Candid Phone Look – Keeps Face & Age
A small LoRA for Qwen-Image 2.1 that keeps adult faces and ages steady, with a light everyday phone-photo feel. Built with Qwen.
🎯 What it does (honest summary)
The phone-snapshot look is subtle: this LoRA does not reliably give flash-on-camera, flat-light or unretouched-skin snapshots. What it does well: it does not make the woman in your prompt look older, and at this checkpoint faces did not collapse into one identity (they did at checkpoint 800), though some of the faces look alike. For a stronger snapshot look, also write "candid phone photo" words in your prompt.
🖼️ Base vs LoRA
Same prompt and seed. Left: Qwen-Image 2.1 without the LoRA. Right: with this LoRA (strength shown on each image).
🚀 Quick start
• Base model: Qwen-Image 2.1 (ComfyUI 0.37 or newer)
• Trigger word: cndphoto (put it first in your prompt)
• Strength: 0.5 to 0.8, start at 0.5
• Negative prompt used for the examples: blurry, low quality, watermark, text, logo, ugly, deformed, extra limbs, extra fingers
⚠️ Limits you should know
• Tested on 13 prompts only: adult women and everyday objects/scenes, all SFW
• Some faces looked similar to each other with bangs and dark hair, so do not expect many different faces
• Text on objects can be garbled, as in the base model
• It is a test-quality LoRA (checkpoint 600 of 1100 steps), not a finished style
✨ Features
• Keeps the age you ask for (tested on "adult Korean woman in her early 20s" prompts)
• Light touch on everyday scenes: bus stop, kitchen, park path, office desk, night street stall, balcony, train seat
• Did not push our non-snapshot test prompts (studio portrait, product shot, landscape) toward a phone-photo look
• One small file (about 80 MB), rank 16
📦 Required Models
• qwen_image_2.1_int8_convrot.safetensors (7.26 GB) — diffusion model → ComfyUI/models/diffusion_models/
• qwen3vl_8b_int8_convrot.safetensors (9.35 GB) — text encoder → ComfyUI/models/text_encoders/
• qwen_image_2.1_vae_bf16.safetensors (0.68 GB) — VAE → ComfyUI/models/vae/
• This LoRA (Qwen21_CandidPhone_v1.safetensors) → ComfyUI/models/loras/
⬇️ Download Links (base model)
• https://huggingface.co/Comfy-Org/Qwen-Image-2.1/resolve/main/vae/qwen_image_2.1_vae_bf16.safetensors
🔧 How to Use in ComfyUI
1) Load Qwen-Image 2.1 as usual and add a "LoraLoaderModelOnly" node after the diffusion model.
2) Select Qwen21_CandidPhone_v1.safetensors and set strength 0.5 to 0.8.
3) Start your prompt with the trigger word cndphoto, then describe the person and scene.
4) For a stronger phone-snapshot look, also write it in the prompt yourself, for example "casual unedited smartphone snapshot, direct on-camera flash".
⚙️ Settings used for the example images
• Strength: 0.8 for the bus stop (cover), train seat and iced tea examples; 0.5 for the kitchen, park, office, night stall and balcony examples
• Size 1344 x 1536 · KSampler euler / simple · 25 steps · CFG 1
• Each example has a base-model picture made with the same prompt and seed without the LoRA
💡 Tips
• Start at 0.5 and compare with the same prompt and seed without the LoRA. At 0.8 the train-seat example (07) came out with a different coat and seat than its base picture, so check each result against the base
• We only tested strengths 0.5, 0.8 and 1.0
• Use "adult" in your prompt; all training and test images were adult subjects
📝 Notes & AI Disclosure
All example images are AI-generated with Qwen-Image 2.1 and this LoRA. All people shown are fictional adults, SFW. The LoRA was trained on 40 AI-generated images (made with Krea 2 turbo) of adult subjects and everyday scenes.
⭐ Found this useful?
• Like if it saved you time
• Comment your results — I read every one
• Follow for new ComfyUI workflows, all tested on 16 GB VRAM
⚖️ Model Attribution & Licensing
Built with Qwen. Qwen-Image 2.1 by the Qwen team (Alibaba), model files from Comfy-Org/Qwen-Image-2.1 on Hugging Face. This LoRA is built on Qwen-Image 2.1, which is licensed under the Qwen Research License Agreement (research / evaluation use; commercial use needs a separate license from the licensor). Please read the license on that repository before using this LoRA. Images you make are your own responsibility; check the platform rules.
Description
v1: checkpoint 600 of a 1100-step run. Trigger word cndphoto, strength 0.5-0.8. Keeps face and age; the phone-photo look is subtle, add candid phone photo words to your prompt for more.