Create README.md

README.md

@@ -0,0 +1,78 @@
+---
+license: apache-2.0
+language:
+- en
+tags:
+- diffusion
+- text-to-image
+- latent-diffusion
+- pytorch
+pipeline_tag: text-to-image
+---
+
+# SykoDiffusion V1.0
+
+İlk versiyon latent diffusion modelim. CLIP text encoder ve VAE kullanarak metinden görüntü üretir.
+
+## Model Detayları
+
+| Özellik | Değer |
+|---|---|
+| Parametre | ~100M |
+| Mimari | Latent Diffusion (U-Net) |
+| Eğitim Verisi | CC3M (~100k görsel) |
+| Eğitim Adımı | 20.000 step |
+| Çözünürlük | 256×256 |
+| Donanım | 2× NVIDIA T4 |
+
+## Kullanım
+
+```python
+import torch
+from diffusers import UNet2DConditionModel, AutoencoderKL, DDIMScheduler
+from transformers import CLIPTextModel, CLIPTokenizer
+from PIL import Image
+import numpy as np
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+unet      = UNet2DConditionModel.from_pretrained("SykoSLM/SykoDiffusion-V1.0").to(device).half()
+vae       = AutoencoderKL.from_pretrained("stabilityai/sd-vae-ft-mse").to(device).half()
+clip      = CLIPTextModel.from_pretrained("openai/clip-vit-large-patch14").to(device).half()
+tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14")
+scheduler = DDIMScheduler(num_train_timesteps=1000, beta_start=0.00085,
+                          beta_end=0.012, beta_schedule="scaled_linear", clip_sample=False)
+
+@torch.no_grad()
+def generate(prompt, steps=30, cfg=7.5, seed=42):
+    torch.manual_seed(seed)
+    tokens     = tokenizer(prompt, padding="max_length", truncation=True, max_length=77, return_tensors="pt").to(device)
+    text_emb   = clip(**tokens).last_hidden_state
+    neg_tokens = tokenizer("", padding="max_length", truncation=True, max_length=77, return_tensors="pt").to(device)
+    neg_emb    = clip(**neg_tokens).last_hidden_state
+    emb        = torch.cat([neg_emb, text_emb])
+    latents    = torch.randn(1, 4, 32, 32, device=device, dtype=torch.float16)
+    scheduler.set_timesteps(steps)
+    for t in scheduler.timesteps:
+        pred          = unet(torch.cat([latents]*2), t, encoder_hidden_states=emb).sample
+        neg_p, text_p = pred.chunk(2)
+        pred          = neg_p + cfg * (text_p - neg_p)
+        latents       = scheduler.step(pred, t, latents).prev_sample
+    image = vae.decode(latents / vae.config.scaling_factor).sample
+    image = (image.clamp(-1,1)+1)/2
+    image = (image[0].permute(1,2,0).cpu().float().numpy()*255).astype("uint8")
+    return Image.fromarray(image)
+
+img = generate("a cat sitting on a chair")
+img.save("output.png")
+```
+
+## Notlar
+
+- Bu model deneysel bir ilk versiyondur, üretim kalitesi sınırlı olabilir.
+- En iyi sonuç için `cfg` değerini 5–10 arasında deneyin.
+- İngilizce prompt önerilir.
+
+## Geliştirici
+
+[@SykoSLM](https://huggingface.co/SykoSLM)