InstaPic
The goal of this LoRa is to generate Post Production images for Instagram.
Note:
A very important detail is that the model suffered a bit of overfitting, so when you put a prompt that deviates greatly from the standard captions used in the dataset, you end up having lower quality than some specific prompts. This must be because the dataset's captions include many words focused on lighting, facial accessories, and more sensualized scenes, and this causes the real quality to be activated when including these types of words in the prompt. You can notice this in the samples I posted; some are very realistic and others have simpler aspects. Other than that, adjusting the number of steps, cfg, sampler, and scheduler are 100% fundamental to achieving good quality, in addition to the prompt, as I said before.
Tests
Model Versions & Training Details
Training Overview:
Four distinct versions were trained during development, each with different approaches and datasets. However, only Version 1 and the Mixed Version (V1+V3) will be released, as the mixed version demonstrates superior results compared to Version 1 alone.
[InstaPic V1 - Original Foundation]
Core Training Specifications:
Dataset: 600 carefully curated real images with professional post-production
Rank: 256 (resulting in ~4.4GB LoRA file)
Training Tool: Diffusion Pipe with optimized parameters
Focus: Instagram-style content and social media aesthetics
Resolution Optimization: Trained for vertical Instagram formats
The high rank (256) was an experimental study I conducted to test quality retention. This original version establishes the foundation for Instagram-style generation.
[InstaPic Mix (V1+V3) - Enhanced Edition]
Advanced Combined Training:
Base: Version 1 foundation dataset
Enhancement: Combined with Version 3 SDXL-enhanced training data
Quality: Superior results compared to V1 alone
Training: Merged training approach for comprehensive style coverage
[Versions V2 & V4 - Experimental Editions]
V2: High volume training experiments (17k images, lower resolution)
V4: Multi-source fusion with StyleGAN and VTON datasets
Status: Development only - Not planned for release
Purpose: Research and development for future iterations
Available Merged Model Formats
Released Versions:
InstaPic V1 (Original):
Rank 256 - 4.4GB - Original foundation model
InstaPic Mix (V1+V3) - Recommended:
FP16 - Full precision version with maximum quality
FP8 E3M4FN - Optimized compression with maintained quality
SDXL Style LoRA:
InstaPic Style SDXL - Enhanced version trained on V1 images processed through Image-to-Image using the Big Love SDXL model, providing improved detail and SDXL-optimized quality
Pre-Merged Qwen Image Base Model:
Ready-to-use merged versions with original LoRAs embedded:
BF16 (Full Precision) - Maximum quality, larger file size
Q8 (High Quality) - Excellent balance of quality and efficiency
Q6 (Balanced) - Good quality with moderate compression
Q4 (Efficient) - Fastest inference with acceptable quality
🧩 Prompt Template (Dataset Style)
Use this template based on the dataset caption style to achieve superior quality:
1nst4p1c Woman with [detailed hair description], wearing [specific clothing items],
[specific pose/position] in/on [detailed location].
She has [expression] and [hand/body positioning].
[Body visibility/clothing details].
The background is [detailed background description with specific elements].
The lighting is [lighting type] with [lighting effects].
The overall aesthetic is [aesthetic description].
The image is well-composed, with [composition details].
The camera angle is [specific angle], looking [direction] on the subject.
The depth of field is [depth description], with [focus details].
Examples (Dataset Style):
1. Latina – Rooftop Party
1nst4p1c Latina woman with long dark wavy hair, wearing a neon pink crop top and ripped denim shorts with glitter details, posing confidently on a rooftop terrace at night. She rests one hand on her hip while holding a plastic cup with the other, her expression bold and playful. Her bronzed skin glows naturally under purple and red neon party lights, showing realistic texture. The background shows blurred silhouettes of people dancing and the distant city skyline. The lighting is vibrant and cinematic. The overall aesthetic is urban, sensual, and social media ready. The image is well-composed, vertical framing, with shallow depth of field isolating her while the rooftop atmosphere fades softly.
