Realstagram_qwen2512_v0.1

Realism / amateur-Instagram style LoRA for Qwen-Image-2512. Trained on 332 hand-curated Pinterest images to push outputs toward a candid, photographic, social-media aesthetic — flushed skin, soft natural lighting, real-world settings, casual composition.

v0.1 — early release. This run was a learning pass — rank 512 trained on a relatively small dataset, so it's stronger on prompts close to the training distribution and can feel rigid on very different scenes. v0.2 will use a larger/more diverse dataset and a more conservative rank. The comparisons below show it's still a meaningful aesthetic shift vs base Qwen 2512.

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 the character.
• Base model: Qwen/Qwen-Image-2512.
• Inference companion: pair with Qwen-Image-2512-Lightning-8steps-V1.0-bf16 at strength ~0.82 for fast 8-step generation. The included sample_workflow.json is set up for this.
• Rank: 512 linear / 16 conv · BF16 · ~4.4 GB.
• Training: 9296 steps · LR 5e-5 cosine · caption_dropout 0.05 · ai-toolkit 0.7.21 on RTX Pro 6000.

Comparisons

Side-by-side: left = REALSTAGRAM (with this LoRA), right = QWEN 2512 (base, same prompt + seed).

Files

File	Purpose
Realstagram_qwen2512_v0.1.safetensors	The LoRA itself (rank 512, ~4.4 GB BF16)
sample_workflow.json	ComfyUI API-format workflow (Qwen-Image-2512 + this LoRA + Lightning + 3-pass ClownsharKSampler refinement)
comparisons/	12 side-by-side stitched comparisons (REALSTAGRAM vs QWEN 2512)
samples/	12 raw with-LoRA outputs
nolora/	12 raw base outputs (Qwen 2512 + Lightning, no style LoRA), matching seeds

Sample workflow

Drop sample_workflow.json into ComfyUI. It expects:

• qwen_image_2512_bf16.safetensors (UNet) in models/diffusion_models/
• qwen_2.5_vl_7b.safetensors (text encoder) in models/text_encoders/
• qwen_image_vae.safetensors (VAE) in models/vae/
• Qwen-Image-2512-Lightning-8steps-V1.0-bf16.safetensors in models/loras/
• Realstagram_qwen2512_v0.1.safetensors in models/loras/
• Custom node: RES4LYF (ClownsharKSampler_Beta)

License

Apache 2.0. Free to use, modify, distribute, including commercial use.