation
Quantity
10–20 images (minimum 10, sweet spot 15–20)
Resolution
1024×1024 or higher (will be scaled down)
Formats
WebP, JPG, PNG
Variety
Different angles, lighting, expressions, settings, clothing
Consistency
Same character/subject across all images — same age, same identity
For Character/Subject LoRAs
Include varied backgrounds (indoor, outdoor, studio, natural)
Show different facial expressions (neutral, smiling, serious, thoughtful)
If captions aren't provided, the trainer auto-generates them using LLaVA.
Trigger Word Selection
The trigger word activates your LoRA concept in prompts.
Rules:
Must be unique — NOT a real word in any language
NOT TOK (clashes with other fine-tunes)
Case-insensitive but ALL CAPS helps visually
Short, memorable
Examples:
ALEXFNCH — Alex Finch character
CXRBRAND — CorreaX brand style
ALEXCOG — Alex cognitive architecture style
Training Workflow
Step 1: Prepare Training Data
# Create a training data folder
New-Item -ItemType Directory -Path "training-data" -Force
# Optionally resize images to 1024px (if oversized)
# Requires ImageMagick:
# macOS: brew install imagemagick
# Windows: winget install ImageMagick.ImageMagick
Get-ChildItem training-data/*.jpg, training-data/*.png | ForEach-Object {
magick $_.FullName -resize "1024x1024>" $_.FullName
}
# Zip the training data
Compress-Archive -Path "training-data/*" -DestinationPath "training-data.zip" -Force
Step 2: Upload Training Data
Upload the zip to a publicly accessible URL, or use the Replicate web UI to upload directly.
import { readFileSync } from"fs";
// Option A: Use Replicate's file uploadconst replicate = newReplicate({ auth: process.env.REPLICATE_API_TOKEN });
// Upload via the web UI at https://replicate.com/replicate/fast-flux-trainer/train// Option B: Use any file hosting (S3, Azure Blob, etc.)// The URL must be publicly accessible during training
Step 3: Create Training via API
importReplicatefrom"replicate";
const replicate = newReplicate({ auth: process.env.REPLICATE_API_TOKEN });
const training = await replicate.trainings.create(
"replicate", // model owner"fast-flux-trainer", // model name"8b10794665aed907bb98a1a5324cd1d3a8bea0e9b31e65210967fb9c9e2e08ed", // version
{
destination: "your-username/alex-brand", // your model destinationinput: {
input_images: "https://your-storage.com/training-data.zip",
trigger_word: "ALEXFNCH",
lora_type: "subject", // "subject" or "style"// training_steps: 1000, // default 1000, range 500–4000
}
}
);
console.log("Training started:", training.id);
console.log("Status URL:", training.urls.get);
// Poll for statusconst status = await replicate.trainings.get(training.id);
console.log("Status:", status.status); // starting → processing → succeeded/failed
# Or via curl
curl -s \
-H "Authorization: Bearer $REPLICATE_API_TOKEN" \
https://api.replicate.com/v1/trainings/<training_id>
Training typically completes in ~2 minutes with the fast trainer.