2. Luxury Car – Night Arrival
1nst4p1c Woman with long straight blonde hair, wearing a short black sequin dress and high heels, stepping out of a black Lamborghini parked in front of a luxury hotel entrance at night. She holds a small designer clutch in her hand, her expression neutral but confident. Her fair skin reflects the golden hotel lights with natural highlights. The background shows blurred chandeliers and hotel staff near the glass doors. The lighting is warm and cinematic, mixing neon reflections from the car with golden tones. The overall aesthetic is glamorous, sensual, and Instagram luxury style. The image is well-composed, vertical framing, with both the woman and the Lamborghini sharply in focus while the background remains softly blurred.
3. Gym – Mirror Selfie
1nst4p1c Brazilian morena woman with long black hair tied in a ponytail, wearing a red sports bra and tight gray leggings, posing for a mirror selfie inside a modern gym. She holds her phone slightly tilted in one hand while flexing her waist, lips slightly parted in a playful smirk. Her tanned skin shows natural highlights under the bright overhead gym lights, with subtle sweat detail across her arms. The background shows blurred dumbbells and cardio equipment. The lighting is harsh but realistic, emphasizing her body definition. The overall aesthetic is fitness influencer style, sensual and social media ready. The image is well-composed, vertical framing, with shallow depth of field focusing on her reflection while the gym remains softly visible.
4. Shopping Bags – Luxury Lifestyle
1nst4p1c Woman with long auburn hair and freckles, wearing a beige crop top and skinny jeans, walking down a luxury shopping street carrying several branded shopping bags. She wears sunglasses and has a confident smile as she looks toward the camera. Her fair skin has soft natural texture under the daylight. The background shows blurred storefronts with luxury logos and glass windows. The lighting is bright natural daylight, giving sharp detail and realistic tones. The overall aesthetic is casual luxury, Instagram influencer style. The image is well-composed, vertical framing, with shallow depth of field isolating her while the high-end shops remain softly blurred.
5. Poolside Summer – Sensual Pose
1nst4p1c Woman with pastel pink hair tied into a messy bun, wearing a turquoise bikini and a gold belly chain, sitting at the edge of a swimming pool with her legs slightly apart. She leans back on her arms, gazing at the camera with a subtle seductive smile. Her fair skin glistens with water droplets reflecting the sunlight. The background shows turquoise pool water and palm trees blurred in the distance. The lighting is bright natural daylight, vibrant and crisp. The overall aesthetic is summery, sensual, and influencer-ready. The image is well-composed, vertical framing, with shallow depth of field focusing on her body while the pool background fades softly.
6. Nightclub Neon – Party Scene
1nst4p1c Black woman with curly hair, wearing a glittery silver mini dress and hoop earrings, standing near the bar in a crowded nightclub. She holds a cocktail in one hand while resting the other on the counter, her lips slightly parted in a playful expression. Her dark skin glows under purple and blue neon reflections with realistic highlights. The background shows blurred silhouettes of dancers and glowing neon signs. The lighting is dramatic and colorful, casting cinematic reflections across her skin and dress. The overall aesthetic is urban, sensual, and vibrant. The image is well-composed, vertical framing, with shallow depth of field highlighting her while the nightclub scene fades softly.
Key Dataset Elements (Very Important for Quality):
Specific clothing details (bikini top/bottom, crop top, etc.)
Precise pose descriptions (sitting cross-legged, kneeling, standing near, etc.)
Body visibility statements ("Her body is mostly visible", "wearing only", etc.)
Industrial/urban backgrounds (construction site, concrete, metal, etc.)
Lighting always "soft and diffused"
"Well-composed" always present
Specific camera angles (slightly elevated, looking down)
Depth of field always mentioned
LoRA Recommendation:
Use the Mixed (V1+V3) versions for best results, as they *********** superior quality compared to the original V1 alone.
Optimal Resolution Settings
Recommended Instagram Resolutions:
Stories/Reels: 1080 x 1920 (9:16 aspect ratio)
Alternative Vertical: 1088 x 1920 (optimized for training)
Posts: 1080 x 1350 (4:5 aspect ratio)
Square Posts: 1080 x 1080 (1:1 aspect ratio)
High-Quality Resolutions (divisible by 16):
1536 x 1024 - Landscape format
1024 x 1536 - Portrait format
1536 x 864 - Wide format
864 x 1536 - Tall format
1152 x 1536 - Alternative portrait
1536 x 1152 - Alternative landscape
Resolution Guidelines:
All resolutions should be divisible by 16 for optimal processing
Avoid excessive high resolutions to prevent screendoor effects
Vertical formats preferred for authentic Instagram aesthetics
Height > Width ratios work best with this model
Test different aspect ratios for varied content types
Recommended Sampler/Scheduler Combinations
Standard ComfyUI (Built-in):
Euler Ancestral + Schedulers:
euler_ancestral
+beta
euler_ancestral
+kl_optimal
euler_ancestral
+simple
DEIS 3M + Schedulers:
deis_3m
+beta
RES4LYF Custom Node Required:
Note: These combinations require the RES4LYF custom node installation in ComfyUI
Res 2S + Schedulers:
res_2s
+simple
res_2s
+beta
res_2s
+beta57
res_2s
+bong_tanget
DEIS 3M + Advanced Schedulers:
deis_3m
+beta57
Lightning Model Integration (8 steps):
Compatible with Lightning 8-step models as demonstrated in sample images - provides ultra-fast generation while maintaining quality.
Installation Note:
To access beta57
, bong_tanget
schedulers and some advanced samplers, install the RES4LYF custom node in your ComfyUI environment.
Quality Considerations:
Beta schedulers: Generally provide smoother gradients
Simple scheduler: Faster inference with good quality
KL_optimal: Best for detailed textures
Beta57: Enhanced beta scheduler (requires RES4LYF)
Bong_tanget: Experimental scheduler for unique artistic effects (requires RES4LYF)
Usage Guidelines
Trigger Word:
1nst4p1c
- Always include at the beginning of your prompts
Instagram-Optimized Prompt Structure:
Trigger Word:
1nst4p1c
Subject & Style: Instagram influencer, casual selfie, lifestyle shot
Composition: Vertical framing, close-up, medium shot, full body
Instagram Elements: Phone visible, ring light, modern background
Lighting: Natural light, soft lighting, golden hour, ring light effect
Aesthetic: Instagram filter look, social media ready, influencer style
Technical Specifications
Training Infrastructure:
Primary Tool: Diffusion Pipe
Base Architecture: Compatible with SD 1.5/SDXL models
Optimization: Instagram-specific styling and composition
Post-Processing: Social media enhancement pipeline
Performance Characteristics:
Memory Usage: 4.4GB (V1 Original) / Variable (Mixed Versions) / Variable (SDXL)
Optimal Resolution: Any resolution divisible by 16
Inference Speed: 30-40 steps standard, 8 steps with Lightning models
Style Consistency: High reliability for Instagram aesthetics
Quality Features
Instagram Aesthetics:
Authentic social media styling
Mobile photography look
Modern composition techniques
Social media color grading
Influencer-style posing
Technical Excellence:
Vertical format optimization
Sharp focus with natural depth of field
Consistent lighting and exposure
Professional mobile photography simulation
Anti-screendoor effect optimization
Lightning model compatibility for fast generation
System Requirements & Dependencies
ComfyUI Requirements:
Standard Installation: Basic ComfyUI setup
RES4LYF Custom Node: Required for advanced schedulers (
beta57
,bong_tanget
) and some samplersInstallation: Follow RES4LYF documentation for proper setup
Screendoor Effect Prevention:
Avoid resolutions above 1920 height
Use recommended sampler/scheduler combinations
Test different CFG scales if artifacts appear
Monitor for texture irregularities at high resolutions