## REALSTAGRAM ZIMG
Subtle realism LoRA for Z-Image Turbo. Designed to stack on top of character LoRAs to enhance photorealism without overpowering the underlying generation — preserves Z-Image's native composition and lighting while pushing the output toward an amateur / candid Instagram look.
🤗 HuggingFace mirror: https://huggingface.co/diobrando0/realstagram-zimg
## Usage
- No trigger word. Stack on top of a character LoRA at strength 0.2 – 0.6 to preserve the character's identity while picking up the realism look. Higher strengths can override the character.
- Base model: Z-Image Turbo / Z-Image De-Turbo
- Rank: 64 · BF16 · 325 MB
## Files
Sample ComfyUI workflow + raw with-LoRA / without-LoRA pairs available on the HuggingFace mirror linked above.
---
## REALSTAGRAM QWEN 2512 v0.1
Same Instagram-style realism approach, ported to Qwen-Image-2512. Trained on a curated 400-image dataset, This qwen version mostly produces young adults but can also be used as a realism lora at lower strength. Personally I'm not much of a fan of this but if you're looking for a unique face to use as a source image for you character lora dataset this could be a good option as it does make some pretty good realistic facial features and doesn't converge to one specific face.
⚠️ **Early release.** Trained on a relatively small dataset (400 images), so it's strongest on prompts close to the training distribution and can feel rigid on very different scenes. I spent a lot of time iterating on this run so I'm releasing it anyway — v0.2 with a larger and more diverse dataset is in the works.
🤗 HuggingFace mirror: https://huggingface.co/diobrando0/Realstagram_qwen2512_v0.1
## Usage
- No trigger word. Style is always-on when the LoRA is loaded.
- Strength 1.0 for full effect. Drop to 0.4 – 0.7 when stacking with character LoRAs to preserve identity.
- Base model: Qwen-Image-2512
- Pair with Qwen-Image-2512-Lightning-8steps at ~0.82 for fast inference (sample workflow included).
- Rank: 256 · BF16 · ~4.4 GB
- Training: 10k+ steps · LR 5e-5 cosine
## Files
Sample ComfyUI workflow + 12 side-by-side comparisons (REALSTAGRAM vs base QWEN 2512) + raw with-LoRA / base outputs all available on the HuggingFace mirror linked above.
Description
version 1.0
FAQ
Comments (11)
I don't mean to be a hater but all these effects you're showing can be achieved solely through prompting.
In ComfyUI I only get black output with your Qwen 2512 LoRA activated. I tried it with the download here and the file on huggingface. Other LoRAs work fine. I use the fp4 Nunchaku version of Qwen Image 2512.
Your sample_workflow.json cannot be opened, too. The file contains JSON but ComfyUI just says "Empty canvas". ComfyUI is up-to-date.
4GB? You sure you got the rank correct?
I'ts rank 256 but I trained on non quantsized so thats why it's so big. Should have just gone on fp8
A 4GB lora?
oeryonMay 15, 2026
Looks great but I don't think i want to have a lora competing for 4GB of my vram
amazingbeautyMay 15, 2026
yes