Step 5: Generate Images with Your LoRA
Once training succeeds, your model is live at the destination:
const output = await replicate.run(
"your-username/alex-brand", // your trained model
{
input: {
prompt: "ALEXFNCH standing in a modern office, wearing a casual blue shirt, " +
"confident smile, natural lighting, professional photography, 8k detail",
aspect_ratio: "3:4",
output_format: "png",
}
}
);
// Save immediately — output URLs expire after 1 hourimport { writeFile } from"node:fs/promises";
awaitwriteFile("alex-office.png", output);
Advanced Training Parameters
ostris/flux-dev-lora-trainer (Full Control)
Parameter
Default
Range
Notes
steps
1000
3–6000
500–4000 recommended
lora_rank
16
1–128
Higher = more complex features, slower training
autocaption
true
—
Uses LLaVA v1.5 13B for auto-captioning
autocaption_prefix
—
—
e.g., "a photo of ALEXFNCH, "
autocaption_suffix
—
—
e.g., " in the style of ALEXFNCH"
hf_repo_id
—
—
Optional: also push to HuggingFace
replicate/fast-flux-trainer (Recommended)
Parameter
Default
Notes
trigger_word
"TOK"
Your unique trigger word
lora_type
—
"subject" or "style"
training_steps
1000
Adjust for dataset size
input_images
—
Zip file URL or direct upload
Using Your LoRA with Other Flux Models
Your trained LoRA weights can be used with any Flux model that supports LoRA:
// Use with Flux Dev (good quality, $0.025/image)const output = await replicate.run("black-forest-labs/flux-dev", {
input: {
prompt: "ALEXFNCH in a coffee shop, reading a book",
extra_lora: "your-username/alex-brand",
extra_lora_scale: 1.0, // 0.0–1.5, default 1.0
}
});
// Use with Flux Schnell (fast iteration, $0.003/image)const output = await replicate.run("black-forest-labs/flux-schnell", {
input: {
prompt: "ALEXFNCH giving a presentation",
extra_lora: "your-username/alex-brand",
}
});
LoRA Scale Tuning
Scale
Effect
0.5
Subtle influence — blends with base model
0.8
Moderate — recognizable but flexible
1.0
Full — strong character likeness (default)
1.2–1.5
Over-fitted — may reduce prompt adherence
Cost Analysis
Operation
Cost
Notes
Training (fast-flux-trainer)
~$1.50
One-time, ~2 minutes
Training (ostris)
~$0.98
One-time, ~10 minutes
Generation (Flux Schnell + LoRA)
$0.003/image
Best for iteration
Generation (Flux Dev + LoRA)
$0.025/image
Best for quality
Generation (Flux 1.1 Pro)
$0.04/image
Highest quality
Break-even vs Nano-Banana Pro: After ~70 images ($0.025 × 70 = $1.75 > $1.50 training), the fine-tune pays for itself and every subsequent image costs only $0.003–0.025 without needing reference photos.
Prompt Engineering for Fine-Tuned Models
Template
TRIGGER_WORD [description of scene], [clothing], [expression], [lighting], [style]
Best Practices
Always lead with trigger word: ALEXFNCH standing in a park not A person standing in a park named ALEXFNCH
Be specific about scene: FLUX follows detailed prompts well
Use quality tokens: "professional photography", "8k detail", "cinematic lighting"
Example Prompts
ALEXFNCH headshot, professional business attire, confident smile, studio lighting, corporate portrait photography
ALEXFNCH sitting at a desk with multiple monitors, coding, casual hoodie, night time, ambient screen glow, documentary photography
ALEXFNCH presenting at a tech conference, standing on stage, gesturing toward a screen, audience in background, event photography
ALEXFNCH in a cozy library, reading a leather-bound book, warm lamplight, bokeh background, editorial photography
Visual Verification (VS Code 1.112+)
After generating with your LoRA, use view_image to evaluate output quality:
Identity fidelity — Does the character match the training data?
Prompt adherence — Does the scene match the prompt, or is the LoRA overfitting?
LoRA scale tuning — Compare outputs at 0.6, 0.8, 1.0 to find the sweet spot
Artifact scan — Check for training artifacts: repeated backgrounds, frozen expressions, texture smearing
When iterating on LoRA scale or prompt structure, compare outputs side-by-side via the image carousel.
Retraining and Iteration
When to Retrain
Character design changed significantly
Original training data was insufficient (< 10 images)
Want to add new scenarios not well-represented
Trigger word conflicts with another LoRA
Iteration Strategy
Generate 10 test images with varied prompts
Evaluate: Does the face/identity hold across scenes?
If weak: add more training images in the weak areas
Retrain with expanded dataset (training is cheap at ~$1.50)
Integration with Visual Memory
If you have a visual-memory.json with reference photos, extract the base64 images back to files for training:
